doc: add LLM section to CONTRIBUTING.md #35264

thomasbuilds commented at 7:48 PM on May 11, 2026: contributor

Maintainers already push back on LLM-generated contributions but there is no written guidance for it. This adds it to CONTRIBUTING.md.

Beyond the reviewer cost argument I've read, I think there's a more important one I have just realized and haven't seen raised: LLMs are built and controlled by a small number of corporations. If enough contributors rely on them, those corporations can gain influence over what gets merged and which direction the project takes. That is a centralization problem, contrary to Bitcoin's foundational principle of decentralization and it will get worse as LLMs improve. Making this explicit gives contributors a principled reason to avoid over-relying on them.

doc: add LLM section to CONTRIBUTING.md

Add a section to CONTRIBUTING.md explaining why contributors should not
substitute LLM output for their own judgment across all forms of
participation.

The section is placed before the Contributor Workflow section so it is
visible to new contributors early, and explicitly covers all forms of
participation rather than pull requests alone.

b5801db739

DrahtBot added the label Docs on May 11, 2026

DrahtBot commented at 7:49 PM on May 11, 2026: contributor

The following sections might be updated with supplementary metadata relevant to reviewers and maintainers.

Code Coverage & Benchmarks

For details see: https://corecheck.dev/bitcoin/bitcoin/pulls/35264.

Reviews

See the guideline for information on the review process.

Type	Reviewers
Concept NACK	kanzure
Concept ACK	sedited
Approach NACK	fjahr

If your review is incorrectly listed, please copy-paste <code></code> into the comment that the bot should ignore.

kanzure commented at 8:01 PM on May 11, 2026: contributor

NACK. The proposed text is overly grandiose.

thomasbuilds commented at 8:05 PM on May 11, 2026: contributor

Are you against the phrasing I used in CONTRIBUTING.md or the idea/arguments ? I can always change the phrasing, that's not a problem

kanzure commented at 8:09 PM on May 11, 2026: contributor

These additions don't really accomplish anything or add anything of value. Anyone who was likely to do a drive-by PR with LLMs will continue to do so, regardless of what is written in the CONTRIBUTING file.

thomasbuilds commented at 8:18 PM on May 11, 2026: contributor

The drive-by spammer argument is fair, but that is not who I am writing this for. Most people submitting LLM-generated contributions are not bad actors, they genuinely want to help but find the codebase too large and technical to contribute without assistance.

Once they understand that relying on corporate-controlled models to shape Bitcoin development is itself a risk to Bitcoin, many will think twice. Written guidance cannot stop bad faith contributors but it can reach good faith ones who simply have not thought it through.

sedited commented at 8:22 PM on May 11, 2026: contributor

Concept ACK on writing something, I had the intention to propose a similar change soon.

We should be concise though. A paragraph warning that a PR may be closed if it is not distinguishable from slop, and that it is up to the author to give the impression that they understand the change, as well as indicate to reviewers that it is of decent quality, could be enough already.

These additions don't really accomplish anything or add anything of value. Anyone who was likely to do a drive-by PR with LLMs will continue to do so, regardless of what is written in the CONTRIBUTING file.

We've had a bunch of non-drive-by LLM contributions too, which is why I'd rather have a short blurb than not.

fjahr commented at 8:59 PM on May 11, 2026: contributor

If enough contributors rely on them, those corporations can gain influence over what gets merged and which direction the project takes.

This is irrelevant. If developers would rely on their tools blindly we would have the same problem with supply chain attacks from compilers, editors, github etc. The whole point of this is that LLMs should not be relied on blindly. This whole part should be dropped.

The current version is way too verbose and repetitive like others have already said.

thomasbuilds commented at 9:21 PM on May 11, 2026: contributor

The supply chain comparison is a bit different. A compromised compiler is an attack vector that can be identified and fixed. LLMs improving over time and becoming the default tool for reviews and discussion is a gradual structural shift, and the more capable they get the less people will notice it happening.

I also think this is Bitcoin-specific in a way that matters. Many other open source projects like Trezor heavily use Copilot for reviews and that is fine for them. Bitcoin is different precisely because decentralization is the whole point.

I will tighten the section, but removing the centralization argument takes away the only explanation of why blind reliance is harmful here specifically. Without it "don't rely on them blindly" is just an instruction with no reason behind it. Do you agree?

fjahr commented at 9:58 PM on May 11, 2026: contributor

A compromised compiler is an attack vector that can be identified and fixed.

That does not make a difference in the context of what I said.

LLMs improving over time and becoming the default tool for reviews and discussion is a gradual structural shift, and the more capable they get the less people will notice it happening.

So the answer is to just not rely on them blindly, the rest is just unnecessary fluff that dilutes the message.

I will tighten the section, but removing the centralization argument takes away the only explanation of why blind reliance is harmful here specifically. Without it "don't rely on them blindly" is just an instruction with no reason behind it. Do you agree?

Approach NACK

maflcko commented at 6:23 AM on May 12, 2026: member

NACK. An essay on your thoughts on LLMs should go on your private blog, not into CONTRIBUTING md.

Low effort pull requests have always been rejected and closed. This is already documented and was last clarified in fa15a8d2d03b0099b97262e47a8cbf685c29dd49. If there is a meaningful re-word of any of the sentences, it can be taken, but recall that the llm bots don't read this file anyway, so I don't really see the point.

Closing for now, due to controversy and the number of NACKs.

maflcko closed this on May 12, 2026

maflcko commented at 7:44 AM on May 12, 2026: member

While there have been a few LLM generated PRs by first-time contributors that were merged, the majority was not. If someone wanted to clarify the docs to document the status-quo that almost all AI-assisted pulls by newcomers will likely be closed, a single sentence can be added somewhere (with the rationale being the review bottleneck).