-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathwiki_ai.html
405 lines (405 loc) · 166 KB
/
wiki_ai.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
<p class="mw-empty-elt">
</p><p><b>Artificial intelligence</b> (<b>AI</b>), in its broadest sense, is <a href="/wiki/Intelligence" title="Intelligence">intelligence</a> exhibited by <a href="/wiki/Machine" title="Machine">machines</a>, particularly <a href="/wiki/Computer" title="Computer">computer systems</a>.
It is a <a class="mw-redirect" href="/wiki/Field_of_research" title="Field of research">field of research</a> in <a href="/wiki/Computer_science" title="Computer science">computer science</a> that develops and studies methods and <a href="/wiki/Software" title="Software">software</a> that enable machines to <a href="/wiki/Machine_perception" title="Machine perception">perceive their environment</a> and use <a href="/wiki/Machine_learning" title="Machine learning">learning</a> and intelligence to take actions that maximize their chances of achieving defined goals.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig20211–4_1-0"><a href="#cite_note-FOOTNOTERussellNorvig20211–4-1">[1]</a></sup> Such machines may be called AIs.
</p><p>Some high-profile <a class="mw-redirect" href="/wiki/Applications_of_AI" title="Applications of AI">applications of AI</a> include advanced <a class="mw-redirect" href="/wiki/Web_search_engine" title="Web search engine">web search engines</a> (e.g., <a href="/wiki/Google_Search" title="Google Search">Google Search</a>); <a href="/wiki/Recommender_system" title="Recommender system">recommendation systems</a> (used by <a href="/wiki/YouTube" title="YouTube">YouTube</a>, <a href="/wiki/Amazon_(company)" title="Amazon (company)">Amazon</a>, and <a href="/wiki/Netflix" title="Netflix">Netflix</a>); interacting <a class="mw-redirect" href="/wiki/Natural-language_understanding" title="Natural-language understanding">via human speech</a> (e.g., <a href="/wiki/Google_Assistant" title="Google Assistant">Google Assistant</a>, <a href="/wiki/Siri" title="Siri">Siri</a>, and <a href="/wiki/Amazon_Alexa" title="Amazon Alexa">Alexa</a>); <a class="mw-redirect" href="/wiki/Autonomous_vehicles" title="Autonomous vehicles">autonomous vehicles</a> (e.g., <a href="/wiki/Waymo" title="Waymo">Waymo</a>); <a href="/wiki/Generative_artificial_intelligence" title="Generative artificial intelligence">generative</a> and <a href="/wiki/Computational_creativity" title="Computational creativity">creative</a> tools (e.g., <a href="/wiki/ChatGPT" title="ChatGPT">ChatGPT</a>, <a href="/wiki/Apple_Intelligence" title="Apple Intelligence">Apple Intelligence</a>, and <a class="mw-redirect" href="/wiki/AI_art" title="AI art">AI art</a>); and <a href="/wiki/Superintelligence" title="Superintelligence">superhuman</a> play and analysis in <a href="/wiki/Strategy_game" title="Strategy game">strategy games</a> (e.g., <a href="/wiki/Chess" title="Chess">chess</a> and <a href="/wiki/Go_(game)" title="Go (game)">Go</a>).<sup class="reference" id="cite_ref-FOOTNOTEGoogle2016_2-0"><a href="#cite_note-FOOTNOTEGoogle2016-2">[2]</a></sup> However, many AI applications are not perceived as AI: "A lot of cutting edge AI has filtered into general applications, often without being called AI because once something becomes useful enough and common enough it's <a href="/wiki/AI_effect" title="AI effect">not labeled AI anymore</a>.
"<sup class="reference" id="cite_ref-3"><a href="#cite_note-3">[3]</a></sup><sup class="reference" id="cite_ref-andreas_4-0"><a href="#cite_note-andreas-4">[4]</a></sup>
</p><p><a href="/wiki/Alan_Turing" title="Alan Turing">Alan Turing</a> was the first person to conduct substantial research in the field that he called machine intelligence.<sup class="reference" id="cite_ref-turing_5-0"><a href="#cite_note-turing-5">[5]</a></sup> Artificial intelligence was founded as an academic discipline in 1956,<sup class="reference" id="cite_ref-Dartmouth_workshop_6-0"><a href="#cite_note-Dartmouth_workshop-6">[6]</a></sup> by those now considered the founding fathers of AI: <a href="/wiki/John_McCarthy_(computer_scientist)" title="John McCarthy (computer scientist)">John McCarthy</a>, <a href="/wiki/Marvin_Minsky" title="Marvin Minsky">Marvin Minksy</a>, <a href="/wiki/Nathaniel_Rochester_(computer_scientist)" title="Nathaniel Rochester (computer scientist)">Nathaniel Rochester</a>, and <a href="/wiki/Claude_Shannon" title="Claude Shannon">Claude Shannon</a>.<sup class="reference" id="cite_ref-7"><a href="#cite_note-7">[7]</a></sup><sup class="reference" id="cite_ref-8"><a href="#cite_note-8">[8]</a></sup> The field went through multiple cycles of optimism,<sup class="reference" id="cite_ref-AI_in_the_60s_9-0"><a href="#cite_note-AI_in_the_60s-9">[9]</a></sup><sup class="reference" id="cite_ref-AI_in_the_80s_10-0"><a href="#cite_note-AI_in_the_80s-10">[10]</a></sup> followed by periods of disappointment and loss of funding, known as <a href="/wiki/AI_winter" title="AI winter">AI winter</a>.<sup class="reference" id="cite_ref-First_AI_winter_11-0"><a href="#cite_note-First_AI_winter-11">[11]</a></sup><sup class="reference" id="cite_ref-Second_AI_winter_12-0"><a href="#cite_note-Second_AI_winter-12">[12]</a></sup> Funding and interest vastly increased after 2012 when <a href="/wiki/Deep_learning" title="Deep learning">deep learning</a> surpassed all previous AI techniques,<sup class="reference" id="cite_ref-Deep_learning_revolution_13-0"><a href="#cite_note-Deep_learning_revolution-13">[13]</a></sup> and after 2017 with the <a class="mw-redirect" href="/wiki/Transformer_(machine_learning_model)" title="Transformer (machine learning model)">transformer architecture</a>.<sup class="reference" id="cite_ref-FOOTNOTEToews2023_14-0"><a href="#cite_note-FOOTNOTEToews2023-14">[14]</a></sup> This led to the <a href="/wiki/AI_boom" title="AI boom">AI boom</a> of the early 2020s, with companies, universities, and laboratories overwhelmingly based in the United States pioneering significant <a class="mw-redirect" href="/wiki/Advances_in_artificial_intelligence" title="Advances in artificial intelligence">advances in artificial intelligence</a>.<sup class="reference" id="cite_ref-FOOTNOTEFrank2023_15-0"><a href="#cite_note-FOOTNOTEFrank2023-15">[15]</a></sup>
</p><p>The growing use of artificial intelligence in the 21st century is influencing a societal and economic shift towards increased <a href="/wiki/Automation" title="Automation">automation</a>, <a class="mw-redirect" href="/wiki/Data-driven_decision-making" title="Data-driven decision-making">data-driven decision-making</a>, and the <a href="/wiki/Artificial_intelligence_systems_integration" title="Artificial intelligence systems integration">integration of AI systems</a> into various economic sectors and areas of life, <a href="/wiki/Workplace_impact_of_artificial_intelligence" title="Workplace impact of artificial intelligence">impacting job markets</a>, <a href="/wiki/Artificial_intelligence_in_healthcare" title="Artificial intelligence in healthcare">healthcare</a>, <a href="/wiki/Artificial_intelligence_in_government" title="Artificial intelligence in government">government</a>, <a href="/wiki/Artificial_intelligence_in_industry" title="Artificial intelligence in industry">industry</a>, education, <a href="/wiki/Propaganda" title="Propaganda">propaganda</a>, and <a href="/wiki/Disinformation" title="Disinformation">disinformation</a>.
This raises questions about <a href="/wiki/AI_aftermath_scenarios" title="AI aftermath scenarios">the long-term effects</a>, <a href="/wiki/Ethics_of_artificial_intelligence" title="Ethics of artificial intelligence">ethical implications</a>, and <a class="mw-redirect" href="/wiki/AI_risk" title="AI risk">risks of AI</a>, prompting discussions about <a href="/wiki/Regulation_of_artificial_intelligence" title="Regulation of artificial intelligence">regulatory policies</a> to ensure the <a href="/wiki/AI_safety" title="AI safety">safety and benefits of the technology</a>.
</p><p>The various subfields of AI research are centered around particular goals and the use of particular tools.
The traditional goals of AI research include <a href="/wiki/Automated_reasoning" title="Automated reasoning">reasoning</a>, <a class="mw-redirect" href="/wiki/Knowledge_representation" title="Knowledge representation">knowledge representation</a>, <a href="/wiki/Automated_planning_and_scheduling" title="Automated planning and scheduling">planning</a>, <a href="/wiki/Machine_learning" title="Machine learning">learning</a>, <a href="/wiki/Natural_language_processing" title="Natural language processing">natural language processing</a>, perception, and support for <a href="/wiki/Robotics" title="Robotics">robotics</a>.<sup class="reference" id="cite_ref-Problems_of_AI_16-0"><a href="#cite_note-Problems_of_AI-16">[a]</a></sup> <a href="/wiki/Artificial_general_intelligence" title="Artificial general intelligence">General intelligence</a>—the ability to complete any task performable by a human on an at least equal level—is among the field's long-term goals.<sup class="reference" id="cite_ref-AGI_17-0"><a href="#cite_note-AGI-17">[16]</a></sup>
</p><p>To reach these goals, AI researchers have adapted and integrated a wide range of techniques, including <a href="/wiki/State_space_search" title="State space search">search</a> and <a href="/wiki/Mathematical_optimization" title="Mathematical optimization">mathematical optimization</a>, <a href="/wiki/Logic#Formal_logic" title="Logic">formal logic</a>, <a class="mw-redirect" href="/wiki/Artificial_neural_network" title="Artificial neural network">artificial neural networks</a>, and methods based on <a href="/wiki/Statistics" title="Statistics">statistics</a>, <a href="/wiki/Operations_research" title="Operations research">operations research</a>, and <a href="/wiki/Economics" title="Economics">economics</a>.<sup class="reference" id="cite_ref-Tools_of_AI_18-0"><a href="#cite_note-Tools_of_AI-18">[b]</a></sup> AI also draws upon <a href="/wiki/Psychology" title="Psychology">psychology</a>, <a href="/wiki/Linguistics" title="Linguistics">linguistics</a>, <a href="/wiki/Philosophy_of_artificial_intelligence" title="Philosophy of artificial intelligence">philosophy</a>, <a href="/wiki/Neuroscience" title="Neuroscience">neuroscience</a>, and other fields.<sup class="reference" id="cite_ref-AI_influences_19-0"><a href="#cite_note-AI_influences-19">[17]</a></sup>
</p><p>The general problem of simulating (or creating) intelligence has been broken into subproblems.
These consist of particular traits or capabilities that researchers expect an intelligent system to display.
The traits described below have received the most attention and cover the scope of AI research.<sup class="reference" id="cite_ref-Problems_of_AI_16-1"><a href="#cite_note-Problems_of_AI-16">[a]</a></sup>
</p><p>Early researchers developed algorithms that imitated step-by-step reasoning that humans use when they solve puzzles or make logical <a href="/wiki/Deductive_reasoning" title="Deductive reasoning">deductions</a>.<sup class="reference" id="cite_ref-20"><a href="#cite_note-20">[18]</a></sup> By the late 1980s and 1990s, methods were developed for dealing with <a href="/wiki/Uncertainty" title="Uncertainty">uncertain</a> or incomplete information, employing concepts from <a href="/wiki/Probability" title="Probability">probability</a> and <a href="/wiki/Economics" title="Economics">economics</a>.<sup class="reference" id="cite_ref-21"><a href="#cite_note-21">[19]</a></sup>
</p><p>Many of these algorithms are insufficient for solving large reasoning problems because they experience a "combinatorial explosion": They become exponentially slower as the problems grow.<sup class="reference" id="cite_ref-Intractability_22-0"><a href="#cite_note-Intractability-22">[20]</a></sup> Even humans rarely use the step-by-step deduction that early AI research could model.
They solve most of their problems using fast, intuitive judgments.<sup class="reference" id="cite_ref-Psychological_evidence_of_sub-symbolic_reasoning_23-0"><a href="#cite_note-Psychological_evidence_of_sub-symbolic_reasoning-23">[21]</a></sup> Accurate and efficient reasoning is an unsolved problem.
</p><p><a class="mw-redirect" href="/wiki/Knowledge_representation" title="Knowledge representation">Knowledge representation</a> and <a href="/wiki/Knowledge_engineering" title="Knowledge engineering">knowledge engineering</a><sup class="reference" id="cite_ref-24"><a href="#cite_note-24">[22]</a></sup> allow AI programs to answer questions intelligently and make deductions about real-world facts.
Formal knowledge representations are used in content-based indexing and retrieval,<sup class="reference" id="cite_ref-FOOTNOTESmoliarZhang1994_25-0"><a href="#cite_note-FOOTNOTESmoliarZhang1994-25">[23]</a></sup> scene interpretation,<sup class="reference" id="cite_ref-FOOTNOTENeumannMöller2008_26-0"><a href="#cite_note-FOOTNOTENeumannMöller2008-26">[24]</a></sup> clinical decision support,<sup class="reference" id="cite_ref-FOOTNOTEKupermanReichleyBailey2006_27-0"><a href="#cite_note-FOOTNOTEKupermanReichleyBailey2006-27">[25]</a></sup> knowledge discovery (mining "interesting" and actionable inferences from large <a href="/wiki/Database" title="Database">databases</a>),<sup class="reference" id="cite_ref-FOOTNOTEMcGarry2005_28-0"><a href="#cite_note-FOOTNOTEMcGarry2005-28">[26]</a></sup> and other areas.<sup class="reference" id="cite_ref-FOOTNOTEBertiniDel_BimboTorniai2006_29-0"><a href="#cite_note-FOOTNOTEBertiniDel_BimboTorniai2006-29">[27]</a></sup>
</p><p>A <a href="/wiki/Knowledge_base" title="Knowledge base">knowledge base</a> is a body of knowledge represented in a form that can be used by a program.
An <a href="/wiki/Ontology_(information_science)" title="Ontology (information science)">ontology</a> is the set of objects, relations, concepts, and properties used by a particular domain of knowledge.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021272_30-0"><a href="#cite_note-FOOTNOTERussellNorvig2021272-30">[28]</a></sup> Knowledge bases need to represent things such as objects, properties, categories, and relations between objects;<sup class="reference" id="cite_ref-Representing_categories_and_relations_31-0"><a href="#cite_note-Representing_categories_and_relations-31">[29]</a></sup> situations, events, states, and time;<sup class="reference" id="cite_ref-Representing_time_32-0"><a href="#cite_note-Representing_time-32">[30]</a></sup> causes and effects;<sup class="reference" id="cite_ref-Representing_causation_33-0"><a href="#cite_note-Representing_causation-33">[31]</a></sup> knowledge about knowledge (what we know about what other people know);<sup class="reference" id="cite_ref-Representing_knowledge_about_knowledge_34-0"><a href="#cite_note-Representing_knowledge_about_knowledge-34">[32]</a></sup> <a class="mw-redirect" href="/wiki/Default_reasoning" title="Default reasoning">default reasoning</a> (things that humans assume are true until they are told differently and will remain true even when other facts are changing);<sup class="reference" id="cite_ref-Default_reasoning_and_non-monotonic_logic_35-0"><a href="#cite_note-Default_reasoning_and_non-monotonic_logic-35">[33]</a></sup> and many other aspects and domains of knowledge.
</p><p>Among the most difficult problems in knowledge representation are the breadth of commonsense knowledge (the set of atomic facts that the average person knows is enormous);<sup class="reference" id="cite_ref-Breadth_of_commonsense_knowledge_36-0"><a href="#cite_note-Breadth_of_commonsense_knowledge-36">[34]</a></sup> and the sub-symbolic form of most commonsense knowledge (much of what people know is not represented as "facts" or "statements" that they could express verbally).<sup class="reference" id="cite_ref-Psychological_evidence_of_sub-symbolic_reasoning_23-1"><a href="#cite_note-Psychological_evidence_of_sub-symbolic_reasoning-23">[21]</a></sup> There is also the difficulty of <a href="/wiki/Knowledge_acquisition" title="Knowledge acquisition">knowledge acquisition</a>, the problem of obtaining knowledge for AI applications.<sup class="reference" id="cite_ref-39"><a href="#cite_note-39">[c]</a></sup>
</p><p>An "agent" is anything that perceives and takes actions in the world.
A <a href="/wiki/Rational_agent" title="Rational agent">rational agent</a> has goals or preferences and takes actions to make them happen.<sup class="reference" id="cite_ref-40"><a href="#cite_note-40">[d]</a></sup><sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021528_41-0"><a href="#cite_note-FOOTNOTERussellNorvig2021528-41">[37]</a></sup> In <a href="/wiki/Automated_planning_and_scheduling" title="Automated planning and scheduling">automated planning</a>, the agent has a specific goal.<sup class="reference" id="cite_ref-42"><a href="#cite_note-42">[38]</a></sup> In <a href="/wiki/Automated_decision-making" title="Automated decision-making">automated decision-making</a>, the agent has preferences—there are some situations it would prefer to be in, and some situations it is trying to avoid.
The decision-making agent assigns a number to each situation (called the "<a class="mw-redirect" href="/wiki/Utility_(economics)" title="Utility (economics)">utility</a>") that measures how much the agent prefers it.
For each possible action, it can calculate the "<a class="mw-redirect" href="/wiki/Expected_utility" title="Expected utility">expected utility</a>": the <a href="/wiki/Utility" title="Utility">utility</a> of all possible outcomes of the action, weighted by the probability that the outcome will occur.
It can then choose the action with the maximum expected utility.<sup class="reference" id="cite_ref-43"><a href="#cite_note-43">[39]</a></sup>
</p><p>In <a href="/wiki/Automated_planning_and_scheduling#classical_planning" title="Automated planning and scheduling">classical planning</a>, the agent knows exactly what the effect of any action will be.<sup class="reference" id="cite_ref-44"><a href="#cite_note-44">[40]</a></sup> In most real-world problems, however, the agent may not be certain about the situation they are in (it is "unknown" or "unobservable") and it may not know for certain what will happen after each possible action (it is not "deterministic").
It must choose an action by making a probabilistic guess and then reassess the situation to see if the action worked.<sup class="reference" id="cite_ref-45"><a href="#cite_note-45">[41]</a></sup>
</p><p>In some problems, the agent's preferences may be uncertain, especially if there are other agents or humans involved.
These can be learned (e.g., with <a class="mw-redirect" href="/wiki/Inverse_reinforcement_learning" title="Inverse reinforcement learning">inverse reinforcement learning</a>), or the agent can seek information to improve its preferences.<sup class="reference" id="cite_ref-46"><a href="#cite_note-46">[42]</a></sup> <a class="mw-redirect" href="/wiki/Information_value_theory" title="Information value theory">Information value theory</a> can be used to weigh the value of exploratory or experimental actions.<sup class="reference" id="cite_ref-47"><a href="#cite_note-47">[43]</a></sup> The space of possible future actions and situations is typically <a class="mw-redirect" href="/wiki/Intractable_problem" title="Intractable problem">intractably</a> large, so the agents must take actions and evaluate situations while being uncertain of what the outcome will be.
</p><p>A <a href="/wiki/Markov_decision_process" title="Markov decision process">Markov decision process</a> has a <a href="/wiki/Finite-state_machine" title="Finite-state machine">transition model</a> that describes the probability that a particular action will change the state in a particular way and a <a class="mw-redirect" href="/wiki/Reward_function" title="Reward function">reward function</a> that supplies the utility of each state and the cost of each action.
A <a href="/wiki/Reinforcement_learning#Policy" title="Reinforcement learning">policy</a> associates a decision with each possible state.
The policy could be calculated (e.g., by <a class="mw-redirect" href="/wiki/Policy_iteration" title="Policy iteration">iteration</a>), be <a href="/wiki/Heuristic" title="Heuristic">heuristic</a>, or it can be learned.<sup class="reference" id="cite_ref-48"><a href="#cite_note-48">[44]</a></sup>
</p><p><a href="/wiki/Game_theory" title="Game theory">Game theory</a> describes the rational behavior of multiple interacting agents and is used in AI programs that make decisions that involve other agents.<sup class="reference" id="cite_ref-49"><a href="#cite_note-49">[45]</a></sup>
</p><p><a href="/wiki/Machine_learning" title="Machine learning">Machine learning</a> is the study of programs that can improve their performance on a given task automatically.<sup class="reference" id="cite_ref-machine_learning_50-0"><a href="#cite_note-machine_learning-50">[46]</a></sup> It has been a part of AI from the beginning.<sup class="reference" id="cite_ref-53"><a href="#cite_note-53">[e]</a></sup>
</p><p>There are several kinds of machine learning.
<a href="/wiki/Unsupervised_learning" title="Unsupervised learning">Unsupervised learning</a> analyzes a stream of data and finds patterns and makes predictions without any other guidance.<sup class="reference" id="cite_ref-54"><a href="#cite_note-54">[49]</a></sup> <a href="/wiki/Supervised_learning" title="Supervised learning">Supervised learning</a> requires a human to label the input data first, and comes in two main varieties: <a href="/wiki/Statistical_classification" title="Statistical classification">classification</a> (where the program must learn to predict what category the input belongs in) and <a href="/wiki/Regression_analysis" title="Regression analysis">regression</a> (where the program must deduce a numeric function based on numeric input).<sup class="reference" id="cite_ref-Supervised_learning_55-0"><a href="#cite_note-Supervised_learning-55">[50]</a></sup>
</p><p>In <a href="/wiki/Reinforcement_learning" title="Reinforcement learning">reinforcement learning</a>, the agent is rewarded for good responses and punished for bad ones.
The agent learns to choose responses that are classified as "good".<sup class="reference" id="cite_ref-56"><a href="#cite_note-56">[51]</a></sup> <a href="/wiki/Transfer_learning" title="Transfer learning">Transfer learning</a> is when the knowledge gained from one problem is applied to a new problem.<sup class="reference" id="cite_ref-57"><a href="#cite_note-57">[52]</a></sup> <a href="/wiki/Deep_learning" title="Deep learning">Deep learning</a> is a type of machine learning that runs inputs through biologically inspired <a class="mw-redirect" href="/wiki/Artificial_neural_networks" title="Artificial neural networks">artificial neural networks</a> for all of these types of learning.<sup class="reference" id="cite_ref-58"><a href="#cite_note-58">[53]</a></sup>
</p><p><a href="/wiki/Computational_learning_theory" title="Computational learning theory">Computational learning theory</a> can assess learners by <a href="/wiki/Computational_complexity" title="Computational complexity">computational complexity</a>, by <a href="/wiki/Sample_complexity" title="Sample complexity">sample complexity</a> (how much data is required), or by other notions of <a class="mw-redirect" href="/wiki/Optimization_theory" title="Optimization theory">optimization</a>.<sup class="reference" id="cite_ref-59"><a href="#cite_note-59">[54]</a></sup>
</p><p><a href="/wiki/Natural_language_processing" title="Natural language processing">Natural language processing</a> (NLP)<sup class="reference" id="cite_ref-60"><a href="#cite_note-60">[55]</a></sup> allows programs to read, write and communicate in human languages such as <a class="mw-redirect" href="/wiki/English_(language)" title="English (language)">English</a>.
Specific problems include <a href="/wiki/Speech_recognition" title="Speech recognition">speech recognition</a>, <a href="/wiki/Speech_synthesis" title="Speech synthesis">speech synthesis</a>, <a href="/wiki/Machine_translation" title="Machine translation">machine translation</a>, <a href="/wiki/Information_extraction" title="Information extraction">information extraction</a>, <a href="/wiki/Information_retrieval" title="Information retrieval">information retrieval</a> and <a href="/wiki/Question_answering" title="Question answering">question answering</a>.<sup class="reference" id="cite_ref-61"><a href="#cite_note-61">[56]</a></sup>
</p><p>Early work, based on <a href="/wiki/Noam_Chomsky" title="Noam Chomsky">Noam Chomsky</a>'s <a href="/wiki/Generative_grammar" title="Generative grammar">generative grammar</a> and <a href="/wiki/Semantic_network" title="Semantic network">semantic networks</a>, had difficulty with <a href="/wiki/Word-sense_disambiguation" title="Word-sense disambiguation">word-sense disambiguation</a><sup class="reference" id="cite_ref-62"><a href="#cite_note-62">[f]</a></sup> unless restricted to small domains called "<a href="/wiki/Blocks_world" title="Blocks world">micro-worlds</a>" (due to the common sense knowledge problem<sup class="reference" id="cite_ref-Breadth_of_commonsense_knowledge_36-1"><a href="#cite_note-Breadth_of_commonsense_knowledge-36">[34]</a></sup>).
<a href="/wiki/Margaret_Masterman" title="Margaret Masterman">Margaret Masterman</a> believed that it was meaning and not grammar that was the key to understanding languages, and that <a class="mw-redirect" href="/wiki/Thesauri" title="Thesauri">thesauri</a> and not dictionaries should be the basis of computational language structure.
</p><p>Modern deep learning techniques for NLP include <a href="/wiki/Word_embedding" title="Word embedding">word embedding</a> (representing words, typically as <a href="/wiki/Vector_space" title="Vector space">vectors</a> encoding their meaning),<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021856–858_63-0"><a href="#cite_note-FOOTNOTERussellNorvig2021856–858-63">[57]</a></sup> <a class="mw-redirect" href="/wiki/Transformer_(machine_learning_model)" title="Transformer (machine learning model)">transformers</a> (a deep learning architecture using an <a href="/wiki/Attention_(machine_learning)" title="Attention (machine learning)">attention</a> mechanism),<sup class="reference" id="cite_ref-FOOTNOTEDickson2022_64-0"><a href="#cite_note-FOOTNOTEDickson2022-64">[58]</a></sup> and others.<sup class="reference" id="cite_ref-65"><a href="#cite_note-65">[59]</a></sup> In 2019, <a href="/wiki/Generative_pre-trained_transformer" title="Generative pre-trained transformer">generative pre-trained transformer</a> (or "GPT") language models began to generate coherent text,<sup class="reference" id="cite_ref-FOOTNOTEVincent2019_66-0"><a href="#cite_note-FOOTNOTEVincent2019-66">[60]</a></sup><sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021875–878_67-0"><a href="#cite_note-FOOTNOTERussellNorvig2021875–878-67">[61]</a></sup> and by 2023, these models were able to get human-level scores on the <a class="mw-redirect" href="/wiki/Bar_exam" title="Bar exam">bar exam</a>, <a class="mw-redirect" href="/wiki/Scholastic_aptitude_test" title="Scholastic aptitude test">SAT</a> test, <a href="/wiki/Graduate_Record_Examinations" title="Graduate Record Examinations">GRE</a> test, and many other real-world applications.<sup class="reference" id="cite_ref-FOOTNOTEBushwick2023_68-0"><a href="#cite_note-FOOTNOTEBushwick2023-68">[62]</a></sup>
</p><p><a href="/wiki/Machine_perception" title="Machine perception">Machine perception</a> is the ability to use input from sensors (such as cameras, microphones, wireless signals, active <a href="/wiki/Lidar" title="Lidar">lidar</a>, sonar, radar, and <a href="/wiki/Tactile_sensor" title="Tactile sensor">tactile sensors</a>) to deduce aspects of the world.
<a href="/wiki/Computer_vision" title="Computer vision">Computer vision</a> is the ability to analyze visual input.<sup class="reference" id="cite_ref-69"><a href="#cite_note-69">[63]</a></sup>
</p><p>The field includes <a href="/wiki/Speech_recognition" title="Speech recognition">speech recognition</a>,<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021849–850_70-0"><a href="#cite_note-FOOTNOTERussellNorvig2021849–850-70">[64]</a></sup> <a class="mw-redirect" href="/wiki/Image_classification" title="Image classification">image classification</a>,<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021895–899_71-0"><a href="#cite_note-FOOTNOTERussellNorvig2021895–899-71">[65]</a></sup> <a href="/wiki/Facial_recognition_system" title="Facial recognition system">facial recognition</a>, <a class="mw-redirect" href="/wiki/Object_recognition" title="Object recognition">object recognition</a>,<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021899–901_72-0"><a href="#cite_note-FOOTNOTERussellNorvig2021899–901-72">[66]</a></sup><a class="mw-redirect mw-disambig" href="/wiki/Object_tracking" title="Object tracking">object tracking</a>,<sup class="reference" id="cite_ref-FOOTNOTEChallaMorelandMušickiEvans2011_73-0"><a href="#cite_note-FOOTNOTEChallaMorelandMušickiEvans2011-73">[67]</a></sup> and <a href="/wiki/Robotic_sensing" title="Robotic sensing">robotic perception</a>.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021931–938_74-0"><a href="#cite_note-FOOTNOTERussellNorvig2021931–938-74">[68]</a></sup>
</p><p><a href="/wiki/Affective_computing" title="Affective computing">Affective computing</a> is an interdisciplinary umbrella that comprises systems that recognize, interpret, process, or simulate human <a href="/wiki/Affect_(psychology)" title="Affect (psychology)">feeling, emotion, and mood</a>.<sup class="reference" id="cite_ref-76"><a href="#cite_note-76">[70]</a></sup> For example, some <a href="/wiki/Virtual_assistant" title="Virtual assistant">virtual assistants</a> are programmed to speak conversationally or even to banter humorously; it makes them appear more sensitive to the emotional dynamics of human interaction, or to otherwise facilitate <a href="/wiki/Human%E2%80%93computer_interaction" title="Human–computer interaction">human–computer interaction</a>.
</p><p>However, this tends to give naïve users an unrealistic conception of the intelligence of existing computer agents.<sup class="reference" id="cite_ref-FOOTNOTEWaddell2018_77-0"><a href="#cite_note-FOOTNOTEWaddell2018-77">[71]</a></sup> Moderate successes related to affective computing include textual <a href="/wiki/Sentiment_analysis" title="Sentiment analysis">sentiment analysis</a> and, more recently, <a href="/wiki/Multimodal_sentiment_analysis" title="Multimodal sentiment analysis">multimodal sentiment analysis</a>, wherein AI classifies the affects displayed by a videotaped subject.<sup class="reference" id="cite_ref-FOOTNOTEPoriaCambriaBajpaiHussain2017_78-0"><a href="#cite_note-FOOTNOTEPoriaCambriaBajpaiHussain2017-78">[72]</a></sup>
</p><p>A machine with <a href="/wiki/Artificial_general_intelligence" title="Artificial general intelligence">artificial general intelligence</a> should be able to solve a wide variety of problems with breadth and versatility similar to <a href="/wiki/Human_intelligence" title="Human intelligence">human intelligence</a>.<sup class="reference" id="cite_ref-AGI_17-1"><a href="#cite_note-AGI-17">[16]</a></sup>
</p><p>AI research uses a wide variety of techniques to accomplish the goals above.<sup class="reference" id="cite_ref-Tools_of_AI_18-1"><a href="#cite_note-Tools_of_AI-18">[b]</a></sup>
</p><p>AI can solve many problems by intelligently searching through many possible solutions.<sup class="reference" id="cite_ref-79"><a href="#cite_note-79">[73]</a></sup> There are two very different kinds of search used in AI: <a href="/wiki/State_space_search" title="State space search">state space search</a> and <a href="/wiki/Local_search_(optimization)" title="Local search (optimization)">local search</a>.
</p><p><a href="/wiki/State_space_search" title="State space search">State space search</a> searches through a tree of possible states to try to find a goal state.<sup class="reference" id="cite_ref-State_space_search_80-0"><a href="#cite_note-State_space_search-80">[74]</a></sup> For example, <a href="/wiki/Automated_planning_and_scheduling" title="Automated planning and scheduling">planning</a> algorithms search through trees of goals and subgoals, attempting to find a path to a target goal, a process called <a class="mw-redirect" href="/wiki/Means-ends_analysis" title="Means-ends analysis">means-ends analysis</a>.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021§11.2_81-0"><a href="#cite_note-FOOTNOTERussellNorvig2021§11.2-81">[75]</a></sup>
</p><p><a class="mw-redirect" href="/wiki/Brute_force_search" title="Brute force search">Simple exhaustive searches</a><sup class="reference" id="cite_ref-Uninformed_search_82-0"><a href="#cite_note-Uninformed_search-82">[76]</a></sup> are rarely sufficient for most real-world problems: the <a href="/wiki/Search_algorithm" title="Search algorithm">search space</a> (the number of places to search) quickly grows to <a class="mw-redirect" href="/wiki/Astronomically_large" title="Astronomically large">astronomical numbers</a>.
The result is a search that is <a class="mw-redirect" href="/wiki/Computation_time" title="Computation time">too slow</a> or never completes.<sup class="reference" id="cite_ref-Intractability_22-1"><a href="#cite_note-Intractability-22">[20]</a></sup> "<a class="mw-redirect" href="/wiki/Heuristics" title="Heuristics">Heuristics</a>" or "rules of thumb" can help prioritize choices that are more likely to reach a goal.<sup class="reference" id="cite_ref-Informed_search_83-0"><a href="#cite_note-Informed_search-83">[77]</a></sup>
</p><p><a class="mw-redirect" href="/wiki/Adversarial_search" title="Adversarial search">Adversarial search</a> is used for <a class="mw-redirect" href="/wiki/Game_AI" title="Game AI">game-playing</a> programs, such as chess or Go.
It searches through a <a href="/wiki/Game_tree" title="Game tree">tree</a> of possible moves and counter-moves, looking for a winning position.<sup class="reference" id="cite_ref-84"><a href="#cite_note-84">[78]</a></sup>
</p><p><a href="/wiki/Local_search_(optimization)" title="Local search (optimization)">Local search</a> uses <a href="/wiki/Mathematical_optimization" title="Mathematical optimization">mathematical optimization</a> to find a solution to a problem.
It begins with some form of guess and refines it incrementally.<sup class="reference" id="cite_ref-Local_search2_85-0"><a href="#cite_note-Local_search2-85">[79]</a></sup>
</p><p><a href="/wiki/Gradient_descent" title="Gradient descent">Gradient descent</a> is a type of local search that optimizes a set of numerical parameters by incrementally adjusting them to minimize a <a href="/wiki/Loss_function" title="Loss function">loss function</a>.
Variants of <a href="/wiki/Gradient_descent" title="Gradient descent">gradient descent</a> are commonly used to train neural networks.<sup class="reference" id="cite_ref-86"><a href="#cite_note-86">[80]</a></sup>
</p><p>Another type of local search is <a href="/wiki/Evolutionary_computation" title="Evolutionary computation">evolutionary computation</a>, which aims to iteratively improve a set of candidate solutions by "mutating" and "recombining" them, <a class="mw-redirect" href="/wiki/Artificial_selection" title="Artificial selection">selecting</a> only the fittest to survive each generation.<sup class="reference" id="cite_ref-87"><a href="#cite_note-87">[81]</a></sup>
</p><p>Distributed search processes can coordinate via <a href="/wiki/Swarm_intelligence" title="Swarm intelligence">swarm intelligence</a> algorithms.
Two popular swarm algorithms used in search are <a href="/wiki/Particle_swarm_optimization" title="Particle swarm optimization">particle swarm optimization</a> (inspired by bird <a class="mw-redirect" href="/wiki/Flocking_(behavior)" title="Flocking (behavior)">flocking</a>) and <a class="mw-redirect" href="/wiki/Ant_colony_optimization" title="Ant colony optimization">ant colony optimization</a> (inspired by <a class="mw-redirect" href="/wiki/Ant_trail" title="Ant trail">ant trails</a>).<sup class="reference" id="cite_ref-FOOTNOTEMerkleMiddendorf2013_88-0"><a href="#cite_note-FOOTNOTEMerkleMiddendorf2013-88">[82]</a></sup>
</p><p>Formal <a href="/wiki/Logic" title="Logic">logic</a> is used for <a class="mw-redirect" href="/wiki/Automatic_reasoning" title="Automatic reasoning">reasoning</a> and <a class="mw-redirect" href="/wiki/Knowledge_representation" title="Knowledge representation">knowledge representation</a>.<sup class="reference" id="cite_ref-Logic_89-0"><a href="#cite_note-Logic-89">[83]</a></sup>
Formal logic comes in two main forms: <a class="mw-redirect" href="/wiki/Propositional_logic" title="Propositional logic">propositional logic</a> (which operates on statements that are true or false and uses <a href="/wiki/Logical_connective" title="Logical connective">logical connectives</a> such as "and", "or", "not" and "implies")<sup class="reference" id="cite_ref-Propositional_logic_90-0"><a href="#cite_note-Propositional_logic-90">[84]</a></sup> and <a class="mw-redirect" href="/wiki/Predicate_logic" title="Predicate logic">predicate logic</a> (which also operates on objects, predicates and relations and uses <a href="/wiki/Quantifier_(logic)" title="Quantifier (logic)">quantifiers</a> such as "<i>Every</i> <i>X</i> is a <i>Y</i>" and "There are <i>some</i> <i>X</i>s that are <i>Y</i>s").<sup class="reference" id="cite_ref-Predicate_logic_91-0"><a href="#cite_note-Predicate_logic-91">[85]</a></sup>
</p><p><a href="/wiki/Deductive_reasoning" title="Deductive reasoning">Deductive reasoning</a> in logic is the process of <a class="mw-redirect" href="/wiki/Logical_proof" title="Logical proof">proving</a> a new statement (<a href="/wiki/Logical_consequence" title="Logical consequence">conclusion</a>) from other statements that are given and assumed to be true (the <a href="/wiki/Premise" title="Premise">premises</a>).<sup class="reference" id="cite_ref-Inference_92-0"><a href="#cite_note-Inference-92">[86]</a></sup> Proofs can be structured as proof <a href="/wiki/Tree_structure" title="Tree structure">trees</a>, in which nodes are labelled by sentences, and children nodes are connected to parent nodes by <a class="mw-redirect" href="/wiki/Inference_rule" title="Inference rule">inference rules</a>.
</p><p>Given a problem and a set of premises, problem-solving reduces to searching for a proof tree whose root node is labelled by a solution of the problem and whose <a class="mw-redirect" href="/wiki/Leaf_nodes" title="Leaf nodes">leaf nodes</a> are labelled by premises or <a href="/wiki/Axiom" title="Axiom">axioms</a>.
In the case of <a href="/wiki/Horn_clause" title="Horn clause">Horn clauses</a>, problem-solving search can be performed by reasoning <a href="/wiki/Forward_chaining" title="Forward chaining">forwards</a> from the premises or <a href="/wiki/Backward_chaining" title="Backward chaining">backwards</a> from the problem.<sup class="reference" id="cite_ref-Logic_as_search_93-0"><a href="#cite_note-Logic_as_search-93">[87]</a></sup> In the more general case of the clausal form of <a href="/wiki/First-order_logic" title="First-order logic">first-order logic</a>, <a href="/wiki/Resolution_(logic)" title="Resolution (logic)">resolution</a> is a single, axiom-free rule of inference, in which a problem is solved by proving a contradiction from premises that include the negation of the problem to be solved.<sup class="reference" id="cite_ref-Resolution_94-0"><a href="#cite_note-Resolution-94">[88]</a></sup>
</p><p>Inference in both Horn clause logic and first-order logic is <a href="/wiki/Undecidable_problem" title="Undecidable problem">undecidable</a>, and therefore <a class="mw-redirect" href="/wiki/Intractable_problem" title="Intractable problem">intractable</a>.
However, backward reasoning with Horn clauses, which underpins computation in the <a href="/wiki/Logic_programming" title="Logic programming">logic programming</a> language <a href="/wiki/Prolog" title="Prolog">Prolog</a>, is <a href="/wiki/Turing_completeness" title="Turing completeness">Turing complete</a>.
Moreover, its efficiency is competitive with computation in other <a href="/wiki/Symbolic_programming" title="Symbolic programming">symbolic programming</a> languages.<sup class="reference" id="cite_ref-95"><a href="#cite_note-95">[89]</a></sup>
</p><p><a href="/wiki/Fuzzy_logic" title="Fuzzy logic">Fuzzy logic</a> assigns a "degree of truth" between 0 and 1.
It can therefore handle propositions that are vague and partially true.<sup class="reference" id="cite_ref-Fuzzy_logic_96-0"><a href="#cite_note-Fuzzy_logic-96">[90]</a></sup>
</p><p><a href="/wiki/Non-monotonic_logic" title="Non-monotonic logic">Non-monotonic logics</a>, including logic programming with <a href="/wiki/Negation_as_failure" title="Negation as failure">negation as failure</a>, are designed to handle <a class="mw-redirect" href="/wiki/Default_reasoning" title="Default reasoning">default reasoning</a>.<sup class="reference" id="cite_ref-Default_reasoning_and_non-monotonic_logic_35-1"><a href="#cite_note-Default_reasoning_and_non-monotonic_logic-35">[33]</a></sup>
Other specialized versions of logic have been developed to describe many complex domains.
</p><p>Many problems in AI (including in reasoning, planning, learning, perception, and robotics) require the agent to operate with incomplete or uncertain information.
AI researchers have devised a number of tools to solve these problems using methods from <a href="/wiki/Probability" title="Probability">probability</a> theory and economics.<sup class="reference" id="cite_ref-Uncertain_reasoning_97-0"><a href="#cite_note-Uncertain_reasoning-97">[91]</a></sup> Precise mathematical tools have been developed that analyze how an agent can make choices and plan, using <a href="/wiki/Decision_theory" title="Decision theory">decision theory</a>, <a href="/wiki/Decision_analysis" title="Decision analysis">decision analysis</a>,<sup class="reference" id="cite_ref-Decisions_theory_and_analysis_98-0"><a href="#cite_note-Decisions_theory_and_analysis-98">[92]</a></sup> and <a class="mw-redirect" href="/wiki/Information_value_theory" title="Information value theory">information value theory</a>.<sup class="reference" id="cite_ref-Information_value_theory_99-0"><a href="#cite_note-Information_value_theory-99">[93]</a></sup> These tools include models such as <a href="/wiki/Markov_decision_process" title="Markov decision process">Markov decision processes</a>,<sup class="reference" id="cite_ref-Markov_decision_process_100-0"><a href="#cite_note-Markov_decision_process-100">[94]</a></sup> dynamic <a class="mw-redirect" href="/wiki/Decision_network" title="Decision network">decision networks</a>,<sup class="reference" id="cite_ref-Stochastic_temporal_models_101-0"><a href="#cite_note-Stochastic_temporal_models-101">[95]</a></sup> <a href="/wiki/Game_theory" title="Game theory">game theory</a> and <a href="/wiki/Mechanism_design" title="Mechanism design">mechanism design</a>.<sup class="reference" id="cite_ref-Game_theory_and_mechanism_design_102-0"><a href="#cite_note-Game_theory_and_mechanism_design-102">[96]</a></sup>
</p><p><a href="/wiki/Bayesian_network" title="Bayesian network">Bayesian networks</a><sup class="reference" id="cite_ref-Bayesian_networks_103-0"><a href="#cite_note-Bayesian_networks-103">[97]</a></sup> are a tool that can be used for <a href="/wiki/Automated_reasoning" title="Automated reasoning">reasoning</a> (using the <a href="/wiki/Bayesian_inference" title="Bayesian inference">Bayesian inference</a> algorithm),<sup class="reference" id="cite_ref-105"><a href="#cite_note-105">[g]</a></sup><sup class="reference" id="cite_ref-Bayesian_inference_106-0"><a href="#cite_note-Bayesian_inference-106">[99]</a></sup> <a href="/wiki/Machine_learning" title="Machine learning">learning</a> (using the <a href="/wiki/Expectation%E2%80%93maximization_algorithm" title="Expectation–maximization algorithm">expectation–maximization algorithm</a>),<sup class="reference" id="cite_ref-108"><a href="#cite_note-108">[h]</a></sup><sup class="reference" id="cite_ref-Bayesian_learning_109-0"><a href="#cite_note-Bayesian_learning-109">[101]</a></sup> <a href="/wiki/Automated_planning_and_scheduling" title="Automated planning and scheduling">planning</a> (using <a class="mw-redirect" href="/wiki/Decision_network" title="Decision network">decision networks</a>)<sup class="reference" id="cite_ref-Bayesian_decision_networks_110-0"><a href="#cite_note-Bayesian_decision_networks-110">[102]</a></sup> and <a href="/wiki/Machine_perception" title="Machine perception">perception</a> (using <a href="/wiki/Dynamic_Bayesian_network" title="Dynamic Bayesian network">dynamic Bayesian networks</a>).<sup class="reference" id="cite_ref-Stochastic_temporal_models_101-1"><a href="#cite_note-Stochastic_temporal_models-101">[95]</a></sup>
</p><p>Probabilistic algorithms can also be used for filtering, prediction, smoothing, and finding explanations for streams of data, thus helping <a href="/wiki/Machine_perception" title="Machine perception">perception</a> systems analyze processes that occur over time (e.g., <a href="/wiki/Hidden_Markov_model" title="Hidden Markov model">hidden Markov models</a> or <a href="/wiki/Kalman_filter" title="Kalman filter">Kalman filters</a>).<sup class="reference" id="cite_ref-Stochastic_temporal_models_101-2"><a href="#cite_note-Stochastic_temporal_models-101">[95]</a></sup>
</p><p>The simplest AI applications can be divided into two types: classifiers (e.g., "if shiny then diamond"), on one hand, and controllers (e.g., "if diamond then pick up"), on the other hand.
<a class="mw-redirect" href="/wiki/Classifier_(mathematics)" title="Classifier (mathematics)">Classifiers</a><sup class="reference" id="cite_ref-Statistical_classifiers_111-0"><a href="#cite_note-Statistical_classifiers-111">[103]</a></sup> are functions that use <a href="/wiki/Pattern_matching" title="Pattern matching">pattern matching</a> to determine the closest match.
They can be fine-tuned based on chosen examples using <a href="/wiki/Supervised_learning" title="Supervised learning">supervised learning</a>.
Each pattern (also called an "<a href="/wiki/Random_variate" title="Random variate">observation</a>") is labeled with a certain predefined class.
All the observations combined with their class labels are known as a <a href="/wiki/Data_set" title="Data set">data set</a>.
When a new observation is received, that observation is classified based on previous experience.<sup class="reference" id="cite_ref-Supervised_learning_55-1"><a href="#cite_note-Supervised_learning-55">[50]</a></sup>
</p><p>There are many kinds of classifiers in use.
The <a href="/wiki/Decision_tree" title="Decision tree">decision tree</a> is the simplest and most widely used symbolic machine learning algorithm.<sup class="reference" id="cite_ref-112"><a href="#cite_note-112">[104]</a></sup> <a class="mw-redirect" href="/wiki/K-nearest_neighbor" title="K-nearest neighbor">K-nearest neighbor</a> algorithm was the most widely used analogical AI until the mid-1990s, and <a class="mw-redirect" href="/wiki/Kernel_methods" title="Kernel methods">Kernel methods</a> such as the <a href="/wiki/Support_vector_machine" title="Support vector machine">support vector machine</a> (SVM) displaced k-nearest neighbor in the 1990s.<sup class="reference" id="cite_ref-113"><a href="#cite_note-113">[105]</a></sup>
The <a href="/wiki/Naive_Bayes_classifier" title="Naive Bayes classifier">naive Bayes classifier</a> is reportedly the "most widely used learner"<sup class="reference" id="cite_ref-FOOTNOTEDomingos2015152_114-0"><a href="#cite_note-FOOTNOTEDomingos2015152-114">[106]</a></sup> at Google, due in part to its scalability.<sup class="reference" id="cite_ref-115"><a href="#cite_note-115">[107]</a></sup>
<a class="mw-redirect" href="/wiki/Artificial_neural_network" title="Artificial neural network">Neural networks</a> are also used as classifiers.<sup class="reference" id="cite_ref-Neural_networks_116-0"><a href="#cite_note-Neural_networks-116">[108]</a></sup>
</p><p>An artificial neural network is based on a collection of nodes also known as <a class="mw-redirect" href="/wiki/Artificial_neurons" title="Artificial neurons">artificial neurons</a>, which loosely model the <a class="mw-redirect" href="/wiki/Neurons" title="Neurons">neurons</a> in a biological brain.
It is trained to recognise patterns; once trained, it can recognise those patterns in fresh data.
There is an input, at least one hidden layer of nodes and an output.
Each node applies a function and once the <a href="/wiki/Weighting" title="Weighting">weight</a> crosses its specified threshold, the data is transmitted to the next layer.
A network is typically called a deep neural network if it has at least 2 hidden layers.<sup class="reference" id="cite_ref-Neural_networks_116-1"><a href="#cite_note-Neural_networks-116">[108]</a></sup>
</p><p>Learning algorithms for neural networks use <a href="/wiki/Local_search_(optimization)" title="Local search (optimization)">local search</a> to choose the weights that will get the right output for each input during training.
The most common training technique is the <a href="/wiki/Backpropagation" title="Backpropagation">backpropagation</a> algorithm.<sup class="reference" id="cite_ref-Backpropagation_117-0"><a href="#cite_note-Backpropagation-117">[109]</a></sup>
Neural networks learn to model complex relationships between inputs and outputs and <a href="/wiki/Pattern_recognition" title="Pattern recognition">find patterns</a> in data.
In theory, a neural network can learn any function.<sup class="reference" id="cite_ref-118"><a href="#cite_note-118">[110]</a></sup>
</p><p>In <a href="/wiki/Feedforward_neural_network" title="Feedforward neural network">feedforward neural networks</a> the signal passes in only one direction.<sup class="reference" id="cite_ref-119"><a href="#cite_note-119">[111]</a></sup> <a href="/wiki/Recurrent_neural_network" title="Recurrent neural network">Recurrent neural networks</a> feed the output signal back into the input, which allows short-term memories of previous input events.
<a class="mw-redirect" href="/wiki/Long_short_term_memory" title="Long short term memory">Long short term memory</a> is the most successful network architecture for recurrent networks.<sup class="reference" id="cite_ref-120"><a href="#cite_note-120">[112]</a></sup>
<a href="/wiki/Perceptron" title="Perceptron">Perceptrons</a><sup class="reference" id="cite_ref-121"><a href="#cite_note-121">[113]</a></sup>
use only a single layer of neurons, deep learning<sup class="reference" id="cite_ref-Deep_learning_122-0"><a href="#cite_note-Deep_learning-122">[114]</a></sup> uses multiple layers.
<a href="/wiki/Convolutional_neural_network" title="Convolutional neural network">Convolutional neural networks</a> strengthen the connection between neurons that are "close" to each other—this is especially important in <a class="mw-redirect" href="/wiki/Image_processing" title="Image processing">image processing</a>, where a local set of neurons must <a href="/wiki/Edge_detection" title="Edge detection">identify an "edge"</a> before the network can identify an object.<sup class="reference" id="cite_ref-123"><a href="#cite_note-123">[115]</a></sup>
</p><p>Deep learning<sup class="reference" id="cite_ref-Deep_learning_122-1"><a href="#cite_note-Deep_learning-122">[114]</a></sup>
uses several layers of neurons between the network's inputs and outputs.
The multiple layers can progressively extract higher-level features from the raw input.
For example, in <a class="mw-redirect" href="/wiki/Image_processing" title="Image processing">image processing</a>, lower layers may identify edges, while higher layers may identify the concepts relevant to a human such as digits, letters, or faces.<sup class="reference" id="cite_ref-FOOTNOTEDengYu2014199–200_124-0"><a href="#cite_note-FOOTNOTEDengYu2014199–200-124">[116]</a></sup>
</p><p>Deep learning has profoundly improved the performance of programs in many important subfields of artificial intelligence, including <a href="/wiki/Computer_vision" title="Computer vision">computer vision</a>, <a href="/wiki/Speech_recognition" title="Speech recognition">speech recognition</a>, <a href="/wiki/Natural_language_processing" title="Natural language processing">natural language processing</a>, <a class="mw-redirect" href="/wiki/Image_classification" title="Image classification">image classification</a>,<sup class="reference" id="cite_ref-FOOTNOTECiresanMeierSchmidhuber2012_125-0"><a href="#cite_note-FOOTNOTECiresanMeierSchmidhuber2012-125">[117]</a></sup> and others.
The reason that deep learning performs so well in so many applications is not known as of 2023.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021751_126-0"><a href="#cite_note-FOOTNOTERussellNorvig2021751-126">[118]</a></sup>
The sudden success of deep learning in 2012–2015 did not occur because of some new discovery or theoretical breakthrough (deep neural networks and <a href="/wiki/Backpropagation" title="Backpropagation">backpropagation</a> had been described by many people, as far back as the 1950s)<sup class="reference" id="cite_ref-134"><a href="#cite_note-134">[i]</a></sup>
but because of two factors: the incredible increase in computer power (including the hundred-fold increase in speed by switching to <a class="mw-redirect" href="/wiki/Graphics_processing_units" title="Graphics processing units">GPUs</a>) and the availability of vast amounts of training data, especially the giant <a href="/wiki/List_of_datasets_for_machine-learning_research" title="List of datasets for machine-learning research">curated datasets</a> used for benchmark testing, such as <a href="/wiki/ImageNet" title="ImageNet">ImageNet</a>.<sup class="reference" id="cite_ref-136"><a href="#cite_note-136">[j]</a></sup>
</p><p><a href="/wiki/Generative_pre-trained_transformer" title="Generative pre-trained transformer">Generative pre-trained transformers</a> (GPT) are <a href="/wiki/Large_language_model" title="Large language model">large language models</a> that are based on the semantic relationships between words in sentences (<a href="/wiki/Natural_language_processing" title="Natural language processing">natural language processing</a>).
Text-based GPT models are pretrained on a large <a class="mw-redirect" href="/wiki/Corpus_of_text" title="Corpus of text">corpus of text</a> that can be from the Internet.
The pretraining consists of predicting the next <a href="/wiki/Lexical_analysis" title="Lexical analysis">token</a> (a token being usually a word, subword, or punctuation).
Throughout this pretraining, GPT models accumulate knowledge about the world and can then generate human-like text by repeatedly predicting the next token.
Typically, a subsequent training phase makes the model more truthful, useful, and harmless, usually with a technique called <a href="/wiki/Reinforcement_learning_from_human_feedback" title="Reinforcement learning from human feedback">reinforcement learning from human feedback</a> (RLHF).
Current GPT models are prone to generating falsehoods called "<a href="/wiki/Hallucination_(artificial_intelligence)" title="Hallucination (artificial intelligence)">hallucinations</a>", although this can be reduced with RLHF and quality data.
They are used in <a href="/wiki/Chatbot" title="Chatbot">chatbots</a>, which allow people to ask a question or request a task in simple text.<sup class="reference" id="cite_ref-FOOTNOTESmith2023_137-0"><a href="#cite_note-FOOTNOTESmith2023-137">[127]</a></sup><sup class="reference" id="cite_ref-138"><a href="#cite_note-138">[128]</a></sup>
</p><p>Current models and services include <a href="/wiki/Gemini_(chatbot)" title="Gemini (chatbot)">Gemini</a> (formerly Bard), <a href="/wiki/ChatGPT" title="ChatGPT">ChatGPT</a>, <a href="/wiki/Grok_(chatbot)" title="Grok (chatbot)">Grok</a>, <a href="/wiki/Anthropic#Claude" title="Anthropic">Claude</a>, <a href="/wiki/Microsoft_Copilot" title="Microsoft Copilot">Copilot</a>, and <a class="mw-redirect" href="/wiki/LLaMA" title="LLaMA">LLaMA</a>.<sup class="reference" id="cite_ref-139"><a href="#cite_note-139">[129]</a></sup> <a href="/wiki/Multimodal_learning" title="Multimodal learning">Multimodal</a> GPT models can process different types of data (<a href="/wiki/Modality_(human%E2%80%93computer_interaction)" title="Modality (human–computer interaction)">modalities</a>) such as images, videos, sound, and text.<sup class="reference" id="cite_ref-FOOTNOTEMarmouyet2023_140-0"><a href="#cite_note-FOOTNOTEMarmouyet2023-140">[130]</a></sup>
</p><p>In the late 2010s, <a href="/wiki/Graphics_processing_unit" title="Graphics processing unit">graphics processing units</a> (GPUs) that were increasingly designed with AI-specific enhancements and used with specialized <a href="/wiki/TensorFlow" title="TensorFlow">TensorFlow</a> software had replaced previously used <a href="/wiki/Central_processing_unit" title="Central processing unit">central processing unit</a> (CPUs) as the dominant means for large-scale (commercial and academic) <a href="/wiki/Machine_learning" title="Machine learning">machine learning</a> models' training.<sup class="reference" id="cite_ref-FOOTNOTEKobielus2019_141-0"><a href="#cite_note-FOOTNOTEKobielus2019-141">[131]</a></sup> Specialized <a href="/wiki/Programming_language" title="Programming language">programming languages</a> such as <a href="/wiki/Prolog" title="Prolog">Prolog</a> were used in early AI research,<sup class="reference" id="cite_ref-142"><a href="#cite_note-142">[132]</a></sup> but <a href="/wiki/General-purpose_programming_language" title="General-purpose programming language">general-purpose programming languages</a> like <a href="/wiki/Python_(programming_language)" title="Python (programming language)">Python</a> have become predominant.<sup class="reference" id="cite_ref-143"><a href="#cite_note-143">[133]</a></sup>
</p><p>AI and machine learning technology is used in most of the essential applications of the 2020s, including: <a class="mw-redirect" href="/wiki/Search_engines" title="Search engines">search engines</a> (such as <a href="/wiki/Google_Search" title="Google Search">Google Search</a>), <a href="/wiki/Targeted_advertising" title="Targeted advertising">targeting online advertisements</a>, <a href="/wiki/Recommender_system" title="Recommender system">recommendation systems</a> (offered by <a href="/wiki/Netflix" title="Netflix">Netflix</a>, <a href="/wiki/YouTube" title="YouTube">YouTube</a> or <a href="/wiki/Amazon_(company)" title="Amazon (company)">Amazon</a>), driving <a href="/wiki/Internet_traffic" title="Internet traffic">internet traffic</a>, <a href="/wiki/Marketing_and_artificial_intelligence" title="Marketing and artificial intelligence">targeted advertising</a> (<a class="mw-redirect" href="/wiki/AdSense" title="AdSense">AdSense</a>, <a href="/wiki/Facebook" title="Facebook">Facebook</a>), <a href="/wiki/Virtual_assistant" title="Virtual assistant">virtual assistants</a> (such as <a href="/wiki/Siri" title="Siri">Siri</a> or <a href="/wiki/Amazon_Alexa" title="Amazon Alexa">Alexa</a>), <a class="mw-redirect" href="/wiki/Autonomous_vehicles" title="Autonomous vehicles">autonomous vehicles</a> (including <a href="/wiki/Unmanned_aerial_vehicle" title="Unmanned aerial vehicle">drones</a>, <a href="/wiki/Advanced_driver-assistance_system" title="Advanced driver-assistance system">ADAS</a> and <a class="mw-redirect" href="/wiki/Self-driving_cars" title="Self-driving cars">self-driving cars</a>), <a href="/wiki/Machine_translation" title="Machine translation">automatic language translation</a> (<a href="/wiki/Microsoft_Translator" title="Microsoft Translator">Microsoft Translator</a>, <a href="/wiki/Google_Translate" title="Google Translate">Google Translate</a>), <a href="/wiki/Facial_recognition_system" title="Facial recognition system">facial recognition</a> (<a class="mw-redirect" href="/wiki/Apple_Computer" title="Apple Computer">Apple</a>'s <a href="/wiki/Face_ID" title="Face ID">Face ID</a> or <a href="/wiki/Microsoft" title="Microsoft">Microsoft</a>'s <a href="/wiki/DeepFace" title="DeepFace">DeepFace</a> and <a href="/wiki/Google" title="Google">Google</a>'s <a href="/wiki/FaceNet" title="FaceNet">FaceNet</a>) and <a href="/wiki/Automatic_image_annotation" title="Automatic image annotation">image labeling</a> (used by <a href="/wiki/Facebook" title="Facebook">Facebook</a>, Apple's <a href="/wiki/IPhoto" title="IPhoto">iPhoto</a> and <a href="/wiki/TikTok" title="TikTok">TikTok</a>).
The deployment of AI may be overseen by a <a href="/wiki/Chief_automation_officer" title="Chief automation officer">Chief automation officer</a> (CAO).
</p><p>The application of AI in <a href="/wiki/Medicine" title="Medicine">medicine</a> and <a href="/wiki/Medical_research" title="Medical research">medical research</a> has the potential to increase patient care and quality of life.<sup class="reference" id="cite_ref-144"><a href="#cite_note-144">[134]</a></sup> Through the lens of the <a href="/wiki/Hippocratic_Oath" title="Hippocratic Oath">Hippocratic Oath</a>, medical professionals are ethically compelled to use AI, if applications can more accurately diagnose and treat patients.
</p><p>For medical research, AI is an important tool for processing and integrating <a href="/wiki/Big_data" title="Big data">big data</a>.
This is particularly important for <a href="/wiki/Organoid" title="Organoid">organoid</a> and <a href="/wiki/Tissue_engineering" title="Tissue engineering">tissue engineering</a> development which use <a href="/wiki/Microscopy" title="Microscopy">microscopy</a> imaging as a key technique in fabrication.<sup class="reference" id="cite_ref-The_future_of_personalized_cardiova_145-0"><a href="#cite_note-The_future_of_personalized_cardiova-145">[135]</a></sup> It has been suggested that AI can overcome discrepancies in funding allocated to different fields of research.<sup class="reference" id="cite_ref-The_future_of_personalized_cardiova_145-1"><a href="#cite_note-The_future_of_personalized_cardiova-145">[135]</a></sup> New AI tools can deepen the understanding of biomedically relevant pathways.
For example, <a class="mw-redirect" href="/wiki/AlphaFold_2" title="AlphaFold 2">AlphaFold 2</a> (2021) demonstrated the ability to approximate, in hours rather than months, the 3D <a href="/wiki/Protein_structure" title="Protein structure">structure of a protein</a>.<sup class="reference" id="cite_ref-146"><a href="#cite_note-146">[136]</a></sup> In 2023, it was reported that AI-guided drug discovery helped find a class of antibiotics capable of killing two different types of drug-resistant bacteria.<sup class="reference" id="cite_ref-147"><a href="#cite_note-147">[137]</a></sup> In 2024, researchers used machine learning to accelerate the search for <a href="/wiki/Parkinson%27s_disease" title="Parkinson's disease">Parkinson's disease</a> drug treatments.
Their aim was to identify compounds that block the clumping, or aggregation, of <a href="/wiki/Alpha-synuclein" title="Alpha-synuclein">alpha-synuclein</a> (the protein that characterises Parkinson's disease).
They were able to speed up the initial screening process ten-fold and reduce the cost by a thousand-fold.<sup class="reference" id="cite_ref-148"><a href="#cite_note-148">[138]</a></sup><sup class="reference" id="cite_ref-149"><a href="#cite_note-149">[139]</a></sup>
</p><p><a class="mw-redirect" href="/wiki/Game_AI" title="Game AI">Game playing</a> programs have been used since the 1950s to demonstrate and test AI's most advanced techniques.<sup class="reference" id="cite_ref-150"><a href="#cite_note-150">[140]</a></sup> <a class="mw-redirect" href="/wiki/IBM_Deep_Blue" title="IBM Deep Blue">Deep Blue</a> became the first computer chess-playing system to beat a reigning world chess champion, <a href="/wiki/Garry_Kasparov" title="Garry Kasparov">Garry Kasparov</a>, on 11 May 1997.<sup class="reference" id="cite_ref-151"><a href="#cite_note-151">[141]</a></sup> In 2011, in a <i><a href="/wiki/Jeopardy!"
title="Jeopardy!
">Jeopardy!</a></i> <a class="mw-redirect" href="/wiki/Quiz_show" title="Quiz show">quiz show</a> exhibition match, <a href="/wiki/IBM" title="IBM">IBM</a>'s <a class="mw-redirect" href="/wiki/Question_answering_system" title="Question answering system">question answering system</a>, <a class="mw-redirect" href="/wiki/Watson_(artificial_intelligence_software)" title="Watson (artificial intelligence software)">Watson</a>, defeated the two greatest <i>Jeopardy!</i> champions, <a href="/wiki/Brad_Rutter" title="Brad Rutter">Brad Rutter</a> and <a href="/wiki/Ken_Jennings" title="Ken Jennings">Ken Jennings</a>, by a significant margin.<sup class="reference" id="cite_ref-152"><a href="#cite_note-152">[142]</a></sup> In March 2016, <a href="/wiki/AlphaGo" title="AlphaGo">AlphaGo</a> won 4 out of 5 games of <a href="/wiki/Go_(game)" title="Go (game)">Go</a> in a match with Go champion <a href="/wiki/Lee_Sedol" title="Lee Sedol">Lee Sedol</a>, becoming the first <a href="/wiki/Computer_Go" title="Computer Go">computer Go</a>-playing system to beat a professional Go player without <a class="mw-redirect" href="/wiki/Go_handicaps" title="Go handicaps">handicaps</a>.
Then in 2017 it <a href="/wiki/AlphaGo_versus_Ke_Jie" title="AlphaGo versus Ke Jie">defeated Ke Jie</a>, who was the best Go player in the world.<sup class="reference" id="cite_ref-153"><a href="#cite_note-153">[143]</a></sup> Other programs handle <a class="mw-redirect" href="/wiki/Imperfect_information" title="Imperfect information">imperfect-information</a> games, such as the <a href="/wiki/Poker" title="Poker">poker</a>-playing program <a href="/wiki/Pluribus_(poker_bot)" title="Pluribus (poker bot)">Pluribus</a>.<sup class="reference" id="cite_ref-154"><a href="#cite_note-154">[144]</a></sup> <a class="mw-redirect" href="/wiki/DeepMind" title="DeepMind">DeepMind</a> developed increasingly generalistic <a href="/wiki/Reinforcement_learning" title="Reinforcement learning">reinforcement learning</a> models, such as with <a href="/wiki/MuZero" title="MuZero">MuZero</a>, which could be trained to play chess, Go, or <a href="/wiki/Atari" title="Atari">Atari</a> games.<sup class="reference" id="cite_ref-155"><a href="#cite_note-155">[145]</a></sup> In 2019, DeepMind's AlphaStar achieved grandmaster level in <a href="/wiki/StarCraft_II" title="StarCraft II">StarCraft II</a>, a particularly challenging real-time strategy game that involves incomplete knowledge of what happens on the map.<sup class="reference" id="cite_ref-156"><a href="#cite_note-156">[146]</a></sup> In 2021, an AI agent competed in a PlayStation <a href="/wiki/Gran_Turismo_(series)" title="Gran Turismo (series)">Gran Turismo</a> competition, winning against four of the world's best Gran Turismo drivers using deep reinforcement learning.<sup class="reference" id="cite_ref-157"><a href="#cite_note-157">[147]</a></sup> In 2024, Google DeepMind introduced SIMA, a type of AI capable of autonomously playing nine previously unseen <a href="/wiki/Open_world" title="Open world">open-world</a> video games by observing screen output, as well as executing short, specific tasks in response to natural language instructions.<sup class="reference" id="cite_ref-158"><a href="#cite_note-158">[148]</a></sup>
</p><p>Finance is one of the fastest growing sectors where applied AI tools are being deployed: from retail online banking to investment advice and insurance, where automated "robot advisers" have been in use for some years.
<sup class="reference" id="cite_ref-159"><a href="#cite_note-159">[149]</a></sup>
</p><p><a href="/wiki/World_Pensions_%26_Investments_Forum" title="World Pensions & Investments Forum">World Pensions</a> experts like Nicolas Firzli insist it may be too early to see the emergence of highly innovative AI-informed financial products and services: "the deployment of AI tools will simply further automatise things: destroying tens of thousands of jobs in banking, financial planning, and pension advice in the process, but I’m not sure it will unleash a new wave of [e.g., sophisticated] pension innovation.
"<sup class="reference" id="cite_ref-160"><a href="#cite_note-160">[150]</a></sup>
</p><p>Various countries are deploying AI military applications.<sup class="reference" id="cite_ref-:22_161-0"><a href="#cite_note-:22-161">[151]</a></sup> The main applications enhance <a href="/wiki/Command_and_control" title="Command and control">command and control</a>, communications, sensors, integration and interoperability.<sup class="reference" id="cite_ref-AI_162-0"><a href="#cite_note-AI-162">[152]</a></sup> Research is targeting intelligence collection and analysis, logistics, cyber operations, information operations, and semiautonomous and <a href="/wiki/Vehicular_automation" title="Vehicular automation">autonomous vehicles</a>.<sup class="reference" id="cite_ref-:22_161-1"><a href="#cite_note-:22-161">[151]</a></sup> AI technologies enable coordination of sensors and effectors, threat detection and identification, marking of enemy positions, <a href="/wiki/Target_acquisition" title="Target acquisition">target acquisition</a>, coordination and deconfliction of distributed <a href="/wiki/Forward_observers_in_the_U.S._military" title="Forward observers in the U.S. military">Joint Fires</a> between networked combat vehicles involving manned and unmanned teams.<sup class="reference" id="cite_ref-AI_162-1"><a href="#cite_note-AI-162">[152]</a></sup> AI was incorporated into military operations in Iraq and Syria.<sup class="reference" id="cite_ref-:22_161-2"><a href="#cite_note-:22-161">[151]</a></sup>
</p><p>In November 2023, US Vice President <a href="/wiki/Kamala_Harris" title="Kamala Harris">Kamala Harris</a> disclosed a declaration signed by 31 nations to set guardrails for the military use of AI.
The commitments include using legal reviews to ensure the compliance of military AI with international laws, and being cautious and transparent in the development of this technology.<sup class="reference" id="cite_ref-163"><a href="#cite_note-163">[153]</a></sup>
</p><p>In the early 2020s, <a class="mw-redirect" href="/wiki/Generative_AI" title="Generative AI">generative AI</a> gained widespread prominence.
In March 2023, 58% of U.S. adults had heard about <a href="/wiki/ChatGPT" title="ChatGPT">ChatGPT</a> and 14% had tried it.<sup class="reference" id="cite_ref-164"><a href="#cite_note-164">[154]</a></sup> The increasing realism and ease-of-use of AI-based <a href="/wiki/Text-to-image_model" title="Text-to-image model">text-to-image</a> generators such as <a href="/wiki/Midjourney" title="Midjourney">Midjourney</a>, <a href="/wiki/DALL-E" title="DALL-E">DALL-E</a>, and <a href="/wiki/Stable_Diffusion" title="Stable Diffusion">Stable Diffusion</a> sparked a trend of <a href="/wiki/Viral_phenomenon" title="Viral phenomenon">viral</a> AI-generated photos.
Widespread attention was gained by a fake photo of <a href="/wiki/Pope_Francis" title="Pope Francis">Pope Francis</a> wearing a white puffer coat, the fictional arrest of <a href="/wiki/Donald_Trump" title="Donald Trump">Donald Trump</a>, and a hoax of an attack on the <a href="/wiki/The_Pentagon" title="The Pentagon">Pentagon</a>, as well as the usage in professional creative arts.<sup class="reference" id="cite_ref-165"><a href="#cite_note-165">[155]</a></sup><sup class="reference" id="cite_ref-166"><a href="#cite_note-166">[156]</a></sup>
</p><p>There are also thousands of successful AI applications used to solve specific problems for specific industries or institutions.
In a 2017 survey, one in five companies reported having incorporated "AI" in some offerings or processes.<sup class="reference" id="cite_ref-167"><a href="#cite_note-167">[157]</a></sup> A few examples are <a href="/wiki/Energy_storage" title="Energy storage">energy storage</a>, medical diagnosis, military logistics, applications that predict the result of judicial decisions, <a href="/wiki/Foreign_policy" title="Foreign policy">foreign policy</a>, or supply chain management.
</p><p>AI applications for evacuation and <a href="/wiki/Disaster" title="Disaster">disaster</a> management are growing.
AI has been used to investigate if and how people evacuated in large scale and small scale evacuations using historical data from GPS, videos or social media.
Further, AI can provide real time information on the real time evacuation conditions.<sup class="reference" id="cite_ref-168"><a href="#cite_note-168">[158]</a></sup><sup class="reference" id="cite_ref-169"><a href="#cite_note-169">[159]</a></sup><sup class="reference" id="cite_ref-170"><a href="#cite_note-170">[160]</a></sup>
</p><p>In agriculture, AI has helped farmers identify areas that need irrigation, fertilization, pesticide treatments or increasing yield.
Agronomists use AI to conduct research and development.
AI has been used to predict the ripening time for crops such as tomatoes, monitor soil moisture, operate agricultural robots, conduct predictive analytics, classify livestock pig call emotions, automate greenhouses, detect diseases and pests, and save water.
</p><p>Artificial intelligence is used in astronomy to analyze increasing amounts of available data and applications, mainly for "classification, regression, clustering, forecasting, generation, discovery, and the development of new scientific insights" for example for discovering exoplanets, forecasting solar activity, and distinguishing between signals and instrumental effects in gravitational wave astronomy.
It could also be used for activities in space such as space exploration, including analysis of data from space missions, real-time science decisions of spacecraft, space debris avoidance, and more autonomous operation.
</p><p>AI has potential benefits and potential risks.
AI may be able to advance science and find solutions for serious problems: <a href="/wiki/Demis_Hassabis" title="Demis Hassabis">Demis Hassabis</a> of <a class="mw-redirect" href="/wiki/DeepMind" title="DeepMind">Deep Mind</a> hopes to "solve intelligence, and then use that to solve everything else".<sup class="reference" id="cite_ref-FOOTNOTESimonite2016_171-0"><a href="#cite_note-FOOTNOTESimonite2016-171">[161]</a></sup> However, as the use of AI has become widespread, several unintended consequences and risks have been identified.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021987_172-0"><a href="#cite_note-FOOTNOTERussellNorvig2021987-172">[162]</a></sup> In-production systems can sometimes not factor ethics and bias into their AI training processes, especially when the AI algorithms are inherently unexplainable in deep learning.<sup class="reference" id="cite_ref-FOOTNOTELaskowski2023_173-0"><a href="#cite_note-FOOTNOTELaskowski2023-173">[163]</a></sup>
</p><p>Machine-learning algorithms require large amounts of data.
The techniques used to acquire this data have raised concerns about <a href="/wiki/Privacy" title="Privacy">privacy</a>, <a href="/wiki/Surveillance" title="Surveillance">surveillance</a> and <a href="/wiki/Copyright" title="Copyright">copyright</a>.
</p><p>Technology companies collect a wide range of data from their users, including online activity, geolocation data, video and audio.<sup class="reference" id="cite_ref-FOOTNOTEGAO2022_174-0"><a href="#cite_note-FOOTNOTEGAO2022-174">[164]</a></sup>
For example, in order to build <a href="/wiki/Speech_recognition" title="Speech recognition">speech recognition</a> algorithms, <a href="/wiki/Amazon_(company)" title="Amazon (company)">Amazon</a> has recorded millions of private conversations and allowed <a class="mw-redirect" href="/wiki/Temporary_worker" title="Temporary worker">temporary workers</a> to listen to and transcribe some of them.<sup class="reference" id="cite_ref-FOOTNOTEValinsky2019_175-0"><a href="#cite_note-FOOTNOTEValinsky2019-175">[165]</a></sup> Opinions about this widespread <a href="/wiki/Surveillance" title="Surveillance">surveillance</a> range from those who see it as a <a href="/wiki/Necessary_evil" title="Necessary evil">necessary evil</a> to those for whom it is clearly <a class="mw-redirect" href="/wiki/Unethical" title="Unethical">unethical</a> and a violation of the <a href="/wiki/Right_to_privacy" title="Right to privacy">right to privacy</a>.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021991_176-0"><a href="#cite_note-FOOTNOTERussellNorvig2021991-176">[166]</a></sup>
</p><p>AI developers argue that this is the only way to deliver valuable applications.
and have developed several techniques that attempt to preserve privacy while still obtaining the data, such as <a href="/wiki/Data_aggregation" title="Data aggregation">data aggregation</a>, <a href="/wiki/De-identification" title="De-identification">de-identification</a> and <a href="/wiki/Differential_privacy" title="Differential privacy">differential privacy</a>.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021991–992_177-0"><a href="#cite_note-FOOTNOTERussellNorvig2021991–992-177">[167]</a></sup> Since 2016, some privacy experts, such as <a href="/wiki/Cynthia_Dwork" title="Cynthia Dwork">Cynthia Dwork</a>, have begun to view privacy in terms of <a href="/wiki/Fairness_(machine_learning)" title="Fairness (machine learning)">fairness</a>.
<a href="/wiki/Brian_Christian" title="Brian Christian">Brian Christian</a> wrote that experts have pivoted "from the question of 'what they know' to the question of 'what they're doing with it'.
"<sup class="reference" id="cite_ref-FOOTNOTEChristian202063_178-0"><a href="#cite_note-FOOTNOTEChristian202063-178">[168]</a></sup>
</p><p>Generative AI is often trained on unlicensed copyrighted works, including in domains such as images or computer code; the output is then used under the rationale of "<a href="/wiki/Fair_use" title="Fair use">fair use</a>".
Experts disagree about how well and under what circumstances this rationale will hold up in courts of law; relevant factors may include "the purpose and character of the use of the copyrighted work" and "the effect upon the potential market for the copyrighted work".<sup class="reference" id="cite_ref-FOOTNOTEVincent2022_179-0"><a href="#cite_note-FOOTNOTEVincent2022-179">[169]</a></sup><sup class="reference" id="cite_ref-180"><a href="#cite_note-180">[170]</a></sup> Website owners who do not wish to have their content scraped can indicate it in a "<a href="/wiki/Robots.txt" title="Robots.txt">robots.txt</a>" file.<sup class="reference" id="cite_ref-181"><a href="#cite_note-181">[171]</a></sup> In 2023, leading authors (including <a href="/wiki/John_Grisham" title="John Grisham">John Grisham</a> and <a href="/wiki/Jonathan_Franzen" title="Jonathan Franzen">Jonathan Franzen</a>) sued AI companies for using their work to train generative AI.<sup class="reference" id="cite_ref-FOOTNOTEReisner2023_182-0"><a href="#cite_note-FOOTNOTEReisner2023-182">[172]</a></sup><sup class="reference" id="cite_ref-FOOTNOTEAlterHarris2023_183-0"><a href="#cite_note-FOOTNOTEAlterHarris2023-183">[173]</a></sup> Another discussed approach is to envision a separate <i><a href="/wiki/Sui_generis" title="Sui generis">sui generis</a></i> system of protection for creations generated by AI to ensure fair attribution and compensation for human authors.<sup class="reference" id="cite_ref-184"><a href="#cite_note-184">[174]</a></sup>
</p><p>The commercial AI scene is dominated by <a href="/wiki/Big_Tech" title="Big Tech">Big Tech</a> companies such as <a href="/wiki/Alphabet_Inc."
title="Alphabet Inc.">Alphabet Inc.</a>, <a href="/wiki/Amazon_(company)" title="Amazon (company)">Amazon</a>, <a href="/wiki/Apple_Inc."
title="Apple Inc.">Apple Inc.</a>, <a href="/wiki/Meta_Platforms" title="Meta Platforms">Meta Platforms</a>, and <a href="/wiki/Microsoft" title="Microsoft">Microsoft</a>.<sup class="reference" id="cite_ref-185"><a href="#cite_note-185">[175]</a></sup><sup class="reference" id="cite_ref-186"><a href="#cite_note-186">[176]</a></sup><sup class="reference" id="cite_ref-187"><a href="#cite_note-187">[177]</a></sup> Some of these players already own the vast majority of existing <a href="/wiki/Cloud_computing" title="Cloud computing">cloud infrastructure</a> and <a href="/wiki/Computing" title="Computing">computing</a> power from <a href="/wiki/Data_center" title="Data center">data centers</a>, allowing them to entrench further in the marketplace.<sup class="reference" id="cite_ref-188"><a href="#cite_note-188">[178]</a></sup><sup class="reference" id="cite_ref-189"><a href="#cite_note-189">[179]</a></sup>
</p><p>In January 2024, the <a href="/wiki/International_Energy_Agency" title="International Energy Agency">International Energy Agency</a> (IEA) released <i>Electricity 2024, Analysis and Forecast to 2026</i>, forecasting electric power use.<sup class="reference" id="cite_ref-190"><a href="#cite_note-190">[180]</a></sup> This is the first IEA report to make projections for data centers and power consumption for artificial intelligence and cryptocurrency.
The report states that power demand for these uses might double by 2026, with additional electric power usage equal to electricity used by the whole Japanese nation.<sup class="reference" id="cite_ref-191"><a href="#cite_note-191">[181]</a></sup>
</p><p>Prodigious power consumption by AI is responsible for the growth of fossil fuels use, and might delay closings of obsolete, carbon-emitting coal energy facilities.
There is a feverish rise in the construction of data centers throughout the US, making large technology firms (e.g., Microsoft, Meta, Google, Amazon) into voracious consumers of electric power.
Projected electric consumption is so immense that there is concern that it will be fulfilled no matter the source.
A ChatGPT search involves the use of 10 times the electrical energy as a Google search.
The large firms are in haste to find power sources – from nuclear energy to geothermal to fusion.
The tech firms argue that – in the long view – AI will be eventually kinder to the environment, but they need the energy now.
AI makes the power grid more efficient and "intelligent", will assist in the growth of nuclear power, and track overall carbon emissions, according to technology firms.<sup class="reference" id="cite_ref-192"><a href="#cite_note-192">[182]</a></sup>
</p><p>A 2024 <a href="/wiki/Goldman_Sachs" title="Goldman Sachs">Goldman Sachs</a> Research Paper, <i>AI Data Centers and the Coming US Power Demand Surge</i>, found "US power demand (is) likely to experience growth not seen in a generation…."
and forecasts that, by 2030, US data centers will consume 8% of US power, as opposed to 3% in 2022, presaging growth for the electrical power generation industry by a variety of means.<sup class="reference" id="cite_ref-193"><a href="#cite_note-193">[183]</a></sup>Data centers' need for more and more electrical power is such that they might max out the electrical grid.
The Big Tech companies counter that AI can be used to maximize the utilization of the grid by all.<sup class="reference" id="cite_ref-194"><a href="#cite_note-194">[184]</a></sup>
</p><p>In 2024, the <i>Wall Street Journal</i> reported that big AI companies have begun negotiations with the US nuclear power providers to provide electricity to the data centers.
In March 2024 Amazon purchased a Pennsylvania nuclear-powered data center for $650 Million (US).<sup class="reference" id="cite_ref-195"><a href="#cite_note-195">[185]</a></sup>
</p><p><a href="/wiki/YouTube" title="YouTube">YouTube</a>, <a href="/wiki/Facebook" title="Facebook">Facebook</a> and others use <a href="/wiki/Recommender_system" title="Recommender system">recommender systems</a> to guide users to more content.
These AI programs were given the goal of <a href="/wiki/Mathematical_optimization" title="Mathematical optimization">maximizing</a> user engagement (that is, the only goal was to keep people watching).
The AI learned that users tended to choose <a href="/wiki/Misinformation" title="Misinformation">misinformation</a>, <a href="/wiki/Conspiracy_theory" title="Conspiracy theory">conspiracy theories</a>, and extreme <a href="/wiki/Partisan_(politics)" title="Partisan (politics)">partisan</a> content, and, to keep them watching, the AI recommended more of it.
Users also tended to watch more content on the same subject, so the AI led people into <a class="mw-redirect" href="/wiki/Filter_bubbles" title="Filter bubbles">filter bubbles</a> where they received multiple versions of the same misinformation.<sup class="reference" id="cite_ref-FOOTNOTENicas2018_196-0"><a href="#cite_note-FOOTNOTENicas2018-196">[186]</a></sup> This convinced many users that the misinformation was true, and ultimately undermined trust in institutions, the media and the government.<sup class="reference" id="cite_ref-197"><a href="#cite_note-197">[187]</a></sup> The AI program had correctly learned to maximize its goal, but the result was harmful to society.
After the U.S. election in 2016, major technology companies took steps to mitigate the problem <sup class="noprint Inline-Template Template-Fact" style="white-space:nowrap;">[<i><a href="/wiki/Wikipedia:Citation_needed" title="Wikipedia:Citation needed"><span title="This assertion is not obviously supported by other information on this page (June 2024)">citation needed</span></a></i>]</sup>.
</p><p>In 2022, <a class="mw-redirect" href="/wiki/Generative_AI" title="Generative AI">generative AI</a> began to create images, audio, video and text that are indistinguishable from real photographs, recordings, films, or human writing.
It is possible for bad actors to use this technology to create massive amounts of misinformation or propaganda.<sup class="reference" id="cite_ref-FOOTNOTEWilliams2023_198-0"><a href="#cite_note-FOOTNOTEWilliams2023-198">[188]</a></sup> AI pioneer <a href="/wiki/Geoffrey_Hinton" title="Geoffrey Hinton">Geoffrey Hinton</a> expressed concern about AI enabling "authoritarian leaders to manipulate their electorates" on a large scale, among other risks.<sup class="reference" id="cite_ref-FOOTNOTETaylorHern2023_199-0"><a href="#cite_note-FOOTNOTETaylorHern2023-199">[189]</a></sup>
</p><p>In statistics, a <a href="/wiki/Bias_(statistics)" title="Bias (statistics)">bias</a> is a systematic error or deviation from the correct value.
But in the context of <a href="/wiki/Fairness_(machine_learning)" title="Fairness (machine learning)">fairness</a>, it often refers to a tendency in favor or against a certain group or individual characteristic, usually in a way that is considered unfair or harmful.
A statistically unbiased AI system that produces disparate outcomes for different demographic groups may thus be viewed as biased in the ethical sense.<sup class="reference" id="cite_ref-:03_200-0"><a href="#cite_note-:03-200">[190]</a></sup>
</p><p>The field of fairness studies how to prevent harms from algorithmic biases.
There are various conflicting definitions and mathematical models of fairness.
These notions depend on ethical assumptions, and are influenced by beliefs about society.
One broad category is distributive fairness, which focuses on the outcomes, often identifying groups and seeking to compensate for statistical disparities.
Representational fairness tries to ensure that AI systems don't reinforce negative <a href="/wiki/Stereotype" title="Stereotype">stereotypes</a> or render certain groups invisible.
Procedural fairness focuses on the decision process rather than the outcome.
The most relevant notions of fairness may depend on the context, notably the type of AI application and the stakeholders.
The subjectivity in the notions of bias and fairness makes it difficult for companies to operationalize them.
Having access to sensitive attributes such as race or gender is also considered by many AI ethicists to be necessary in order to compensate for biases, but it may conflict with <a href="/wiki/Anti-discrimination_law" title="Anti-discrimination law">anti-discrimination laws</a>.<sup class="reference" id="cite_ref-:03_200-1"><a href="#cite_note-:03-200">[190]</a></sup>
</p><p>Machine learning applications will be biased if they learn from biased data.<sup class="reference" id="cite_ref-FOOTNOTERose2023_201-0"><a href="#cite_note-FOOTNOTERose2023-201">[191]</a></sup> The developers may not be aware that the bias exists.<sup class="reference" id="cite_ref-FOOTNOTECNA2019_202-0"><a href="#cite_note-FOOTNOTECNA2019-202">[192]</a></sup>
Bias can be introduced by the way <a class="mw-redirect" href="/wiki/Training_data" title="Training data">training data</a> is selected and by the way a model is deployed.<sup class="reference" id="cite_ref-FOOTNOTEGoffrey200817_203-0"><a href="#cite_note-FOOTNOTEGoffrey200817-203">[193]</a></sup><sup class="reference" id="cite_ref-FOOTNOTERose2023_201-1"><a href="#cite_note-FOOTNOTERose2023-201">[191]</a></sup> If a biased algorithm is used to make decisions that can seriously <a href="/wiki/Harm" title="Harm">harm</a> people (as it can in <a href="/wiki/Health_equity" title="Health equity">medicine</a>, <a href="/wiki/Credit_rating" title="Credit rating">finance</a>, <a href="/wiki/Recruitment" title="Recruitment">recruitment</a>, <a href="/wiki/Public_housing" title="Public housing">housing</a> or <a class="mw-redirect" href="/wiki/Policing" title="Policing">policing</a>) then the algorithm may cause <a href="/wiki/Discrimination" title="Discrimination">discrimination</a>.<sup class="reference" id="cite_ref-204"><a href="#cite_note-204">[194]</a></sup>
</p><p>On June 28, 2015, <a href="/wiki/Google_Photos" title="Google Photos">Google Photos</a>'s new image labeling feature mistakenly identified Jacky Alcine and a friend as "gorillas" because they were black.
The system was trained on a dataset that contained very few images of black people,<sup class="reference" id="cite_ref-FOOTNOTEChristian202025_205-0"><a href="#cite_note-FOOTNOTEChristian202025-205">[195]</a></sup> a problem called "sample size disparity".<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021995_206-0"><a href="#cite_note-FOOTNOTERussellNorvig2021995-206">[196]</a></sup> Google "fixed" this problem by preventing the system from labelling <i>anything</i> as a "gorilla".
Eight years later, in 2023, Google Photos still could not identify a gorilla, and neither could similar products from Apple, Facebook, Microsoft and Amazon.<sup class="reference" id="cite_ref-FOOTNOTEGrantHill2023_207-0"><a href="#cite_note-FOOTNOTEGrantHill2023-207">[197]</a></sup>
</p><p><a href="/wiki/COMPAS_(software)" title="COMPAS (software)">COMPAS</a> is a commercial program widely used by <a class="mw-redirect" href="/wiki/U.S._court" title="U.S. court">U.S.
courts</a> to assess the likelihood of a <a href="/wiki/Defendant" title="Defendant">defendant</a> becoming a <a class="mw-redirect" href="/wiki/Recidivist" title="Recidivist">recidivist</a>.
In 2016, <a href="/wiki/Julia_Angwin" title="Julia Angwin">Julia Angwin</a> at <a href="/wiki/ProPublica" title="ProPublica">ProPublica</a> discovered that COMPAS exhibited racial bias, despite the fact that the program was not told the races of the defendants.
Although the error rate for both whites and blacks was calibrated equal at exactly 61%, the errors for each race were different—the system consistently overestimated the chance that a black person would re-offend and would underestimate the chance that a white person would not re-offend.<sup class="reference" id="cite_ref-FOOTNOTELarsonAngwin2016_208-0"><a href="#cite_note-FOOTNOTELarsonAngwin2016-208">[198]</a></sup> In 2017, several researchers<sup class="reference" id="cite_ref-210"><a href="#cite_note-210">[k]</a></sup> showed that it was mathematically impossible for COMPAS to accommodate all possible measures of fairness when the base rates of re-offense were different for whites and blacks in the data.<sup class="reference" id="cite_ref-211"><a href="#cite_note-211">[200]</a></sup>
</p><p>A program can make biased decisions even if the data does not explicitly mention a problematic feature (such as "race" or "gender").
The feature will correlate with other features (like "address", "shopping history" or "first name"), and the program will make the same decisions based on these features as it would on "race" or "gender".<sup class="reference" id="cite_ref-212"><a href="#cite_note-212">[201]</a></sup>
Moritz Hardt said "the most robust fact in this research area is that fairness through blindness doesn't work.
"<sup class="reference" id="cite_ref-213"><a href="#cite_note-213">[202]</a></sup>
</p><p>Criticism of COMPAS highlighted that machine learning models are designed to make "predictions" that are only valid if we assume that the future will resemble the past.
If they are trained on data that includes the results of racist decisions in the past, machine learning models must predict that racist decisions will be made in the future.
If an application then uses these predictions as <i>recommendations</i>, some of these "recommendations" will likely be racist.<sup class="reference" id="cite_ref-214"><a href="#cite_note-214">[203]</a></sup> Thus, machine learning is not well suited to help make decisions in areas where there is hope that the future will be <i>better</i> than the past.
It is descriptive rather than prescriptive.<sup class="reference" id="cite_ref-216"><a href="#cite_note-216">[l]</a></sup>
</p><p>Bias and unfairness may go undetected because the developers are overwhelmingly white and male: among AI engineers, about 4% are black and 20% are women.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021995_206-1"><a href="#cite_note-FOOTNOTERussellNorvig2021995-206">[196]</a></sup>
</p><p>At its 2022 <a href="/wiki/ACM_Conference_on_Fairness,_Accountability,_and_Transparency" title="ACM Conference on Fairness, Accountability, and Transparency">Conference on Fairness, Accountability, and Transparency</a> (ACM FAccT 2022), the <a href="/wiki/Association_for_Computing_Machinery" title="Association for Computing Machinery">Association for Computing Machinery</a>, in Seoul, South Korea, presented and published findings that recommend that until AI and robotics systems are demonstrated to be free of bias mistakes, they are unsafe, and the use of self-learning neural networks trained on vast, unregulated sources of flawed internet data should be curtailed.<sup class="noprint Inline-Template" style="white-space:nowrap;">[<i><a href="/wiki/Wikipedia:Accuracy_dispute#Disputed_statement" title="Wikipedia:Accuracy dispute"><span title="Depending on what is meant by "free of bias", it may be impossible in practice to demonstrate it.
Additionally, the study evaluates the priors (initial assumptions) of the robots, rather than their decision-making in scenarios where there is a correct choice.
For example, it may not be sexist to have the prior that most doctors are males (it's actually an accurate statistical prior in the world we currently live in, so the bias may arguably be to not have this prior).
If forced to choose which one is the doctor based solely on gender, a rational person seeking to maximize the number of correct answers would choose the man 100% of the time.
The real issue arises when such priors lead to significant discrimination.
(July 2024)">dubious</span></a> – <a href="/wiki/Talk:Artificial_intelligence#Dubious" title="Talk:Artificial intelligence">discuss</a></i>]</sup><sup class="reference" id="cite_ref-FOOTNOTEDockrill2022_217-0"><a href="#cite_note-FOOTNOTEDockrill2022-217">[205]</a></sup>
</p><p>Many AI systems are so complex that their designers cannot explain how they reach their decisions.<sup class="reference" id="cite_ref-FOOTNOTESample2017_218-0"><a href="#cite_note-FOOTNOTESample2017-218">[206]</a></sup> Particularly with <a class="mw-redirect" href="/wiki/Deep_neural_networks" title="Deep neural networks">deep neural networks</a>, in which there are a large amount of non-<a class="mw-redirect" href="/wiki/Linear" title="Linear">linear</a> relationships between inputs and outputs.
But some popular explainability techniques exist.<sup class="reference" id="cite_ref-219"><a href="#cite_note-219">[207]</a></sup>
</p><p>It is impossible to be certain that a program is operating correctly if no one knows how exactly it works.
There have been many cases where a machine learning program passed rigorous tests, but nevertheless learned something different than what the programmers intended.
For example, a system that could identify skin diseases better than medical professionals was found to actually have a strong tendency to classify images with a <a href="/wiki/Ruler" title="Ruler">ruler</a> as "cancerous", because pictures of malignancies typically include a ruler to show the scale.<sup class="reference" id="cite_ref-FOOTNOTEChristian2020110_220-0"><a href="#cite_note-FOOTNOTEChristian2020110-220">[208]</a></sup> Another machine learning system designed to help effectively allocate medical resources was found to classify patients with asthma as being at "low risk" of dying from pneumonia.
Having asthma is actually a severe risk factor, but since the patients having asthma would usually get much more medical care, they were relatively unlikely to die according to the training data.
The correlation between asthma and low risk of dying from pneumonia was real, but misleading.<sup class="reference" id="cite_ref-FOOTNOTEChristian202088–91_221-0"><a href="#cite_note-FOOTNOTEChristian202088–91-221">[209]</a></sup>
</p><p>People who have been harmed by an algorithm's decision have a right to an explanation.<sup class="reference" id="cite_ref-222"><a href="#cite_note-222">[210]</a></sup> Doctors, for example, are expected to clearly and completely explain to their colleagues the reasoning behind any decision they make.
Early drafts of the European Union's <a href="/wiki/General_Data_Protection_Regulation" title="General Data Protection Regulation">General Data Protection Regulation</a> in 2016 included an explicit statement that this right exists.<sup class="reference" id="cite_ref-223"><a href="#cite_note-223">[m]</a></sup> Industry experts noted that this is an unsolved problem with no solution in sight.
Regulators argued that nevertheless the harm is real: if the problem has no solution, the tools should not be used.<sup class="reference" id="cite_ref-FOOTNOTEChristian202091_224-0"><a href="#cite_note-FOOTNOTEChristian202091-224">[211]</a></sup>
</p><p><a href="/wiki/DARPA" title="DARPA">DARPA</a> established the <a class="mw-redirect" href="/wiki/Explainable_Artificial_Intelligence" title="Explainable Artificial Intelligence">XAI</a> ("Explainable Artificial Intelligence") program in 2014 to try and solve these problems.<sup class="reference" id="cite_ref-FOOTNOTEChristian202083_225-0"><a href="#cite_note-FOOTNOTEChristian202083-225">[212]</a></sup>
</p><p>Several approaches aim to address the transparency problem.
SHAP enables to visualise the contribution of each feature to the output.<sup class="reference" id="cite_ref-FOOTNOTEVerma2021_226-0"><a href="#cite_note-FOOTNOTEVerma2021-226">[213]</a></sup> LIME can locally approximate a model's outputs with a simpler, interpretable model.<sup class="reference" id="cite_ref-FOOTNOTERothman2020_227-0"><a href="#cite_note-FOOTNOTERothman2020-227">[214]</a></sup> <a class="mw-redirect" href="/wiki/Multitask_learning" title="Multitask learning">Multitask learning</a> provides a large number of outputs in addition to the target classification.
These other outputs can help developers deduce what the network has learned.<sup class="reference" id="cite_ref-FOOTNOTEChristian2020105–108_228-0"><a href="#cite_note-FOOTNOTEChristian2020105–108-228">[215]</a></sup> <a href="/wiki/Deconvolution" title="Deconvolution">Deconvolution</a>, <a href="/wiki/DeepDream" title="DeepDream">DeepDream</a> and other <a class="mw-redirect" href="/wiki/Generative_AI" title="Generative AI">generative</a> methods can allow developers to see what different layers of a deep network for computer vision have learned, and produce output that can suggest what the network is learning.<sup class="reference" id="cite_ref-FOOTNOTEChristian2020108–112_229-0"><a href="#cite_note-FOOTNOTEChristian2020108–112-229">[216]</a></sup> For <a href="/wiki/Generative_pre-trained_transformer" title="Generative pre-trained transformer">generative pre-trained transformers</a>, <a href="/wiki/Anthropic" title="Anthropic">Anthropic</a> developed a technique based on <a class="mw-redirect" href="/wiki/Dictionary_learning" title="Dictionary learning">dictionary learning</a> that associates patterns of neuron activations with human-understandable concepts.<sup class="reference" id="cite_ref-230"><a href="#cite_note-230">[217]</a></sup>
</p><p>Artificial intelligence provides a number of tools that are useful to <a class="mw-redirect" href="/wiki/Bad_actor" title="Bad actor">bad actors</a>, such as <a class="mw-redirect" href="/wiki/Authoritarian" title="Authoritarian">authoritarian governments</a>, <a class="mw-redirect" href="/wiki/Terrorist" title="Terrorist">terrorists</a>, <a class="mw-redirect" href="/wiki/Criminals" title="Criminals">criminals</a> or <a class="mw-redirect" href="/wiki/Rogue_states" title="Rogue states">rogue states</a>.
</p><p>A lethal autonomous weapon is a machine that locates, selects and engages human targets without human supervision.<sup class="reference" id="cite_ref-232"><a href="#cite_note-232">[n]</a></sup> Widely available AI tools can be used by bad actors to develop inexpensive autonomous weapons and, if produced at scale, they are potentially <a class="mw-redirect" href="/wiki/Weapons_of_mass_destruction" title="Weapons of mass destruction">weapons of mass destruction</a>.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021987–990_233-0"><a href="#cite_note-FOOTNOTERussellNorvig2021987–990-233">[219]</a></sup> Even when used in conventional warfare, it is unlikely that they will be unable to reliably choose targets and could potentially <a href="/wiki/Murder" title="Murder">kill an innocent person</a>.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021987–990_233-1"><a href="#cite_note-FOOTNOTERussellNorvig2021987–990-233">[219]</a></sup> In 2014, 30 nations (including China) supported a ban on autonomous weapons under the <a href="/wiki/United_Nations" title="United Nations">United Nations</a>' <a href="/wiki/Convention_on_Certain_Conventional_Weapons" title="Convention on Certain Conventional Weapons">Convention on Certain Conventional Weapons</a>, however the <a href="/wiki/United_States" title="United States">United States</a> and others disagreed.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021988_234-0"><a href="#cite_note-FOOTNOTERussellNorvig2021988-234">[220]</a></sup> By 2015, over fifty countries were reported to be researching battlefield robots.<sup class="reference" id="cite_ref-235"><a href="#cite_note-235">[221]</a></sup>
</p><p>AI tools make it easier for <a class="mw-redirect" href="/wiki/Authoritarian" title="Authoritarian">authoritarian governments</a> to efficiently control their citizens in several ways.
<a href="/wiki/Facial_recognition_system" title="Facial recognition system">Face</a> and <a href="/wiki/Speaker_recognition" title="Speaker recognition">voice recognition</a> allow widespread <a href="/wiki/Surveillance" title="Surveillance">surveillance</a>.
<a href="/wiki/Machine_learning" title="Machine learning">Machine learning</a>, operating this data, can <a class="mw-redirect" href="/wiki/Classifier_(machine_learning)" title="Classifier (machine learning)">classify</a> potential enemies of the state and prevent them from hiding.
<a href="/wiki/Recommender_system" title="Recommender system">Recommendation systems</a> can precisely target <a href="/wiki/Propaganda" title="Propaganda">propaganda</a> and <a href="/wiki/Misinformation" title="Misinformation">misinformation</a> for maximum effect.
<a class="mw-redirect" href="/wiki/Deepfakes" title="Deepfakes">Deepfakes</a> and <a class="mw-redirect" href="/wiki/Generative_AI" title="Generative AI">generative AI</a> aid in producing misinformation.
Advanced AI can make authoritarian <a href="/wiki/Technocracy" title="Technocracy">centralized decision making</a> more competitive than liberal and decentralized systems such as <a href="/wiki/Market_(economics)" title="Market (economics)">markets</a>.
It lowers the cost and difficulty of <a class="mw-redirect" href="/wiki/Digital_warfare" title="Digital warfare">digital warfare</a> and <a href="/wiki/Spyware" title="Spyware">advanced spyware</a>.<sup class="reference" id="cite_ref-FOOTNOTEHarari2018_236-0"><a href="#cite_note-FOOTNOTEHarari2018-236">[222]</a></sup> All these technologies have been available since 2020 or earlier—AI <a href="/wiki/Facial_recognition_system" title="Facial recognition system">facial recognition systems</a> are already being used for <a href="/wiki/Mass_surveillance" title="Mass surveillance">mass surveillance</a> in China.<sup class="reference" id="cite_ref-237"><a href="#cite_note-237">[223]</a></sup><sup class="reference" id="cite_ref-238"><a href="#cite_note-238">[224]</a></sup>
</p><p>There many other ways that AI is expected to help bad actors, some of which can not be foreseen.
For example, machine-learning AI is able to design tens of thousands of toxic molecules in a matter of hours.<sup class="reference" id="cite_ref-FOOTNOTEUrbinaLentzosInvernizziEkins2022_239-0"><a href="#cite_note-FOOTNOTEUrbinaLentzosInvernizziEkins2022-239">[225]</a></sup>
</p><p>Economists have frequently highlighted the risks of redundancies from AI, and speculated about unemployment if there is no adequate social policy for full employment.<sup class="reference" id="cite_ref-auto1_240-0"><a href="#cite_note-auto1-240">[226]</a></sup>
</p><p>In the past, technology has tended to increase rather than reduce total employment, but economists acknowledge that "we're in uncharted territory" with AI.<sup class="reference" id="cite_ref-241"><a href="#cite_note-241">[227]</a></sup> A survey of economists showed disagreement about whether the increasing use of robots and AI will cause a substantial increase in long-term <a href="/wiki/Unemployment" title="Unemployment">unemployment</a>, but they generally agree that it could be a net benefit if <a href="/wiki/Productivity" title="Productivity">productivity</a> gains are <a href="/wiki/Redistribution_of_income_and_wealth" title="Redistribution of income and wealth">redistributed</a>.<sup class="reference" id="cite_ref-FOOTNOTEIGM_Chicago2017_242-0"><a href="#cite_note-FOOTNOTEIGM_Chicago2017-242">[228]</a></sup> Risk estimates vary; for example, in the 2010s, Michael Osborne and <a href="/wiki/Carl_Benedikt_Frey" title="Carl Benedikt Frey">Carl Benedikt Frey</a> estimated 47% of U.S. jobs are at "high risk" of potential automation, while an OECD report classified only 9% of U.S. jobs as "high risk".<sup class="reference" id="cite_ref-244"><a href="#cite_note-244">[o]</a></sup><sup class="reference" id="cite_ref-245"><a href="#cite_note-245">[230]</a></sup> The methodology of speculating about future employment levels has been criticised as lacking evidential foundation, and for implying that technology, rather than social policy, creates unemployment, as opposed to redundancies.<sup class="reference" id="cite_ref-auto1_240-1"><a href="#cite_note-auto1-240">[226]</a></sup> In April 2023, it was reported that 70% of the jobs for Chinese video game illustrators had been eliminated by generative artificial intelligence.<sup class="reference" id="cite_ref-246"><a href="#cite_note-246">[231]</a></sup><sup class="reference" id="cite_ref-247"><a href="#cite_note-247">[232]</a></sup>
</p><p>Unlike previous waves of automation, many middle-class jobs may be eliminated by artificial intelligence; <i><a href="/wiki/The_Economist" title="The Economist">The Economist</a></i> stated in 2015 that "the worry that AI could do to white-collar jobs what steam power did to blue-collar ones during the Industrial Revolution" is "worth taking seriously".<sup class="reference" id="cite_ref-FOOTNOTEMorgenstern2015_248-0"><a href="#cite_note-FOOTNOTEMorgenstern2015-248">[233]</a></sup> Jobs at extreme risk range from <a href="/wiki/Paralegal" title="Paralegal">paralegals</a> to fast food cooks, while job demand is likely to increase for care-related professions ranging from personal healthcare to the clergy.<sup class="reference" id="cite_ref-249"><a href="#cite_note-249">[234]</a></sup>
</p><p>From the early days of the development of artificial intelligence, there have been arguments, for example, those put forward by <a href="/wiki/Joseph_Weizenbaum" title="Joseph Weizenbaum">Joseph Weizenbaum</a>, about whether tasks that can be done by computers actually should be done by them, given the difference between computers and humans, and between quantitative calculation and qualitative, value-based judgement.<sup class="reference" id="cite_ref-250"><a href="#cite_note-250">[235]</a></sup>
</p><p>It has been argued AI will become so powerful that humanity may irreversibly lose control of it.
This could, as physicist <a href="/wiki/Stephen_Hawking" title="Stephen Hawking">Stephen Hawking</a> stated, "<a href="/wiki/Global_catastrophic_risk" title="Global catastrophic risk">spell the end of the human race</a>".<sup class="reference" id="cite_ref-FOOTNOTECellan-Jones2014_251-0"><a href="#cite_note-FOOTNOTECellan-Jones2014-251">[236]</a></sup> This scenario has been common in science fiction, when a computer or robot suddenly develops a human-like "self-awareness" (or "sentience" or "consciousness") and becomes a malevolent character.<sup class="reference" id="cite_ref-253"><a href="#cite_note-253">[p]</a></sup> These sci-fi scenarios are misleading in several ways.
</p><p>First, AI does not require human-like "<a href="/wiki/Sentience" title="Sentience">sentience</a>" to be an existential risk.
Modern AI programs are given specific goals and use learning and intelligence to achieve them.
Philosopher <a href="/wiki/Nick_Bostrom" title="Nick Bostrom">Nick Bostrom</a> argued that if one gives <i>almost any</i> goal to a sufficiently powerful AI, it may choose to destroy humanity to achieve it (he used the example of a <a href="/wiki/Instrumental_convergence#Paperclip_maximizer" title="Instrumental convergence">paperclip factory manager</a>).<sup class="reference" id="cite_ref-FOOTNOTEBostrom2014_254-0"><a href="#cite_note-FOOTNOTEBostrom2014-254">[238]</a></sup> <a href="/wiki/Stuart_J._Russell" title="Stuart J. Russell">Stuart Russell</a> gives the example of household robot that tries to find a way to kill its owner to prevent it from being unplugged, reasoning that "you can't fetch the coffee if you're dead.
"<sup class="reference" id="cite_ref-FOOTNOTERussell2019_255-0"><a href="#cite_note-FOOTNOTERussell2019-255">[239]</a></sup> In order to be safe for humanity, a <a href="/wiki/Superintelligence" title="Superintelligence">superintelligence</a> would have to be genuinely <a href="/wiki/AI_alignment" title="AI alignment">aligned</a> with humanity's morality and values so that it is "fundamentally on our side".<sup class="reference" id="cite_ref-256"><a href="#cite_note-256">[240]</a></sup>
</p><p>Second, <a href="/wiki/Yuval_Noah_Harari" title="Yuval Noah Harari">Yuval Noah Harari</a> argues that AI does not require a robot body or physical control to pose an existential risk.
The essential parts of civilization are not physical.
Things like <a href="/wiki/Ideology" title="Ideology">ideologies</a>, <a href="/wiki/Law" title="Law">law</a>, <a href="/wiki/Government" title="Government">government</a>, <a href="/wiki/Money" title="Money">money</a> and the <a href="/wiki/Economy" title="Economy">economy</a> are made of <a href="/wiki/Language" title="Language">language</a>; they exist because there are stories that billions of people believe.
The current prevalence of <a href="/wiki/Misinformation" title="Misinformation">misinformation</a> suggests that an AI could use language to convince people to believe anything, even to take actions that are destructive.<sup class="reference" id="cite_ref-FOOTNOTEHarari2023_257-0"><a href="#cite_note-FOOTNOTEHarari2023-257">[241]</a></sup>
</p><p>The opinions amongst experts and industry insiders are mixed, with sizable fractions both concerned and unconcerned by risk from eventual superintelligent AI.<sup class="reference" id="cite_ref-FOOTNOTEMüllerBostrom2014_258-0"><a href="#cite_note-FOOTNOTEMüllerBostrom2014-258">[242]</a></sup> Personalities such as <a href="/wiki/Stephen_Hawking" title="Stephen Hawking">Stephen Hawking</a>, <a href="/wiki/Bill_Gates" title="Bill Gates">Bill Gates</a>, and <a href="/wiki/Elon_Musk" title="Elon Musk">Elon Musk</a>,<sup class="reference" id="cite_ref-259"><a href="#cite_note-259">[243]</a></sup> as well as AI pioneers such as <a href="/wiki/Yoshua_Bengio" title="Yoshua Bengio">Yoshua Bengio</a>, <a href="/wiki/Stuart_J._Russell" title="Stuart J. Russell">Stuart Russell</a>, <a href="/wiki/Demis_Hassabis" title="Demis Hassabis">Demis Hassabis</a>, and <a href="/wiki/Sam_Altman" title="Sam Altman">Sam Altman</a>, have expressed concerns about existential risk from AI.
</p><p>In May 2023, <a href="/wiki/Geoffrey_Hinton" title="Geoffrey Hinton">Geoffrey Hinton</a> announced his resignation from Google in order to be able to "freely speak out about the risks of AI" without "considering how this impacts Google.
"<sup class="reference" id="cite_ref-:0_260-0"><a href="#cite_note-:0-260">[244]</a></sup> He notably mentioned risks of an <a href="/wiki/AI_takeover" title="AI takeover">AI takeover</a>,<sup class="reference" id="cite_ref-261"><a href="#cite_note-261">[245]</a></sup> and stressed that in order to avoid the worst outcomes, establishing safety guidelines will require cooperation among those competing in use of AI.<sup class="reference" id="cite_ref-262"><a href="#cite_note-262">[246]</a></sup>
</p><p>In 2023, many leading AI experts issued <a href="/wiki/Statement_on_AI_risk_of_extinction" title="Statement on AI risk of extinction">the joint statement</a> that "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war".<sup class="reference" id="cite_ref-FOOTNOTEValance2023_263-0"><a href="#cite_note-FOOTNOTEValance2023-263">[247]</a></sup>
</p><p>Other researchers, however, spoke in favor of a less dystopian view.
AI pioneer <a class="mw-redirect" href="/wiki/Juergen_Schmidhuber" title="Juergen Schmidhuber">Juergen Schmidhuber</a> did not sign the joint statement, emphasising that in 95% of all cases, AI research is about making "human lives longer and healthier and easier.
"<sup class="reference" id="cite_ref-guardian2023_264-0"><a href="#cite_note-guardian2023-264">[248]</a></sup> While the tools that are now being used to improve lives can also be used by bad actors, "they can also be used against the bad actors.
"<sup class="reference" id="cite_ref-foxnews2023_265-0"><a href="#cite_note-foxnews2023-265">[249]</a></sup><sup class="reference" id="cite_ref-forbes2023_266-0"><a href="#cite_note-forbes2023-266">[250]</a></sup> <a href="/wiki/Andrew_Ng" title="Andrew Ng">Andrew Ng</a> also argued that "it's a mistake to fall for the doomsday hype on AI—and that regulators who do will only benefit vested interests.
"<sup class="reference" id="cite_ref-andrewng2023_267-0"><a href="#cite_note-andrewng2023-267">[251]</a></sup> <a href="/wiki/Yann_LeCun" title="Yann LeCun">Yann LeCun</a> "scoffs at his peers' dystopian scenarios of supercharged misinformation and even, eventually, human extinction.
"<sup class="reference" id="cite_ref-lecun2023_268-0"><a href="#cite_note-lecun2023-268">[252]</a></sup> In the early 2010s, experts argued that the risks are too distant in the future to warrant research or that humans will be valuable from the perspective of a superintelligent machine.<sup class="reference" id="cite_ref-269"><a href="#cite_note-269">[253]</a></sup> However, after 2016, the study of current and future risks and possible solutions became a serious area of research.<sup class="reference" id="cite_ref-FOOTNOTEChristian202067,_73_270-0"><a href="#cite_note-FOOTNOTEChristian202067,_73-270">[254]</a></sup>
</p><p>Friendly AI are machines that have been designed from the beginning to minimize risks and to make choices that benefit humans.
<a href="/wiki/Eliezer_Yudkowsky" title="Eliezer Yudkowsky">Eliezer Yudkowsky</a>, who coined the term, argues that developing friendly AI should be a higher research priority: it may require a large investment and it must be completed before AI becomes an existential risk.<sup class="reference" id="cite_ref-FOOTNOTEYudkowsky2008_271-0"><a href="#cite_note-FOOTNOTEYudkowsky2008-271">[255]</a></sup>
</p><p>Machines with intelligence have the potential to use their intelligence to make ethical decisions.
The field of machine ethics provides machines with ethical principles and procedures for resolving ethical dilemmas.<sup class="reference" id="cite_ref-FOOTNOTEAndersonAnderson2011_272-0"><a href="#cite_note-FOOTNOTEAndersonAnderson2011-272">[256]</a></sup>
The field of machine ethics is also called computational morality,<sup class="reference" id="cite_ref-FOOTNOTEAndersonAnderson2011_272-1"><a href="#cite_note-FOOTNOTEAndersonAnderson2011-272">[256]</a></sup>
and was founded at an <a class="mw-redirect" href="/wiki/AAAI" title="AAAI">AAAI</a> symposium in 2005.<sup class="reference" id="cite_ref-FOOTNOTEAAAI2014_273-0"><a href="#cite_note-FOOTNOTEAAAI2014-273">[257]</a></sup>
</p><p>Other approaches include <a href="/wiki/Wendell_Wallach" title="Wendell Wallach">Wendell Wallach</a>'s "artificial moral agents"<sup class="reference" id="cite_ref-FOOTNOTEWallach2010_274-0"><a href="#cite_note-FOOTNOTEWallach2010-274">[258]</a></sup> and <a href="/wiki/Stuart_J._Russell" title="Stuart J. Russell">Stuart J. Russell</a>'s <a href="/wiki/Human_Compatible#Russell's_three_principles" title="Human Compatible">three principles</a> for developing provably beneficial machines.<sup class="reference" id="cite_ref-FOOTNOTERussell2019173_275-0"><a href="#cite_note-FOOTNOTERussell2019173-275">[259]</a></sup>
</p><p>Active organizations in the AI open-source community include <a href="/wiki/Hugging_Face" title="Hugging Face">Hugging Face</a>,<sup class="reference" id="cite_ref-276"><a href="#cite_note-276">[260]</a></sup> <a href="/wiki/Google" title="Google">Google</a>,<sup class="reference" id="cite_ref-277"><a href="#cite_note-277">[261]</a></sup> <a href="/wiki/EleutherAI" title="EleutherAI">EleutherAI</a> and <a href="/wiki/Meta_Platforms" title="Meta Platforms">Meta</a>.<sup class="reference" id="cite_ref-278"><a href="#cite_note-278">[262]</a></sup> Various AI models, such as <a class="mw-redirect" href="/wiki/LLaMA" title="LLaMA">Llama 2</a>, <a href="/wiki/Mistral_AI" title="Mistral AI">Mistral</a> or <a href="/wiki/Stable_Diffusion" title="Stable Diffusion">Stable Diffusion</a>, have been made open-weight,<sup class="reference" id="cite_ref-279"><a href="#cite_note-279">[263]</a></sup><sup class="reference" id="cite_ref-280"><a href="#cite_note-280">[264]</a></sup> meaning that their architecture and trained parameters (the "weights") are publicly available.
Open-weight models can be freely <a href="/wiki/Fine-tuning_(deep_learning)" title="Fine-tuning (deep learning)">fine-tuned</a>, which allows companies to specialize them with their own data and for their own use-case.<sup class="reference" id="cite_ref-281"><a href="#cite_note-281">[265]</a></sup> Open-weight models are useful for research and innovation but can also be misused.
Since they can be fine-tuned, any built-in security measure, such as objecting to harmful requests, can be trained away until it becomes ineffective.
Some researchers warn that future AI models may develop dangerous capabilities (such as the potential to drastically facilitate <a href="/wiki/Bioterrorism" title="Bioterrorism">bioterrorism</a>) and that once released on the Internet, they can't be deleted everywhere if needed.
They recommend pre-release audits and cost-benefit analyses.<sup class="reference" id="cite_ref-282"><a href="#cite_note-282">[266]</a></sup>
</p><p>Artificial Intelligence projects can have their ethical permissibility tested while designing, developing, and implementing an AI system.
An AI framework such as the Care and Act Framework containing the SUM values—developed by the <a href="/wiki/Alan_Turing_Institute" title="Alan Turing Institute">Alan Turing Institute</a> tests projects in four main areas:<sup class="reference" id="cite_ref-283"><a href="#cite_note-283">[267]</a></sup><sup class="reference" id="cite_ref-284"><a href="#cite_note-284">[268]</a></sup>
</p><p>Other developments in ethical frameworks include those decided upon during the <a href="/wiki/Asilomar_Conference_on_Beneficial_AI" title="Asilomar Conference on Beneficial AI">Asilomar Conference</a>, the Montreal Declaration for Responsible AI, and the IEEE's Ethics of Autonomous Systems initiative, among others;<sup class="reference" id="cite_ref-285"><a href="#cite_note-285">[269]</a></sup> however, these principles do not go without their criticisms, especially regards to the people chosen contributes to these frameworks.<sup class="reference" id="cite_ref-286"><a href="#cite_note-286">[270]</a></sup>
</p><p>Promotion of the wellbeing of the people and communities that these technologies affect requires consideration of the social and ethical implications at all stages of AI system design, development and implementation, and collaboration between job roles such as data scientists, product managers, data engineers, domain experts, and delivery managers.<sup class="reference" id="cite_ref-287"><a href="#cite_note-287">[271]</a></sup>
</p><p>The <a class="mw-redirect" href="/wiki/AI_Safety_Institute_(United_Kingdom)" title="AI Safety Institute (United Kingdom)">UK AI Safety Institute</a> released in 2024 a testing toolset called 'Inspect' for AI safety evaluations available under a MIT open-source licence which is freely available on GitHub and can be improved with third-party packages.
It can be used to evaluate AI models in a range of areas including core knowledge, ability to reason, and autonomous capabilities.<sup class="reference" id="cite_ref-288"><a href="#cite_note-288">[272]</a></sup>
</p><p>The regulation of artificial intelligence is the development of public sector policies and laws for promoting and regulating artificial intelligence (AI); it is therefore related to the broader regulation of algorithms.<sup class="reference" id="cite_ref-289"><a href="#cite_note-289">[273]</a></sup> The regulatory and policy landscape for AI is an emerging issue in jurisdictions globally.<sup class="reference" id="cite_ref-FOOTNOTELaw_Library_of_Congress_(U.S.)._Global_Legal_Research_Directorate2019_290-0"><a href="#cite_note-FOOTNOTELaw_Library_of_Congress_(U.S.)._Global_Legal_Research_Directorate2019-290">[274]</a></sup> According to AI Index at <a class="mw-redirect" href="/wiki/Stanford" title="Stanford">Stanford</a>, the annual number of AI-related laws passed in the 127 survey countries jumped from one passed in 2016 to 37 passed in 2022 alone.<sup class="reference" id="cite_ref-FOOTNOTEVincent2023_291-0"><a href="#cite_note-FOOTNOTEVincent2023-291">[275]</a></sup><sup class="reference" id="cite_ref-FOOTNOTEStanford_University2023_292-0"><a href="#cite_note-FOOTNOTEStanford_University2023-292">[276]</a></sup> Between 2016 and 2020, more than 30 countries adopted dedicated strategies for AI.<sup class="reference" id="cite_ref-FOOTNOTEUNESCO2021_293-0"><a href="#cite_note-FOOTNOTEUNESCO2021-293">[277]</a></sup> Most EU member states had released national AI strategies, as had Canada, China, India, Japan, Mauritius, the Russian Federation, Saudi Arabia, United Arab Emirates, U.S., and Vietnam.
Others were in the process of elaborating their own AI strategy, including Bangladesh, Malaysia and Tunisia.<sup class="reference" id="cite_ref-FOOTNOTEUNESCO2021_293-1"><a href="#cite_note-FOOTNOTEUNESCO2021-293">[277]</a></sup> The <a href="/wiki/Global_Partnership_on_Artificial_Intelligence" title="Global Partnership on Artificial Intelligence">Global Partnership on Artificial Intelligence</a> was launched in June 2020, stating a need for AI to be developed in accordance with human rights and democratic values, to ensure public confidence and trust in the technology.<sup class="reference" id="cite_ref-FOOTNOTEUNESCO2021_293-2"><a href="#cite_note-FOOTNOTEUNESCO2021-293">[277]</a></sup> <a href="/wiki/Henry_Kissinger" title="Henry Kissinger">Henry Kissinger</a>, <a href="/wiki/Eric_Schmidt" title="Eric Schmidt">Eric Schmidt</a>, and <a class="mw-redirect" href="/wiki/Daniel_P._Huttenlocher" title="Daniel P. Huttenlocher">Daniel Huttenlocher</a> published a joint statement in November 2021 calling for a government commission to regulate AI.<sup class="reference" id="cite_ref-FOOTNOTEKissinger2021_294-0"><a href="#cite_note-FOOTNOTEKissinger2021-294">[278]</a></sup> In 2023, OpenAI leaders published recommendations for the governance of superintelligence, which they believe may happen in less than 10 years.<sup class="reference" id="cite_ref-FOOTNOTEAltmanBrockmanSutskever2023_295-0"><a href="#cite_note-FOOTNOTEAltmanBrockmanSutskever2023-295">[279]</a></sup> In 2023, the United Nations also launched an advisory body to provide recommendations on AI governance; the body comprises technology company executives, governments officials and academics.<sup class="reference" id="cite_ref-296"><a href="#cite_note-296">[280]</a></sup>
</p><p>In a 2022 <a href="/wiki/Ipsos" title="Ipsos">Ipsos</a> survey, attitudes towards AI varied greatly by country; 78% of Chinese citizens, but only 35% of Americans, agreed that "products and services using AI have more benefits than drawbacks".<sup class="reference" id="cite_ref-FOOTNOTEVincent2023_291-1"><a href="#cite_note-FOOTNOTEVincent2023-291">[275]</a></sup> A 2023 <a href="/wiki/Reuters" title="Reuters">Reuters</a>/Ipsos poll found that 61% of Americans agree, and 22% disagree, that AI poses risks to humanity.<sup class="reference" id="cite_ref-FOOTNOTEEdwards2023_297-0"><a href="#cite_note-FOOTNOTEEdwards2023-297">[281]</a></sup> In a 2023 <a href="/wiki/Fox_News" title="Fox News">Fox News</a> poll, 35% of Americans thought it "very important", and an additional 41% thought it "somewhat important", for the federal government to regulate AI, versus 13% responding "not very important" and 8% responding "not at all important".<sup class="reference" id="cite_ref-FOOTNOTEKasperowicz2023_298-0"><a href="#cite_note-FOOTNOTEKasperowicz2023-298">[282]</a></sup><sup class="reference" id="cite_ref-FOOTNOTEFox_News2023_299-0"><a href="#cite_note-FOOTNOTEFox_News2023-299">[283]</a></sup>
</p><p>In November 2023, the first global <a class="mw-redirect" href="/wiki/2023_AI_Safety_Summit" title="2023 AI Safety Summit">AI Safety Summit</a> was held in <a href="/wiki/Bletchley_Park" title="Bletchley Park">Bletchley Park</a> in the UK to discuss the near and far term risks of AI and the possibility of mandatory and voluntary regulatory frameworks.<sup class="reference" id="cite_ref-300"><a href="#cite_note-300">[284]</a></sup> 28 countries including the United States, China, and the European Union issued a declaration at the start of the summit, calling for international co-operation to manage the challenges and risks of artificial intelligence.<sup class="reference" id="cite_ref-2023-11-01-bletchley-declaration-full_301-0"><a href="#cite_note-2023-11-01-bletchley-declaration-full-301">[285]</a></sup><sup class="reference" id="cite_ref-302"><a href="#cite_note-302">[286]</a></sup> In May 2024 at the <a href="/wiki/AI_Summit_Seoul" title="AI Summit Seoul">AI Seoul Summit</a>, 16 global AI tech companies agreed to safety commitments on the development of AI.<sup class="reference" id="cite_ref-303"><a href="#cite_note-303">[287]</a></sup><sup class="reference" id="cite_ref-304"><a href="#cite_note-304">[288]</a></sup>
</p><p>The study of mechanical or "formal" reasoning began with philosophers and mathematicians in antiquity.
The study of logic led directly to <a href="/wiki/Alan_Turing" title="Alan Turing">Alan Turing</a>'s <a href="/wiki/Theory_of_computation" title="Theory of computation">theory of computation</a>, which suggested that a machine, by shuffling symbols as simple as "0" and "1", could simulate any conceivable form of mathematical reasoning.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig20219_305-0"><a href="#cite_note-FOOTNOTERussellNorvig20219-305">[289]</a></sup><sup class="reference" id="cite_ref-turing_5-1"><a href="#cite_note-turing-5">[5]</a></sup> This, along with concurrent discoveries in <a href="/wiki/Cybernetics" title="Cybernetics">cybernetics</a>, <a href="/wiki/Information_theory" title="Information theory">information theory</a> and <a class="mw-redirect" href="/wiki/Neurobiology" title="Neurobiology">neurobiology</a>, led researchers to consider the possibility of building an "electronic brain".<sup class="reference" id="cite_ref-307"><a href="#cite_note-307">[q]</a></sup>
They developed several areas of research that would become part of AI,<sup class="reference" id="cite_ref-308"><a href="#cite_note-308">[291]</a></sup>
such as <a class="mw-redirect" href="/wiki/Warren_McCullouch" title="Warren McCullouch">McCullouch</a> and <a href="/wiki/Walter_Pitts" title="Walter Pitts">Pitts</a> design for "artificial neurons" in 1943,<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig202117_309-0"><a href="#cite_note-FOOTNOTERussellNorvig202117-309">[292]</a></sup> and Turing's influential 1950 paper '<a href="/wiki/Computing_Machinery_and_Intelligence" title="Computing Machinery and Intelligence">Computing Machinery and Intelligence</a>', which introduced the <a href="/wiki/Turing_test" title="Turing test">Turing test</a> and showed that "machine intelligence" was plausible.<sup class="reference" id="cite_ref-Turing_test_310-0"><a href="#cite_note-Turing_test-310">[293]</a></sup><sup class="reference" id="cite_ref-turing_5-2"><a href="#cite_note-turing-5">[5]</a></sup>
</p><p>The field of AI research was founded at <a href="/wiki/Dartmouth_workshop" title="Dartmouth workshop">a workshop</a> at <a href="/wiki/Dartmouth_College" title="Dartmouth College">Dartmouth College</a> in 1956.<sup class="reference" id="cite_ref-312"><a href="#cite_note-312">[r]</a></sup><sup class="reference" id="cite_ref-Dartmouth_workshop_6-1"><a href="#cite_note-Dartmouth_workshop-6">[6]</a></sup> The attendees became the leaders of AI research in the 1960s.<sup class="reference" id="cite_ref-314"><a href="#cite_note-314">[s]</a></sup> They and their students produced programs that the press described as "astonishing":<sup class="reference" id="cite_ref-316"><a href="#cite_note-316">[t]</a></sup> computers were learning <a class="mw-redirect" href="/wiki/Draughts" title="Draughts">checkers</a> strategies, solving word problems in algebra, proving <a href="/wiki/Theorem" title="Theorem">logical theorems</a> and speaking English.<sup class="reference" id="cite_ref-317"><a href="#cite_note-317">[u]</a></sup><sup class="reference" id="cite_ref-AI_in_the_60s_9-1"><a href="#cite_note-AI_in_the_60s-9">[9]</a></sup> Artificial intelligence laboratories were set up at a number of British and U.S. universities in the latter 1950s and early 1960s.<sup class="reference" id="cite_ref-turing_5-3"><a href="#cite_note-turing-5">[5]</a></sup>
</p><p>Researchers in the 1960s and the 1970s were convinced that their methods would eventually succeed in creating a machine with <a href="/wiki/Artificial_general_intelligence" title="Artificial general intelligence">general intelligence</a> and considered this the goal of their field.<sup class="reference" id="cite_ref-FOOTNOTENewquist199486–86_318-0"><a href="#cite_note-FOOTNOTENewquist199486–86-318">[297]</a></sup> <a href="/wiki/Herbert_A._Simon" title="Herbert A. Simon">Herbert Simon</a> predicted, "machines will be capable, within twenty years, of doing any work a man can do".<sup class="reference" id="cite_ref-319"><a href="#cite_note-319">[298]</a></sup> <a href="/wiki/Marvin_Minsky" title="Marvin Minsky">Marvin Minsky</a> agreed, writing, "within a generation ... the problem of creating 'artificial intelligence' will substantially be solved".<sup class="reference" id="cite_ref-320"><a href="#cite_note-320">[299]</a></sup> They had, however, underestimated the difficulty of the problem.<sup class="reference" id="cite_ref-322"><a href="#cite_note-322">[v]</a></sup> In 1974, both the U.S. and British governments cut off exploratory research in response to the <a href="/wiki/Lighthill_report" title="Lighthill report">criticism</a> of <a class="mw-redirect" href="/wiki/Sir_James_Lighthill" title="Sir James Lighthill">Sir James Lighthill</a><sup class="reference" id="cite_ref-FOOTNOTELighthill1973_323-0"><a href="#cite_note-FOOTNOTELighthill1973-323">[301]</a></sup> and ongoing pressure from the U.S. Congress to <a class="mw-redirect" href="/wiki/Mansfield_Amendment" title="Mansfield Amendment">fund more productive projects</a>.<sup class="reference" id="cite_ref-FOOTNOTENRC1999212–213_324-0"><a href="#cite_note-FOOTNOTENRC1999212–213-324">[302]</a></sup> <a href="/wiki/Marvin_Minsky" title="Marvin Minsky">Minsky</a>'s and <a href="/wiki/Seymour_Papert" title="Seymour Papert">Papert</a>'s book <i><a href="/wiki/Perceptron" title="Perceptron">Perceptrons</a></i> was understood as proving that <a class="mw-redirect" href="/wiki/Artificial_neural_networks" title="Artificial neural networks">artificial neural networks</a> would never be useful for solving real-world tasks, thus discrediting the approach altogether.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig202122_325-0"><a href="#cite_note-FOOTNOTERussellNorvig202122-325">[303]</a></sup> The "<a href="/wiki/AI_winter" title="AI winter">AI winter</a>", a period when obtaining funding for AI projects was difficult, followed.<sup class="reference" id="cite_ref-First_AI_winter_11-1"><a href="#cite_note-First_AI_winter-11">[11]</a></sup>
</p><p>In the early 1980s, AI research was revived by the commercial success of <a href="/wiki/Expert_system" title="Expert system">expert systems</a>,<sup class="reference" id="cite_ref-326"><a href="#cite_note-326">[304]</a></sup> a form of AI program that simulated the knowledge and analytical skills of human experts.
By 1985, the market for AI had reached over a billion dollars.
At the same time, Japan's <a class="mw-redirect" href="/wiki/Fifth_generation_computer" title="Fifth generation computer">fifth generation computer</a> project inspired the U.S. and British governments to restore funding for <a class="mw-redirect" href="/wiki/Academic_research" title="Academic research">academic research</a>.<sup class="reference" id="cite_ref-AI_in_the_80s_10-1"><a href="#cite_note-AI_in_the_80s-10">[10]</a></sup> However, beginning with the collapse of the <a class="mw-redirect" href="/wiki/Lisp_Machine" title="Lisp Machine">Lisp Machine</a> market in 1987, AI once again fell into disrepute, and a second, longer-lasting winter began.<sup class="reference" id="cite_ref-Second_AI_winter_12-1"><a href="#cite_note-Second_AI_winter-12">[12]</a></sup>
</p><p>Up to this point, most of AI's funding had gone to projects that used high-level <a class="mw-redirect" href="/wiki/Symbolic_AI" title="Symbolic AI">symbols</a> to represent <a class="mw-redirect" href="/wiki/Mental_objects" title="Mental objects">mental objects</a> like plans, goals, beliefs, and known facts.
In the 1980s, some researchers began to doubt that this approach would be able to imitate all the processes of human cognition, especially <a href="/wiki/Machine_perception" title="Machine perception">perception</a>, <a href="/wiki/Robotics" title="Robotics">robotics</a>, <a href="/wiki/Machine_learning" title="Machine learning">learning</a> and <a href="/wiki/Pattern_recognition" title="Pattern recognition">pattern recognition</a>,<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig202124_327-0"><a href="#cite_note-FOOTNOTERussellNorvig202124-327">[305]</a></sup> and began to look into "sub-symbolic" approaches.<sup class="reference" id="cite_ref-FOOTNOTENilsson19987_328-0"><a href="#cite_note-FOOTNOTENilsson19987-328">[306]</a></sup> <a href="/wiki/Rodney_Brooks" title="Rodney Brooks">Rodney Brooks</a> rejected "representation" in general and focussed directly on engineering machines that move and survive.<sup class="reference" id="cite_ref-333"><a href="#cite_note-333">[w]</a></sup> <a href="/wiki/Judea_Pearl" title="Judea Pearl">Judea Pearl</a>, <a class="mw-redirect" href="/wiki/Lofti_Zadeh" title="Lofti Zadeh">Lofti Zadeh</a> and others developed methods that handled incomplete and uncertain information by making reasonable guesses rather than precise logic.<sup class="reference" id="cite_ref-Uncertain_reasoning_97-1"><a href="#cite_note-Uncertain_reasoning-97">[91]</a></sup><sup class="reference" id="cite_ref-FOOTNOTERussellNorvig202125_334-0"><a href="#cite_note-FOOTNOTERussellNorvig202125-334">[311]</a></sup> But the most important development was the revival of "<a href="/wiki/Connectionism" title="Connectionism">connectionism</a>", including neural network research, by <a href="/wiki/Geoffrey_Hinton" title="Geoffrey Hinton">Geoffrey Hinton</a> and others.<sup class="reference" id="cite_ref-335"><a href="#cite_note-335">[312]</a></sup> In 1990, <a href="/wiki/Yann_LeCun" title="Yann LeCun">Yann LeCun</a> successfully showed that <a class="mw-redirect" href="/wiki/Convolutional_neural_networks" title="Convolutional neural networks">convolutional neural networks</a> can recognize handwritten digits, the first of many successful applications of neural networks.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig202126_336-0"><a href="#cite_note-FOOTNOTERussellNorvig202126-336">[313]</a></sup>
</p><p>AI gradually restored its reputation in the late 1990s and early 21st century by exploiting formal mathematical methods and by finding specific solutions to specific problems.
This "<a class="mw-redirect" href="/wiki/Narrow_AI" title="Narrow AI">narrow</a>" and "formal" focus allowed researchers to produce verifiable results and collaborate with other fields (such as <a href="/wiki/Statistics" title="Statistics">statistics</a>, <a href="/wiki/Economics" title="Economics">economics</a> and <a href="/wiki/Mathematical_optimization" title="Mathematical optimization">mathematics</a>).<sup class="reference" id="cite_ref-AI_1990s_337-0"><a href="#cite_note-AI_1990s-337">[314]</a></sup> By 2000, solutions developed by AI researchers were being widely used, although in the 1990s they were rarely described as "artificial intelligence".<sup class="reference" id="cite_ref-AI_widely_used_1990s_338-0"><a href="#cite_note-AI_widely_used_1990s-338">[315]</a></sup>
However, several academic researchers became concerned that AI was no longer pursuing its original goal of creating versatile, fully intelligent machines.
Beginning around 2002, they founded the subfield of <a href="/wiki/Artificial_general_intelligence" title="Artificial general intelligence">artificial general intelligence</a> (or "AGI"), which had several well-funded institutions by the 2010s.<sup class="reference" id="cite_ref-AGI_17-2"><a href="#cite_note-AGI-17">[16]</a></sup>
</p><p><a href="/wiki/Deep_learning" title="Deep learning">Deep learning</a> began to dominate industry benchmarks in 2012 and was adopted throughout the field.<sup class="reference" id="cite_ref-Deep_learning_revolution_13-1"><a href="#cite_note-Deep_learning_revolution-13">[13]</a></sup>
For many specific tasks, other methods were abandoned.<sup class="reference" id="cite_ref-340"><a href="#cite_note-340">[x]</a></sup>
Deep learning's success was based on both hardware improvements (<a href="/wiki/Moore%27s_law" title="Moore's law">faster computers</a>,<sup class="reference" id="cite_ref-Moore's_Law_341-0"><a href="#cite_note-Moore's_Law-341">[317]</a></sup> <a href="/wiki/Graphics_processing_unit" title="Graphics processing unit">graphics processing units</a>, <a href="/wiki/Cloud_computing" title="Cloud computing">cloud computing</a><sup class="reference" id="cite_ref-FOOTNOTEClark2015b_342-0"><a href="#cite_note-FOOTNOTEClark2015b-342">[318]</a></sup>) and access to <a href="/wiki/Big_data" title="Big data">large amounts of data</a><sup class="reference" id="cite_ref-Big_data_343-0"><a href="#cite_note-Big_data-343">[319]</a></sup> (including curated datasets,<sup class="reference" id="cite_ref-FOOTNOTEClark2015b_342-1"><a href="#cite_note-FOOTNOTEClark2015b-342">[318]</a></sup> such as <a href="/wiki/ImageNet" title="ImageNet">ImageNet</a>).
Deep learning's success led to an enormous increase in interest and funding in AI.<sup class="reference" id="cite_ref-344"><a href="#cite_note-344">[y]</a></sup> The amount of machine learning research (measured by total publications) increased by 50% in the years 2015–2019.<sup class="reference" id="cite_ref-FOOTNOTEUNESCO2021_293-3"><a href="#cite_note-FOOTNOTEUNESCO2021-293">[277]</a></sup>
</p><p>In 2016, issues of <a class="mw-redirect" href="/wiki/Algorithmic_fairness" title="Algorithmic fairness">fairness</a> and the misuse of technology were catapulted into center stage at machine learning conferences, publications vastly increased, funding became available, and many researchers re-focussed their careers on these issues.
The <a href="/wiki/AI_alignment" title="AI alignment">alignment problem</a> became a serious field of academic study.<sup class="reference" id="cite_ref-FOOTNOTEChristian202067,_73_270-1"><a href="#cite_note-FOOTNOTEChristian202067,_73-270">[254]</a></sup>
</p><p>In the late teens and early 2020s, <a href="/wiki/Artificial_general_intelligence" title="Artificial general intelligence">AGI</a> companies began to deliver programs that created enormous interest.
In 2015, <a href="/wiki/AlphaGo" title="AlphaGo">AlphaGo</a>, developed by <a class="mw-redirect" href="/wiki/DeepMind" title="DeepMind">DeepMind</a>, beat the world champion <a class="mw-redirect" href="/wiki/Go_player" title="Go player">Go player</a>.
The program was taught only the rules of the game and developed strategy by itself.
<a href="/wiki/GPT-3" title="GPT-3">GPT-3</a> is a <a href="/wiki/Large_language_model" title="Large language model">large language model</a> that was released in 2020 by <a href="/wiki/OpenAI" title="OpenAI">OpenAI</a> and is capable of generating high-quality human-like text.<sup class="reference" id="cite_ref-345"><a href="#cite_note-345">[320]</a></sup> These programs, and others, inspired an aggressive <a href="/wiki/AI_boom" title="AI boom">AI boom</a>, where large companies began investing billions in AI research.
According to AI Impacts, about $50 billion annually was invested in "AI" around 2022 in the U.S. alone and about 20% of the new U.S. Computer Science PhD graduates have specialized in "AI".<sup class="reference" id="cite_ref-FOOTNOTEDiFeliciantonio2023_346-0"><a href="#cite_note-FOOTNOTEDiFeliciantonio2023-346">[321]</a></sup>
About 800,000 "AI"-related U.S. job openings existed in 2022.<sup class="reference" id="cite_ref-FOOTNOTEGoswami2023_347-0"><a href="#cite_note-FOOTNOTEGoswami2023-347">[322]</a></sup>
</p><p><a href="/wiki/Alan_Turing" title="Alan Turing">Alan Turing</a> wrote in 1950 "I propose to consider the question 'can machines think'?
"<sup class="reference" id="cite_ref-FOOTNOTETuring19501_348-0"><a href="#cite_note-FOOTNOTETuring19501-348">[323]</a></sup> He advised changing the question from whether a machine "thinks", to "whether or not it is possible for machinery to show intelligent behaviour".<sup class="reference" id="cite_ref-FOOTNOTETuring19501_348-1"><a href="#cite_note-FOOTNOTETuring19501-348">[323]</a></sup> He devised the Turing test, which measures the ability of a machine to simulate human conversation.<sup class="reference" id="cite_ref-Turing_test_310-1"><a href="#cite_note-Turing_test-310">[293]</a></sup> Since we can only observe the behavior of the machine, it does not matter if it is "actually" thinking or literally has a "mind".
Turing notes that <a href="/wiki/Problem_of_other_minds" title="Problem of other minds">we can not determine these things about other people</a> but "it is usual to have a polite convention that everyone thinks.
"<sup class="reference" id='cite_ref-FOOTNOTETuring1950Under_"The_Argument_from_Consciousness"_349-0'><a href='#cite_note-FOOTNOTETuring1950Under_"The_Argument_from_Consciousness"-349'>[324]</a></sup>
</p><p><a href="/wiki/Stuart_J._Russell" title="Stuart J. Russell">Russell</a> and <a href="/wiki/Peter_Norvig" title="Peter Norvig">Norvig</a> agree with Turing that intelligence must be defined in terms of external behavior, not internal structure.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig20211–4_1-1"><a href="#cite_note-FOOTNOTERussellNorvig20211–4-1">[1]</a></sup> However, they are critical that the test requires the machine to imitate humans.
"<a href="/wiki/Aeronautics" title="Aeronautics">Aeronautical engineering</a> texts," they wrote, "do not define the goal of their field as making 'machines that fly so exactly like <a class="mw-redirect" href="/wiki/Pigeon" title="Pigeon">pigeons</a> that they can fool other pigeons.<span style="padding-right:.15em;">'</span>"<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig20213_350-0"><a href="#cite_note-FOOTNOTERussellNorvig20213-350">[325]</a></sup> AI founder <a href="/wiki/John_McCarthy_(computer_scientist)" title="John McCarthy (computer scientist)">John McCarthy</a> agreed, writing that "Artificial intelligence is not, by definition, simulation of human intelligence".<sup class="reference" id="cite_ref-FOOTNOTEMaker2006_351-0"><a href="#cite_note-FOOTNOTEMaker2006-351">[326]</a></sup>
</p><p>McCarthy defines intelligence as "the computational part of the ability to achieve goals in the world".<sup class="reference" id="cite_ref-FOOTNOTEMcCarthy1999_352-0"><a href="#cite_note-FOOTNOTEMcCarthy1999-352">[327]</a></sup> Another AI founder, <a href="/wiki/Marvin_Minsky" title="Marvin Minsky">Marvin Minsky</a> similarly describes it as "the ability to solve hard problems".<sup class="reference" id="cite_ref-FOOTNOTEMinsky1986_353-0"><a href="#cite_note-FOOTNOTEMinsky1986-353">[328]</a></sup> The leading AI textbook defines it as the study of agents that perceive their environment and take actions that maximize their chances of achieving defined goals.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig20211–4_1-2"><a href="#cite_note-FOOTNOTERussellNorvig20211–4-1">[1]</a></sup> These definitions view intelligence in terms of well-defined problems with well-defined solutions, where both the difficulty of the problem and the performance of the program are direct measures of the "intelligence" of the machine—and no other philosophical discussion is required, or may not even be possible.
</p><p>Another definition has been adopted by Google,<sup class="reference" id="cite_ref-354"><a href="#cite_note-354">[329]</a></sup> a major practitioner in the field of AI.
This definition stipulates the ability of systems to synthesize information as the manifestation of intelligence, similar to the way it is defined in biological intelligence.
</p><p>Some authors have suggested in practice, that the definition of AI is vague and difficult to define, with contention as to whether classical algorithms should be categorised as AI,<sup class="reference" id="cite_ref-355"><a href="#cite_note-355">[330]</a></sup> with many companies during the early 2020s AI boom using the term as a marketing <a href="/wiki/Buzzword" title="Buzzword">buzzword</a>, often even if they did "not actually use AI in a material way".<sup class="reference" id="cite_ref-356"><a href="#cite_note-356">[331]</a></sup>
</p><p>No established unifying theory or <a href="/wiki/Paradigm" title="Paradigm">paradigm</a> has guided AI research for most of its history.<sup class="reference" id="cite_ref-358"><a href="#cite_note-358">[z]</a></sup> The unprecedented success of statistical machine learning in the 2010s eclipsed all other approaches (so much so that some sources, especially in the business world, use the term "artificial intelligence" to mean "machine learning with neural networks").
This approach is mostly <a class="mw-redirect" href="/wiki/Sub-symbolic" title="Sub-symbolic">sub-symbolic</a>, <a href="/wiki/Soft_computing" title="Soft computing">soft</a> and <a href="/wiki/Artificial_general_intelligence" title="Artificial general intelligence">narrow</a>.
Critics argue that these questions may have to be revisited by future generations of AI researchers.
</p><p><a href="/wiki/Symbolic_artificial_intelligence" title="Symbolic artificial intelligence">Symbolic AI</a> (or "<a href="/wiki/GOFAI" title="GOFAI">GOFAI</a>")<sup class="reference" id="cite_ref-FOOTNOTEHaugeland1985112–117_359-0"><a href="#cite_note-FOOTNOTEHaugeland1985112–117-359">[333]</a></sup> simulated the high-level conscious reasoning that people use when they solve puzzles, express legal reasoning and do mathematics.
They were highly successful at "intelligent" tasks such as algebra or IQ tests.
In the 1960s, Newell and Simon proposed the <a class="mw-redirect" href="/wiki/Physical_symbol_systems_hypothesis" title="Physical symbol systems hypothesis">physical symbol systems hypothesis</a>: "A physical symbol system has the necessary and sufficient means of general intelligent action.
"<sup class="reference" id="cite_ref-Physical_symbol_system_hypothesis_360-0"><a href="#cite_note-Physical_symbol_system_hypothesis-360">[334]</a></sup>
</p><p>However, the symbolic approach failed on many tasks that humans solve easily, such as learning, recognizing an object or commonsense reasoning.
<a href="/wiki/Moravec%27s_paradox" title="Moravec's paradox">Moravec's paradox</a> is the discovery that high-level "intelligent" tasks were easy for AI, but low level "instinctive" tasks were extremely difficult.<sup class="reference" id="cite_ref-361"><a href="#cite_note-361">[335]</a></sup> Philosopher <a href="/wiki/Hubert_Dreyfus" title="Hubert Dreyfus">Hubert Dreyfus</a> had <a class="mw-redirect" href="/wiki/Dreyfus%27_critique_of_AI" title="Dreyfus' critique of AI">argued</a> since the 1960s that human expertise depends on unconscious instinct rather than conscious symbol manipulation, and on having a "feel" for the situation, rather than explicit symbolic knowledge.<sup class="reference" id="cite_ref-Dreyfus'_critique_362-0"><a href="#cite_note-Dreyfus'_critique-362">[336]</a></sup> Although his arguments had been ridiculed and ignored when they were first presented, eventually, AI research came to agree with him.<sup class="reference" id="cite_ref-364"><a href="#cite_note-364">[aa]</a></sup><sup class="reference" id="cite_ref-Psychological_evidence_of_sub-symbolic_reasoning_23-2"><a href="#cite_note-Psychological_evidence_of_sub-symbolic_reasoning-23">[21]</a></sup>
</p><p>The issue is not resolved: <a class="mw-redirect" href="/wiki/Sub-symbolic" title="Sub-symbolic">sub-symbolic</a> reasoning can make many of the same inscrutable mistakes that human intuition does, such as <a href="/wiki/Algorithmic_bias" title="Algorithmic bias">algorithmic bias</a>.
Critics such as <a href="/wiki/Noam_Chomsky" title="Noam Chomsky">Noam Chomsky</a> argue continuing research into symbolic AI will still be necessary to attain general intelligence,<sup class="reference" id="cite_ref-FOOTNOTELangley2011_365-0"><a href="#cite_note-FOOTNOTELangley2011-365">[338]</a></sup><sup class="reference" id="cite_ref-FOOTNOTEKatz2012_366-0"><a href="#cite_note-FOOTNOTEKatz2012-366">[339]</a></sup> in part because sub-symbolic AI is a move away from <a class="mw-redirect" href="/wiki/Explainable_AI" title="Explainable AI">explainable AI</a>: it can be difficult or impossible to understand why a modern statistical AI program made a particular decision.
The emerging field of <a href="/wiki/Neuro-symbolic_AI" title="Neuro-symbolic AI">neuro-symbolic artificial intelligence</a> attempts to bridge the two approaches.
</p><p>"Neats" hope that intelligent behavior is described using simple, elegant principles (such as <a href="/wiki/Logic" title="Logic">logic</a>, <a class="mw-redirect" href="/wiki/Optimization_(mathematics)" title="Optimization (mathematics)">optimization</a>, or <a class="mw-redirect" href="/wiki/Artificial_neural_network" title="Artificial neural network">neural networks</a>).
"Scruffies" expect that it necessarily requires solving a large number of unrelated problems.
Neats defend their programs with theoretical rigor, scruffies rely mainly on incremental testing to see if they work.
This issue was actively discussed in the 1970s and 1980s,<sup class="reference" id="cite_ref-Neats_vs._scruffies_367-0"><a href="#cite_note-Neats_vs._scruffies-367">[340]</a></sup> but eventually was seen as irrelevant.
Modern AI has elements of both.
</p><p>Finding a provably correct or optimal solution is <a class="mw-redirect" href="/wiki/Intractability_(complexity)" title="Intractability (complexity)">intractable</a> for many important problems.<sup class="reference" id="cite_ref-Intractability_22-2"><a href="#cite_note-Intractability-22">[20]</a></sup> Soft computing is a set of techniques, including <a class="mw-redirect" href="/wiki/Genetic_algorithms" title="Genetic algorithms">genetic algorithms</a>, <a href="/wiki/Fuzzy_logic" title="Fuzzy logic">fuzzy logic</a> and neural networks, that are tolerant of imprecision, uncertainty, partial truth and approximation.
Soft computing was introduced in the late 1980s and most successful AI programs in the 21st century are examples of soft computing with neural networks.
</p><p>AI researchers are divided as to whether to pursue the goals of artificial general intelligence and <a href="/wiki/Superintelligence" title="Superintelligence">superintelligence</a> directly or to solve as many specific problems as possible (narrow AI) in hopes these solutions will lead indirectly to the field's long-term goals.<sup class="reference" id="cite_ref-FOOTNOTEPennachinGoertzel2007_368-0"><a href="#cite_note-FOOTNOTEPennachinGoertzel2007-368">[341]</a></sup><sup class="reference" id="cite_ref-FOOTNOTERoberts2016_369-0"><a href="#cite_note-FOOTNOTERoberts2016-369">[342]</a></sup> General intelligence is difficult to define and difficult to measure, and modern AI has had more verifiable successes by focusing on specific problems with specific solutions.
The experimental sub-field of artificial general intelligence studies this area exclusively.
</p><p>The <a href="/wiki/Philosophy_of_mind" title="Philosophy of mind">philosophy of mind</a> does not know whether a machine can have a <a href="/wiki/Mind" title="Mind">mind</a>, <a href="/wiki/Consciousness" title="Consciousness">consciousness</a> and <a href="/wiki/Philosophy_of_mind" title="Philosophy of mind">mental states</a>, in the same sense that human beings do.
This issue considers the internal experiences of the machine, rather than its external behavior.
Mainstream AI research considers this issue irrelevant because it does not affect the goals of the field: to build machines that can solve problems using intelligence.
<a href="/wiki/Stuart_J._Russell" title="Stuart J. Russell">Russell</a> and <a href="/wiki/Peter_Norvig" title="Peter Norvig">Norvig</a> add that "[t]he additional project of making a machine conscious in exactly the way humans are is not one that we are equipped to take on.
"<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig2021986_370-0"><a href="#cite_note-FOOTNOTERussellNorvig2021986-370">[343]</a></sup> However, the question has become central to the philosophy of mind.
It is also typically the central question at issue in <a href="/wiki/Artificial_intelligence_in_fiction" title="Artificial intelligence in fiction">artificial intelligence in fiction</a>.
</p><p><a href="/wiki/David_Chalmers" title="David Chalmers">David Chalmers</a> identified two problems in understanding the mind, which he named the "hard" and "easy" problems of consciousness.<sup class="reference" id="cite_ref-FOOTNOTEChalmers1995_371-0"><a href="#cite_note-FOOTNOTEChalmers1995-371">[344]</a></sup> The easy problem is understanding how the brain processes signals, makes plans and controls behavior.
The hard problem is explaining how this <i>feels</i> or why it should feel like anything at all, assuming we are right in thinking that it truly does feel like something (Dennett's consciousness illusionism says this is an illusion).
While human <a href="/wiki/Information_processing_(psychology)" title="Information processing (psychology)">information processing</a> is easy to explain, human <a class="mw-redirect" href="/wiki/Subjective_experience" title="Subjective experience">subjective experience</a> is difficult to explain.
For example, it is easy to imagine a color-blind person who has learned to identify which objects in their field of view are red, but it is not clear what would be required for the person to <i>know what red looks like</i>.<sup class="reference" id="cite_ref-FOOTNOTEDennett1991_372-0"><a href="#cite_note-FOOTNOTEDennett1991-372">[345]</a></sup>
</p><p>Computationalism is the position in the <a href="/wiki/Philosophy_of_mind" title="Philosophy of mind">philosophy of mind</a> that the human mind is an information processing system and that thinking is a form of computing.
Computationalism argues that the relationship between mind and body is similar or identical to the relationship between software and hardware and thus may be a solution to the <a href="/wiki/Mind%E2%80%93body_problem" title="Mind–body problem">mind–body problem</a>.
This philosophical position was inspired by the work of AI researchers and cognitive scientists in the 1960s and was originally proposed by philosophers <a href="/wiki/Jerry_Fodor" title="Jerry Fodor">Jerry Fodor</a> and <a href="/wiki/Hilary_Putnam" title="Hilary Putnam">Hilary Putnam</a>.<sup class="reference" id="cite_ref-FOOTNOTEHorst2005_373-0"><a href="#cite_note-FOOTNOTEHorst2005-373">[346]</a></sup>
</p><p>Philosopher <a href="/wiki/John_Searle" title="John Searle">John Searle</a> characterized this position as "<a class="mw-redirect" href="/wiki/Strong_AI_hypothesis" title="Strong AI hypothesis">strong AI</a>": "The appropriately programmed computer with the right inputs and outputs would thereby have a mind in exactly the same sense human beings have minds.
"<sup class="reference" id="cite_ref-Searle's_strong_AI_377-0"><a href="#cite_note-Searle's_strong_AI-377">[ab]</a></sup> Searle counters this assertion with his Chinese room argument, which attempts to show that, even if a machine perfectly simulates human behavior, there is still no reason to suppose it also has a mind.<sup class="reference" id="cite_ref-Chinese_room_378-0"><a href="#cite_note-Chinese_room-378">[350]</a></sup>
</p><p>It is difficult or impossible to reliably evaluate whether an advanced <a class="mw-redirect" href="/wiki/Sentient_AI" title="Sentient AI">AI is sentient</a> (has the ability to feel), and if so, to what degree.<sup class="reference" id="cite_ref-379"><a href="#cite_note-379">[351]</a></sup> But if there is a significant chance that a given machine can feel and suffer, then it may be entitled to certain rights or welfare protection measures, similarly to animals.<sup class="reference" id="cite_ref-:02_380-0"><a href="#cite_note-:02-380">[352]</a></sup><sup class="reference" id="cite_ref-:12_381-0"><a href="#cite_note-:12-381">[353]</a></sup> <a class="mw-redirect" href="/wiki/Sapience" title="Sapience">Sapience</a> (a set of capacities related to high intelligence, such as discernment or <a href="/wiki/Self-awareness" title="Self-awareness">self-awareness</a>) may provide another moral basis for AI rights.<sup class="reference" id="cite_ref-:02_380-1"><a href="#cite_note-:02-380">[352]</a></sup> <a class="mw-redirect" href="/wiki/Robot_rights" title="Robot rights">Robot rights</a> are also sometimes proposed as a practical way to integrate autonomous agents into society.<sup class="reference" id="cite_ref-382"><a href="#cite_note-382">[354]</a></sup>
</p><p>In 2017, the European Union considered granting "electronic personhood" to some of the most capable AI systems.
Similarly to the legal status of companies, it would have conferred rights but also responsibilities.<sup class="reference" id="cite_ref-383"><a href="#cite_note-383">[355]</a></sup> Critics argued in 2018 that granting rights to AI systems would downplay the importance of <a href="/wiki/Human_rights" title="Human rights">human rights</a>, and that legislation should focus on user needs rather than speculative futuristic scenarios.
They also noted that robots lacked the autonomy to take part to society on their own.<sup class="reference" id="cite_ref-384"><a href="#cite_note-384">[356]</a></sup><sup class="reference" id="cite_ref-385"><a href="#cite_note-385">[357]</a></sup>
</p><p>Progress in AI increased interest in the topic.
Proponents of AI welfare and rights often argue that AI sentience, if it emerges, would be particularly easy to deny.
They warn that this may be a <a href="/wiki/Moral_blindness" title="Moral blindness">moral blind spot</a> analogous to <a href="/wiki/Slavery" title="Slavery">slavery</a> or <a class="mw-redirect" href="/wiki/Factory_farming" title="Factory farming">factory farming</a>, which could lead to <a href="/wiki/Suffering_risks" title="Suffering risks">large-scale suffering</a> if sentient AI is created and carelessly exploited.<sup class="reference" id="cite_ref-:12_381-1"><a href="#cite_note-:12-381">[353]</a></sup><sup class="reference" id="cite_ref-:02_380-2"><a href="#cite_note-:02-380">[352]</a></sup>
</p><p>A <a href="/wiki/Superintelligence" title="Superintelligence">superintelligence</a> is a hypothetical agent that would possess intelligence far surpassing that of the brightest and most gifted human mind.<sup class="reference" id="cite_ref-FOOTNOTERoberts2016_369-1"><a href="#cite_note-FOOTNOTERoberts2016-369">[342]</a></sup>
</p><p>If research into <a href="/wiki/Artificial_general_intelligence" title="Artificial general intelligence">artificial general intelligence</a> produced sufficiently intelligent software, it might be able to <a href="/wiki/Recursive_self-improvement" title="Recursive self-improvement">reprogram and improve itself</a>.
The improved software would be even better at improving itself, leading to what <a href="/wiki/I._J._Good" title="I. J.
Good">I.
J. Good</a> called an "<a class="mw-redirect" href="/wiki/Intelligence_explosion" title="Intelligence explosion">intelligence explosion</a>" and <a href="/wiki/Vernor_Vinge" title="Vernor Vinge">Vernor Vinge</a> called a "<a href="/wiki/Technological_singularity" title="Technological singularity">singularity</a>".<sup class="reference" id="cite_ref-Singularity_386-0"><a href="#cite_note-Singularity-386">[358]</a></sup>
</p><p>However, technologies cannot improve exponentially indefinitely, and typically follow an <a class="mw-redirect" href="/wiki/S-shaped_curve" title="S-shaped curve">S-shaped curve</a>, slowing when they reach the physical limits of what the technology can do.<sup class="reference" id="cite_ref-FOOTNOTERussellNorvig20211005_387-0"><a href="#cite_note-FOOTNOTERussellNorvig20211005-387">[359]</a></sup>
</p><p>Robot designer <a href="/wiki/Hans_Moravec" title="Hans Moravec">Hans Moravec</a>, cyberneticist <a href="/wiki/Kevin_Warwick" title="Kevin Warwick">Kevin Warwick</a>, and inventor <a href="/wiki/Ray_Kurzweil" title="Ray Kurzweil">Ray Kurzweil</a> have predicted that humans and machines will merge in the future into <a href="/wiki/Cyborg" title="Cyborg">cyborgs</a> that are more capable and powerful than either.
This idea, called transhumanism, has roots in <a href="/wiki/Aldous_Huxley" title="Aldous Huxley">Aldous Huxley</a> and <a href="/wiki/Robert_Ettinger" title="Robert Ettinger">Robert Ettinger</a>.<sup class="reference" id="cite_ref-388"><a href="#cite_note-388">[360]</a></sup>
</p><p><a href="/wiki/Edward_Fredkin" title="Edward Fredkin">Edward Fredkin</a> argues that "artificial intelligence is the next stage in evolution", an idea first proposed by <a href="/wiki/Samuel_Butler_(novelist)" title="Samuel Butler (novelist)">Samuel Butler</a>'s "<a href="/wiki/Darwin_among_the_Machines" title="Darwin among the Machines">Darwin among the Machines</a>" as far back as 1863, and expanded upon by <a href="/wiki/George_Dyson_(science_historian)" title="George Dyson (science historian)">George Dyson</a> in his 1998 book <i><a class="mw-redirect" href="/wiki/Darwin_Among_the_Machines#Evolution_of_Global_Intelligence" title="Darwin Among the Machines">Darwin Among the Machines: The Evolution of Global Intelligence</a></i>.<sup class="reference" id="cite_ref-389"><a href="#cite_note-389">[361]</a></sup>
</p><p>Thought-capable artificial beings have appeared as storytelling devices since antiquity,<sup class="reference" id="cite_ref-AI_in_myth_390-0"><a href="#cite_note-AI_in_myth-390">[362]</a></sup> and have been a persistent theme in <a href="/wiki/Science_fiction" title="Science fiction">science fiction</a>.<sup class="reference" id="cite_ref-FOOTNOTEMcCorduck2004340–400_391-0"><a href="#cite_note-FOOTNOTEMcCorduck2004340–400-391">[363]</a></sup>
</p><p>A common <a href="/wiki/Trope_(literature)" title="Trope (literature)">trope</a> in these works began with <a href="/wiki/Mary_Shelley" title="Mary Shelley">Mary Shelley</a>'s <i><a href="/wiki/Frankenstein" title="Frankenstein">Frankenstein</a></i>, where a human creation becomes a threat to its masters.
This includes such works as <a href="/wiki/2001:_A_Space_Odyssey_(novel)" title="2001: A Space Odyssey (novel)">Arthur C. Clarke's</a> and <a class="mw-redirect" href="/wiki/2001:_A_Space_Odyssey_(film)" title="2001: A Space Odyssey (film)">Stanley Kubrick's</a> <i>2001: A Space Odyssey</i> (both 1968), with <a href="/wiki/HAL_9000" title="HAL 9000">HAL 9000</a>, the murderous computer in charge of the <i><a href="/wiki/Discovery_One" title="Discovery One">Discovery One</a></i> spaceship, as well as <i><a href="/wiki/The_Terminator" title="The Terminator">The Terminator</a></i> (1984) and <i><a href="/wiki/The_Matrix" title="The Matrix">The Matrix</a></i> (1999).
In contrast, the rare loyal robots such as Gort from <i><a href="/wiki/The_Day_the_Earth_Stood_Still" title="The Day the Earth Stood Still">The Day the Earth Stood Still</a></i> (1951) and Bishop from <i><a href="/wiki/Aliens_(film)" title="Aliens (film)">Aliens</a></i> (1986) are less prominent in popular culture.<sup class="reference" id="cite_ref-FOOTNOTEButtazzo2001_392-0"><a href="#cite_note-FOOTNOTEButtazzo2001-392">[364]</a></sup>
</p><p><a href="/wiki/Isaac_Asimov" title="Isaac Asimov">Isaac Asimov</a> introduced the <a href="/wiki/Three_Laws_of_Robotics" title="Three Laws of Robotics">Three Laws of Robotics</a> in many stories, most notably with the "<a href="/wiki/Multivac" title="Multivac">Multivac</a>" super-intelligent computer.
Asimov's laws are often brought up during lay discussions of machine ethics;<sup class="reference" id="cite_ref-FOOTNOTEAnderson2008_393-0"><a href="#cite_note-FOOTNOTEAnderson2008-393">[365]</a></sup> while almost all artificial intelligence researchers are familiar with Asimov's laws through popular culture, they generally consider the laws useless for many reasons, one of which is their ambiguity.<sup class="reference" id="cite_ref-FOOTNOTEMcCauley2007_394-0"><a href="#cite_note-FOOTNOTEMcCauley2007-394">[366]</a></sup>
</p><p>Several works use AI to force us to confront the fundamental question of what makes us human, showing us artificial beings that have <a href="/wiki/Sentience" title="Sentience">the ability to feel</a>, and thus to suffer.
This appears in <a href="/wiki/Karel_%C4%8Capek" title="Karel Čapek">Karel Čapek</a>'s <i><a href="/wiki/R.U.R."
title="R.U.R.
">R.U.R.</a></i>, the films <i><a href="/wiki/A.I._Artificial_Intelligence" title="A.I.
Artificial Intelligence">A.I.
Artificial Intelligence</a></i> and <i><a href="/wiki/Ex_Machina_(film)" title="Ex Machina (film)">Ex Machina</a></i>, as well as the novel <i><a href="/wiki/Do_Androids_Dream_of_Electric_Sheep%3F" title="Do Androids Dream of Electric Sheep?
">Do Androids Dream of Electric Sheep?</a></i>, by <a href="/wiki/Philip_K._Dick" title="Philip K. Dick">Philip K. Dick</a>.
Dick considers the idea that our understanding of human subjectivity is altered by technology created with artificial intelligence.<sup class="reference" id="cite_ref-FOOTNOTEGalvan1997_395-0"><a href="#cite_note-FOOTNOTEGalvan1997-395">[367]</a></sup>
</p><p>The two most widely used textbooks in 2023 (see the <a class="external text" href="https://explorer.opensyllabus.org/result/field?id=Computer+Science" rel="nofollow">Open Syllabus</a>):
</p><p>These were the four of the most widely used AI textbooks in 2008:
</p><p>Later editions:
</p>