doc: refactor the AI guidelines - #65269
Conversation
The earlier version of the document contains duplications and has become logically disconnected after iteration. This patch wordsmiths it a bit to make it more coherent: - Merge policy points that have been repeated in the guidelines into the opening policy. - Remove redundant taglines. - Reword the paragraphs so that they appear more logical: -> decision making and accountability is human-only -> AI assistance is not prohibited -> disclosure and licensing -> verification and review requirement -> other off limits items (automation, good first issues) - Reword the guidelines to contextualize the suggestions with the use of AI instead of being generic. - Change the tone of the document to be more in line with other documents we have (e.g. avoid being too imperative or calling out "you" constantly) Signed-off-by: Joyee Cheung <joyeec9h3@gmail.com>
|
Review requested:
|
4b185d4 to
4783c93
Compare
| The Node.js project does not prohibit the use of AI tools in contributions, but | ||
| when the contribution is generated with AI, the contributor must disclose the use | ||
| of such tools, and the extent to which the contributor has personally verified | ||
| the generated output. The changes submitted to the Node.js code base must still satisfy |
There was a problem hiding this comment.
The paragraph above states that one takes full responsibility of what they send. A consequence of that statement is that one has to verify the generated output. They should not state they do, they have to do it.
There was a problem hiding this comment.
The intent is more about "explain how you verify it" instead of "just declare you've verified it (but not explaining what you actually did)". Updated to
the contributor should disclose the use of such tools, and what the contributor does to personally verify the generated output.
|
@legendecas @mcollina I think I've addressed your comments. Can you take a look again? Thanks! |
| never be "I'm not sure. The AI did it." | ||
| The Node.js project does not prohibit the use of AI tools in contributions, but | ||
| when the contribution is generated with AI, the contributor should disclose the use | ||
| of such tools, and what the contributor does to personally verify the generated output. |
There was a problem hiding this comment.
I find the second part of the sentence redundant. I don't see there are any other way than reading all the code. Literally:
I was assisted by AI in proposing this change, and I reviewed it myself.
Should be enough. We should likely include it in a checkbox for the PR template.
There was a problem hiding this comment.
I think that depends on the working style, the person can edit the code to various degree - e.g. lay out the skeleton and let AI fill in some pieces, or keep the AI draft, but edit a lot of it. "Let the AI generate all the code and human only read it but not touch it" is only one of the many ways to use it, and IMO isn't even something we should encourage. And review is only one (weak) form of verification, there can also be other better ways to verify it e.g. benchmarks, tests, measuring impact otherwise etc.
6 participants
The earlier version of the document contains duplications and has become logically disconnected after iteration. This patch wordsmiths it a bit to make it more coherent:
-> decision making and responsibility is human-only
-> AI assistance is not prohibited
-> disclosure and licensing
-> verification and review requirement
-> other off limits items (automation, good first issues)