[openstreetmap/openstreetmap-website] How should we deal with zero-effort AI PRs? (Issue #6954)

[Tom Hughes]

tomhughes left a comment (openstreetmap/openstreetmap-website#6954)

So here's where I am on this...

Whilst I'm personally unconvinced of the utility of current LLMs for this kind of work I'm prepared to work with people that use them so long as they are in the loop as a sentient human being that is only using the LLM as a tool.

What I'm not interested in is PRs where somebody is sending us unreviewed output from an LLM and I am talking directly to an agent or to a human that is simple acting as a stenographer and copying everything I say to the LLM and then pasting back the result with no kind of review or filtration at all.

Simply put we have ample evidence already that such PRs are impossible to deal with - you just go round and round in circles trying to coax them into shape with every round of review fixing things and breaking others.

Fundamentally an LLM which is just given review notes doesn't act in the way you want:

* It treats every comment as an order and tries to act on it - it will never ask for clarification or push back and suggest that another solution, or the original solution, might be better, even if you phrase the review comment as a question. Humans struggle with this and LLMs are ten times worse.
* It focuses very narrowly on the comment and never considers the wider impacts that might have been implicit so it will do things like changing the way a piece of code works but not remove the function that it originally added but which is then no longer needed because you never mentioned that.

But to me the biggest thing as a maintainer is that every time I try and help a new contributor to improve by reviewing their work that is in fact completely wasted if there is no human learning from the comments. Even if the LLM uses the entire discussion as context when answering review comments, so that it (sort of) learns for this PR all of that will be entirely lost when the next PR comes round as nobody is including all their previous PRs in the context window for their future PRs.

The end result is that every time you review such a PR you're effectively starting from scratch with trying to teach a new contributor even if that person has nominally opened 100 previous PRs so I am just wasting my time reviewing such work and have already effectively stopped bothering with such PRs.

-- 
Reply to this email directly or view it on GitHub:
https://github.com/openstreetmap/openstreetmap-website/issues/6954#issuecomment-4172019605
You are receiving this because you are subscribed to this thread.

Message ID: <openstreetmap/openstreetmap-website/issues/6954/4172019605 at github.com>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstreetmap.org/pipermail/rails-dev/attachments/20260401/949ba7e7/attachment.htm>