But I also think that if a maintainer asks you to jump before submitting a PR, you politely ask, “how high?”
If trust didn't matter, there wouldn't have been a need for the Linux Kernel team to ban the University of Minnesota for attempting to intentionally smuggle bugs through the PR process as part of an unauthorized social experiment. As it stands, if you / your PRs can't be trusted, they should not even be admitted to the review process.
Otherwise, what’s the harm in saying AI guides you to the solution if you can attest to it being a good solution?
If I just vibe-coded something and haven't looked at the code myself, that seems like a necessary thing to disclose. But beyond that, if the code is well understood and solid, I feel that I'd be clouding the conversation by unnecessarily bringing the tools I used into it. If I understand the code and feel confident in it, whether I used AI or not seems irrelevant and distracting.
This policy is just shoving the real problem under the rug. Generative AI is going to require us to come up with better curation/filtering/selection tooling, in general. This heuristic of "whether or not someone self-disclosed using LLMs" just doesn't seem very useful in the long run. Maybe it's a piece of the puzzle but I'm pretty sure there are more useful ways to sift through PRs than that. Line count differences, for example. Whether it was a person with an LLM or a 10x coder without one, a PR that adds 15000 lines is just not likely to be it.
This is the core problem with AI that makes so many people upset. In the old days, if you get a substantial submission, you know a substantial amount of effort went into it. You know that someone at some point had a mental model of what the submission was. Even if they didn't translate that perfectly, you can still try to figure out what they meant and we're thinking. You know the submitter put forth significant effort. That is a real signal that they are both willing and able to do so to address going forward to address issues you raise.
The existence of AI slop fundamentally breaks these assumptions. That is why we need enforced social norms around disclosure.
10x engineers create so many bugs without AI, and vibe coding could multiply that to 100x. But let's not distract from the source of that, which is rewarding the false confidence it takes to pretend we understand stuff that we actually don't.
but maybe those don't need to be about "whether or not you used LLMs" and might have more to do
with "how well you understand the code you are opening a PR for" (or are reviewing, for that matter)
If they had used AI, their PRs might have been more understandable / less buggy, and ultimately I would have preferred that.
The only reason one may not want disclosure is if one can’t write anything by themselves, thus they will have to label all code as AI generated and everyone will see their real skill level.
If they had used AI, their PRs might have been more understandable / less buggy, and ultimately I would have preferred that.
To address your claim about AI slop improving the output of these mythical 10x coders: doubtful. LLMs can only approximate meaningful output if they've already indexed the solution. If your vaunted 10x coders are working on already solved problems you're likely wasting their time. If they're working on something novel LLMs are of little use. For instance: I've had the pleasure of working with a notoriously poorly documented crate that's also got a reputation for frequently making breaking changes. I used DDG and Google to see if I could track down someone with a similar use case. If I forgot to append "-ai" to the query I'd get back absolutely asinine results typically along the line of "here's an answer with rust and one of the words in your query". At best first sentence would explain something entirely unrelated about the crate.
Potentially LLMs could be improved by ingesting more and more data, but that's an arms race they're destined to lose. People are already turning to Cloudflare and Anubis en masse to avoid being billed for training LLMs. If Altman and co. had to pay market rate for their training data nobody could afford to use these AI doodads.
Exactly! The code used double as "proof of work". Well-formed language used to double as "proof of thinking". And that's what AI breaks: it speaks, but doesn't think. And my core point is that language that does not originate from well-reasoned human effort (i.e., from either writing the language directly, or from writing such code manually that generates the language deterministically, and for known reasons/intents), does not deserve human attention. Even if the "observable behavior" of such language (when executed as code) looks "alright".
And because I further think that no code should be accepted without human review (which excludes both not reviewing AI-generated code at all and having some other AI review the AI-generated code), I conclude that AI-generated code can never be accepted.
Such behaviors can only be normalized in a classroom / ramp-up / mentorship-like setting. Which is very valid, BUT:
- Your reviewers are always overloaded, so they need some official mandate / approval to mentor newcomers. This is super important, and should be done everywhere.
- Even with the above in place: because you're being mentored with great attention to detail, you owe it to your reviewer not to drown them in AI slop. You must honor them by writing every single line that you ask them to spend their attention on yourself. Ultimately, their educative efforts are invested IN YOU, not (only) in the code that may finally be merged. I absolutely refuse to review or otherwise correct AI slop, while at the same time I'm 100% committed to transfer whatever knowledge I may have to another human.
Fuck AI.