atlas-of-ai.html.pm

#lang pollen

◊define-meta[page-title]{◊@{Book Review: ◊em{Atlas of AI}}}
◊define-meta[short-title]{◊@{◊em{Atlas of AI}}}
◊define-meta[original-date]{2021-08-22}
◊define-meta[edited-date]{2021-08-25}
◊define-meta[snippet]{A book review of Kate Crawford's 'Atlas of AI': This book situates AI in the physical and social world. Crawford shows us the ways that AI exploits and changes our relationships with the Earth and with each other.}

◊declare-work[#:id "Crawford" #:type "book" #:author "Kate Crawford"
#:title "Atlas of AI" #:year "2021" #:publisher "Yale University
Press" #:publisher-location "New Haven" #:url
"https://yalebooks.yale.edu/book/9780300209570/atlas-ai"]

◊small{Review of: ◊b{◊format-work[#:type "book" #:author "Kate Crawford" #:title
"Atlas of AI" #:year "2021" #:publisher "Yale University Press"
#:publisher-location "New Haven" #:url
"https://yalebooks.yale.edu/book/9780300209570/atlas-ai"]}}

◊fig[#:src "assets/atlas-of-ai.png" #:width "300px"]

◊em{Atlas of AI} is a timely book by Kate Crawford covering the
various implications of AI today and throughout history. There is
intense public scrutiny on the issues covered in this book: labour,
energy consumption, climate change, discrimination, privacy, and
government accountability. However, these concerns are often
overshadowed by the general enamourment with technological artifacts
like Amazon Alexa, Google Nest, Tesla, and mobile devices. ◊b{This
book situates AI in the physical and social world. Crawford shows us
the ways that AI exploits and changes our relationships with the Earth
and with each other.}

Crawford covers many of the canonical examples and issues flowing from
AI. You will see many of the examples from this book also referenced
in any current writing that views the field through a legal, social,
or political lens. The latter chapters are an especially good
presentation of AI as a tool of state power. While I found this to be
a valuable, scope-defining work for people newly thinking about the
consequences of AI use and development, I did find portions of it
repetitive, vague, or not well-connected to a central thesis. These
are easy enough to avoid or skim over.

◊heading{Summary}

This book opens with an image of two people holding up what appears to
be a mainframe computer on their shoulders, themselves crouched on top
of a fragment of the Earth. The computer is projecting, through a
lens, a distorted image of the Earth's surface. That projection sits
above the scene.

◊fig[#:src "assets/atlas-of-ai-projection.png" #:width "300px"]{The
opening image from ◊em{Atlas of AI}.}

This image neatly captures the themes in this book. AI exploits human
labour and resources extracted from the Earth. The conception of AI as
◊em{disembodied}, while sometimes a helpful abstraction---and
highlights one way in which AI might never be ◊em{Really
Intelligent}---serves to obscure the material needs and effects of
AI. This is the central theme of the first two chapters, "Earth" and
"Labor." The middle of the book focuses on the computer science of AI
and machine learning. AI, especially machine learning, transforms data
through its operation. It creates projections of reality into digital
representations. These are necessarily an approximation, with the goal
that the representation preserves features that are useful for the
intended task. In the image above, this approximation appears as a
◊a[#:href
"https://en.wikipedia.org/wiki/Bernard_J._S._Cahill"]{◊em{butterfly
map}}, a projection that presents more accurately the distances
between locations in northern latitudes. The projection exaggerates
the distance and introduces stark discontinuities between land masses
in the southern hemisphere. Of course all map projections onto a flat
surface must make tradeoffs, but Crawford's selection of this
particular projection reflects a theme developed in the latter
portions of this book: "jingoistic idea of sovereign
AI"◊note-cite["Crawford" #:pinpoint "208"] and ethics frameworks that
are "overwhelmingly produced by economically developed countries, with
little representation from Africa, South and Central America, or
Central Asia."◊note-cite["Crawford" #:pinpoint "224"]

◊sub-heading{Earth}

The first chapter highlights the physical damage to the Earth when we
extract the materials needed for our technology: e.g. lithium mines in
Nevada, rare earth elements often located in emerging markets and
conflict zones, and latex needed for transatlantic cables that is
extracted from a species of tree in Southest Asia. That is just some
of what is required to create our physical technological
tools. Crawford also highlights the ongoing energy needs to use these
tools. Training very large deep neural networks is a prominent
example.

◊sub-heading{Labor}

The chapter on Labor shows us how human labour is both controlled by AI
and exploited in the development of AI. AI is used to monitor, track,
and predict demand and capacity. It is used to make hiring
decisions. AI also relies on labour-intensive, crowdsourced, manual
labelling of data that is needed for its training data. Crawford
reviews Amazon's Mechanical Turk and Google's reCAPTCHA technologies.

◊sub-heading{Data and Classification}

◊b{I have grouped these chapters together in this review because they
encompass the general focus of current-day research into AI fairness
and they can't be considered separate from each other. These two
chapters are a highlight of the book for me.} One of my only
criticisms of these chapters is how much of each might possibly have
been also discussed in the other chapter.

The chapter on data is the most coherent in the entire book, with a
great collection of on-point examples. It does slightly underplay the
current awareness of dataset issues, though.◊note{See e.g. the
[NeurIPS 2021 Paper
Checklist](https://neurips.cc/Conferences/2021/PaperInformation/PaperChecklist)
which asks authors to report on the copyright, licence, and consent
relating to the creation of the dataset(s) that are used.} The data
chapter has a good discussion of privacy and ownership issues and the
practice of using public data to create private models.

The classification chapter highlights the circularity and cycle
of feedback between data generation, training, classification, and
back to data generation. Classification risks replicating in the
future the biases exhibited in the past and this is tied up both with
the data used to train an AI and the systems surrounding an AI.

◊q{ Making these choices about which information feeds AI systems to
produce new classifications is a powerful moment of decision making:
but who gets to choose and on what basis? The problem for computer
science is that justice in AI systems will never be something that can
be coded or computed. It requires a shift to assessing systems beyond
optimization metrics and statistical parity and an understanding of
where the frameworks of mathematics and engineering are causing the
problems. This also means understanding how AI systems interact with
data, workers, the environment, and the individuals whose lives will
be affected by its use and deciding where AI should not be
used.◊note-cite["Crawford" #:pinpoint "147--78"]}

◊sub-heading{Affect}

This chapter is a focused critique on a particular application of AI:
affect (roughly, emotion) recognition. It echoes Crawford's critique
from [AI Now's 2019
Report](https://ainowinstitute.org/AI_Now_2019_Report.pdf) and
elsewhere. It points out the dubious foundations underlying these
attempts at classifying people's emotional state based on their facial
expressions. This is another good chapter in terms of coherence and
writing style. I would not be surprised if it had existed as a
standalone piece before being used in this book.

◊sub-heading{State}

While state use of AI was discussed tangentially throughout the book
(COMPAS, law enforcement generally, border crossings, mug shot
analysis), this chapter is ostensibly focused on state intelligence,
surveillance, and decision-making. However, it unavoidably weaves
together a story that includes private research (e.g. Google,
Microsoft) and decision-making spanning NSA intelligence gathering to
apps like "Neighbors, Citizen, and Nextdoor." Crawford discusses the
Snowden documents, the goal of militaries around the world to gain the
AI edge, and enlistment of private corporations in these efforts. The
enamourment with AI surveillance and decision-making has also worked
its way down to local police forces, municipalities, and
administrative bodies. Crawford presents a commonly referenced example
of government decision-making gone wrong: a system used in Michigan to
attempt identify unemployment-insurance fraud. It "inaccurately
identified more than forty thousand Michigan residents of suspected
fraud."◊note-cite["Crawford" #:pinpoint "206"]

◊sub-heading{Power}

This chapter is the book's conclusion but it presents a new idea, too:
◊em{enchanted determinism}. This is the paradoxical idea that AI
systems are both "enchanted" in that they can do what we haven't told
them to do, but also deterministic enough to be safe and ethical for
deployment in high-stakes decision-making. I have also argued that you
can't have it both ways, especially when we expect a decision-maker to
act with discretion. An algorithm will either be following
deterministic instructions, which is a fettering of discretion, or it
will incorporate randomness, and that also isn’t what we expect from
discretion.

I like Crawford's reminder that we must attend to the committments we
are making as we develop and deploy AI system in private and through
the state.

◊heading{Critiques}

"AI is neither ◊em{artificial} nor
◊em{intelligent}."◊note{◊cite["Crawford" #:pinpoint "8"]} This is a
nice pithy point, but I don't think it is one that anyone disagrees
with, at least with the sense that Crawford sets up through contrast
in the following sentence.◊note{The quote goes on: "Rather, artificial
intelligence is both embodied and material, made from natural
resources, fuel, human labour..."} While some people have optimism
that we might someday create ◊em{real} intelligence, many people doubt
that this is possible, and nobody today suggests that we have done
so. This is why it is called ◊em{artificial} intelligence. The term
◊em{artificial} is used not to imply that there are no real artifacts
or effects of AI, but merely to indicate that they are not truly
intelligent. They are ◊em{artifices} of intelligence. They are also
"artificial" in another sense: that they are created by us, rather
than occuring naturally. I think this is wholly consistent with the
point that Crawford is trying to make with this aphorism but without
the need for Crawford's linguistic attack on the term itself.

◊declare-work[#:id "Strubell" #:type "article" #:author "Emma
Strubell" #:author2-given "Anaya" #:author2-family "Ganesh"
#:author3-given "Andrew" #:author3-family "McCallum" #:title "Energy
and Policy Considerations for Deep Learning in NLP" #:journal
"arXiv.org" #:year "2019" #:url "https://arxiv.org/abs/1906.02243"]

◊declare-work[#:id "Brown" #:type "article" #:author-given "Tom B"
#:author-family "Brown" #:etal? "yes" #:title "Language Models are
Few-Shot Learners" #:journal "arXiv.org" #:url
"https://arxiv.org/abs/2005.14165" #:year "2020"]

In the ◊em{Earth} chapter, Crawford cites the work of Emma Strubell et
al. for the proposition that "running only a single NLP model produced
more than 660,000 pounds of carbon dioxide
emissions."◊note{◊cite["Crawford" #:pinpoint "42" #:terminal ","]
citing ◊cite["Strubell"]} Strubell et al. were reporting the cost to
◊em{train} such a network. It is now well-recognized that
◊em{training} large neural networks has a large cost to our
environment, but also that the cost of running them at ◊em{inference
time} is not nearly costly.◊note-see-eg["Brown" #:pinpoint "s 6.3
(Energy usage)"]

I found the continued reference to ◊em{logics} to be unhelpfully
vague. Crawford often does not elaborate on what a particular
◊em{logic} actually entails.◊note{E.g. "the logics of capital,
policing, and militarization"; "the logics of extraction, a constant
drawdown of minerals, water and fossil fuels"; "Logics of efficiency,
surveillance, and automation"; "the response is to accept regulation
at the margins but to leave untouched the underlying logics of
production"; "new logics of information at a planetary level";
"cross-breeding corporate and state logics"; "Silicon Valley had
already built and monetized the logics and infrastructures of AI."}
It seems that ◊em{logics} is used variably as a synonym for "methods,"
"effects," or "values." But the intended meaning isn't clear and seems
to shift depending on the usage.

Last, it seems like the editor may have demanded additional content to
fill out some of the chapters, especially the chapter on labour. The
Google TrueTime example seems shoehorned into this chapter and it is
not a good fit. [Google
TrueTime](https://cloud.google.com/spanner/docs/true-time-external-consistency)
is a distributed, consistent clock that is available on Google
servers. This is presented in a chapter devoted to highlighting the
ways in which AI relies on and enables exploitative labour
practices. It is hard for me to see the argument that TrueTime does
this. TrueTime is used to consistently create timestamps for events
that occur across multiple servers and datacenters, separated by
thousands of kilometres. This kind of consistency is critical to
programs that deal with any sort of accounting of resources: e.g. bank
accounts, selling concert tickets. Many people might be attempting to
interact with such a system simultaneously through separate servers
and there has to be a way for those separate servers to determine
which interaction happened first. Otherwise, two people might be sold
the same final ticket to a concert. Crawford fails to connect this
technological example to the thesis developed in this chapter: that AI
serves the goals of controlling "bodies in space and
time"◊note-cite["Crawford" #:pinpoint "16"] and of "controlling the
political order."◊note-cite["Crawford" #:pinpoint "82"] Other examples
that Crawford provides are compelling; this one is not.

Again, these are minor points that can be skimmed past or set aside
while you read this work. The bigger picture is worth seeing.

◊heading{Conclusion}

For a general audience unversed on the full scope of the use and
development of AI, this is a helpful book. Much of the material
transfers to concerns with other large-scale technology: batteries,
blockchain, digital currency, NFTs. And AI as a tool of state power
will need public regulation.