@@ -130,6 +130,78 @@ are interested in, it is a good idea to:
130
130
used for searching the mailing list and linking to previous
131
131
discussions.)
132
132
133
+ ## AI guidelines
134
+
135
+ The use of Artificial Intelligence related tools ("AI" or "AIs" for
136
+ short), like Large Language Models ("LLMs"), by developers is becoming
137
+ more and more popular, but that's not a reason to blindly use them,
138
+ especially when working on the Git project.
139
+
140
+ ### Legal issues
141
+
142
+ In general, it's not clear if AI generated code or documentation are
143
+ acceptable in the Git project. Authors contributing to Git are
144
+ required to "sign off" the patches they contribute, and to agree with
145
+ the Developer’s Certificate of Origin, also called DCO, (see
146
+ https://git-scm.com/docs/SubmittingPatches#sign-off ), while the DCO
147
+ says basically that authors of a patch should be sure about the origin
148
+ of its content.
149
+
150
+ As we cannot usually be sure how AIs have been trained and that they
151
+ are not just repeating proprietary existing code or docuementation
152
+ they saw during their training, we cannot accept much AI generated
153
+ code or documentation in general. Now, if it's only a few lines to fix
154
+ a bug or to implement a common pattern or summarize something, and if
155
+ that looks specific enough to a current concrete problem a developer
156
+ is working on that might be OK.
157
+
158
+ Anyway as the situation is not clear for the Git project and probably
159
+ many other open source projects, you should be very prudent regarding
160
+ this.
161
+
162
+ ### Be very careful with AI output
163
+
164
+ For a number of reasons, not just legal ones, developers should really
165
+ make sure that they treat what AIs produce very carefully. They
166
+ should:
167
+
168
+ - triple check everything, especially regarding our guidelines
169
+ (indent, style, commit message guidelines, etc, see especially the
170
+ SubmittingPatches and CodingGuidelines docs as well as the other pages
171
+ on this website) and the feedback reviewers already gave them or
172
+ others,
173
+
174
+ - doubt anything they don't fully understand, or anything that might
175
+ not match our guidelines or feedback, and
176
+
177
+ - fix, simplify, adapt or change anything that is suspicious,
178
+ bloated, too formal, or that they don't understand, or that doesn't
179
+ match our guidelines or our feedback.
180
+
181
+ Yeah, AIs still often hallucinate or just produce bad outputs, even
182
+ when you point out their mistakes.
183
+
184
+ It's unacceptable to send us something that is obvious AI slop, or
185
+ that sounds overly formal or bloated, or that looks good on the
186
+ surface but makes no sense, or that senders don’t understand or cannot
187
+ explain. It just wastes our time and we cannot accept it.
188
+
189
+ For example, it's unacceptable to have an AI generate a commit message
190
+ where it just describes what the code in the commit does, instead of
191
+ the purpose of the change, and then send that to us.
192
+
193
+ ### Better ways to use AIs
194
+
195
+ Developers would often likely get better results, learn more and have
196
+ a better overall experience by asking AIs to only explain things, and
197
+ guide them step by step towards producing a solution by themselves,
198
+ rather than by asking for a full solution that they would then mostly
199
+ copy-paste.
200
+
201
+ They can also use AIs to help with debugging, or with checking for
202
+ obvious mistakes, things that can be improved, things that don't match
203
+ our style, guidelines or our feedback, before sending it to us.
204
+
133
205
## Note about the number of slots
134
206
135
207
The Git organization usually has very limited mentoring capacity.
0 commit comments