Quick question about your AI policy!

[cloudskater]

Hi! I've been a user of BizHawk for quite a while now and it's quickly become my favorite emulator I've ever used. It's also the main reason I can use RetroAchievements for most of my games. I'm really grateful for the project. ^^

I had a quick question about your AI policy, though. It says contributions "authored" by AI/LLMs won't be accepted, which is great to hear, but does that signal a complete ban, or will "human reviewed" AI code be accepted at any point? I ask because there's a growing number of projects that have pledged to reject the current tech bro "AI" movement, and I'd be ecstatic to add my favorite emulator to this list!

[YoshiRulz]

Still haven't gotten around to writing LLM usage and provenance docs, so the official answer is just that one line in the contrib guidelines. Effectively, if we catch you "ferrying" error messages and review comments back and forth to an LLM, that will block your PR from getting merged. Some high-profile FOSS projects have been inundated with low-quality automated PRs, and while we've dodged that so far, it's on everyone's minds.

My personal opinion is that (as of last year or so) it's possible to train an LLM which is good at programming and isn't copyright-infringing, but most models aren't that. And I suspect even models which are good at C# will assume .NET Core, which is useless to us. But maybe we could use them to write unit tests or something.

edit 2026-06-26: Regarding a blanket ban like described here: I think that approach is a non-starter. I favour the self-identification approach because it works with, rather than against, the AI('s overseer). It would be more effort for the overseer to remove all explicit mentions of the model than to rm AGENTS.md and re-run.
Our goal with adopting AGENTS.md would primarily be to prevent human maintainers from wasting time on code review of slop PRs. I don't think we should be discouraging individuals from experimenting with the codebase or making customisations (one of the 4 freedoms). Some of them may end up with a patchset worth contributing. Case in point: #4677

[cloudskater]

My personal opinion is that (as of last year or so) it's possible to train an LLM which is good at programming and isn't copyright-infringing, but most models aren't that. And I suspect even models which are good at C# will assume .NET Core, which is useless to us. But maybe we could use them to write unit tests or something.

I see. I have little programming knowledge so I can't comment much on the technical details, but even if a model was available that was verifiably not infringing copyright, there are still many other issues that come with this kind of thing, many of which are outlined here. I'm glad that BizHawk is at least resistant for the time being, but aren't LLMs pretty unethical by design? From my understanding, they're called large language models because they are trained on public data that, copyrighted or not, nobody gave permission to be scraped. Additionally, even if there existed a form of this technology made in a "just" way, would you really want to hand even part of the process over to such a device? Especially if it's closed source, which goes against the reason FLOSS software inherently?

I just don't see a reason not to ban AI outright, is all, especially given the environmental concerns around it. Again, BizHawk is one of the most useful emulators out there, and it would be really assuring to know it was going to take a firm stance against this sort of thing.

[SuuperW]

I think it's great that the list of potential issues you listed gives specific examples. But most of the issues with AI use are best solved by doing it better (both users and companies making them). Some of what should be done better is best enforced with well-enforced laws and policies that guard against misuse/abuse, not outright bans. It's important to realize that there are many, many options besides ban or not ban.

aren't LLMs pretty unethical by design?

Not at all. There's nothing inherently unethical about LLMs in theory. Some or maybe all popular ones used unethical methods of gathering training data or do other unethical things to varying degrees, but it is absolutely possible to make and use an LLM while being ethical.

Additionally, even if there existed a form of this technology made in a "just" way, would you really want to hand even part of the process over to such a device?

Yes! I am all for tools that make boring things easier and faster. It's really easy to use LLMs in a bad way, in ways that lead to problems. But proper use would be beneficial. It being closed source doesn't matter either, software developers routinely use closed source tools and even non-software tools as part of their process.

especially given the environmental concerns

Personally I find this to be one of the weakest arguments, because the environmental problems seem to me to be among the easiest problems to solve. (Legislation can demand they recycle water, not use too much public potable water, and contribute financially to clean power generation and grid infrastructure if they consume wildly more than anyone else.)

infringing copyright

Also I'd like to point out that using bits of someone's code as training data isn't copyright infringement. But there is an issue with companies scraping public data sets, in that gathering these data sets requires significant time, effort, and money from the companies that do it; yet LLMs can offer a service that somewhat obsoletes those data sets (such as Stack Overflow) by providing a replacement service, without paying the companies that originally collected the data.

[YoshiRulz]

I'm not that interested in debating (this is an issue tracker...), so I'll be brief:

• On power consumption, reports are still conflicting, but even the higher estimates aren't a problem because renewables won. Even here in Australia we're finally going green.
• I don't know whether software can be unethical by design. (Maybe a disk wiping virus would be the digital equivalent of a firearm and thus unethical?) Clearly LLMs have at least some legitimate uses, so you can't make that generalisation.
• Releasing code under a licence like CC0 would, in fact, include permission to use it as training data. Many other licences would allow it under the condition that the licence, or just the authors' names, were preserved when an LLM regurgitates the code. Amazon made a POC.^{[citation needed]} GitHub didn't care, breaking the terms of those projects' licences, which yes in most jurisdictions amounts to copyright infringement. (I made this analogy and this other comment at the time.) There's an ongoing class-action suit.
• I would not want to entrust an LLM with control over the codebase and/or repo, no. I barely trust humans to keep it in working order. If we were to allow LLM contributions, I would consider them tools whose actions (and code) their human managers must take responsibility for. And my position remains that all tooling needs to be FOSS.

[aliusentity]

Chiming in as an outsider:

While it's true that LLMs can be ethical, the current reality is that any competent LLM is trained on the work of people that did not consent and who are not credited.

Additionally, natural gas supplies around 40% of electricity used by US data centers. Coal supplies 15%. Renewable energy does not seem to have won, yet.

Because of the current state of things, I think it makes sense to completely ban LLM use in the repo. The policy can always be changed when things get better.

[YoshiRulz]

Really tempted to hide your comment.
45% renewables is higher than the global figure of 32% in 2024, so that doesn't even refute my point. I also question whether you can extrapolate anything from statistics about the USA because it's an outlier in so many respects.
Also, did you make a GitHub account just to picket for anti-AI policies? Good on you for activism I guess.

[aliusentity]

Really tempted to hide your comment. 45% renewables is higher than the global figure of 32% in 2024, so that doesn't even refute my point. I also question whether you can extrapolate anything from statistics about the USA because it's an outlier in so many respects. Also, did you make a GitHub account just to picket for anti-AI policies? Good on you for activism I guess.

Thank you.
It was wrong of me to leave out that the remaining 45% is split between nuclear and renewables (20% nuclear and 24% renewables). The majority of the world's AI compute power is housed in the United states, which is why I brought that up.

Here's another article about the environmental impacts of genAI that might be of interest.

[jessebot]

I wanted to kindly chime in as a casual gamer and professional engineer at gaming company with more info on the environmental impact of AI:

• MIT News: Generative AI’s environmental impact
• MIT Technology Review: We did the math on AI’s energy footprint. Here’s the story you haven’t heard.

And a large reason why the US is relevant btw, is that the bulk of the LLM companies that people use are based in the US and that's often where the bulk of their datacenters are. Yes, they're also all over the world, but they are often located in low income areas of the US, because that's where it's cheap to buy land, and they often use things like diesel generators which then poison the air around the neighborhoods they are based in.

There are many ethical reasons to avoid LLMs, but the environmental impact is a very big reason that I think is very hard to ignore. Also, for what it's worth, I didn't make a GitHub account to advocate against AI. I just think it's important to mention when projects I think are neat are having the discussion.

Another reason to ban them is the hardware prices they drive up:

• The RAM shortage is coming for everything you care about

• Western Digital is already sold out of hard drives for all of 2026 — chief says some long-term agreements for 2027 and 2028 already in place

• The 2026 storage crisis: Why AI data centers are hoarding every hard drive on the market

This then makes it harder for people, especially those with not a lot of money, to get even the older computer parts they need to build small machines to run even the most basic of things.

Finally, there's the AI company's ties to the war industry, which are significant and contribute to some unfortunately sad things. Here's an example of what I mean: NYT: Palantir, Anthropic and small start-ups are generating rewards from their investments in defense tech. A lot of that gets re-invested in defense contracts with the US department of War, which is not going green.

I hope you'll consider a no AI policy to help keep the community a little more ethical.

[lumi-me-not]

I would say generative "AI" is unethical from the start. Taking a massive amount of human creative outputs just to soullessly regurgitate them is offensive. Doesn't matter whether it's for code, art, writing, or whatever else.

There are many, many more reasons it is unethical, and others have talked about some of them. I don't think accepting it in any form is a good idea.

So, I would like to see free software projects and their communities band together in solidarity with everyone affected, and go against this deeply unethical technology.

[YoshiRulz]

I think that's enough debating for now.
This can stay open as a nag until one of us (probably me) writes more detailed guidelines.

[vadosnaprimer]

I'll reread the whole thread but it's probably time to make a decision given that tasvideos the site and tasvideos the codebase now have a GenAI policy

• https://tasvideos.org/AIPolicy
• TASVideos/tasvideos@fbad245

It's clear to most people that bizhawk is not directly obliged to follow the above thing directly. BUT. As CPP said the site and bizhawk can't be separated either:

• bizhawk was created by a tasvideos admin (adelikat) and a person with server access (zeromus) to make life easier for people creating tas tools for people on tasvideos. people in charge of tasvideos are still in charge of bizhawk.
• bizhawk is actively developed with tasvideos needs in mind. both adelikat and natt are also the main people behind tasvideos the codebase
• some active bizhawk devs are not super active on tasvideos, but that doesn't mean any decision making life harder for the site has a chance of survival in bizhawk. and I don't believe site decisions are going to harm bizhawk either.

The tasvideos mantra with the AI policy is "we'll have to allow tas tools that contain GenAI code until there's a no-GenAI alternative". That sounds like a reasonable compromise for things already existing, and something to work with regarding things yet to come.

Regarding https://gitlab.com/YoshiRulz/yoshis-hawk-thoughts/-/work_items/228

Linux kernel has how many devs and how many money circling around just because how many industries depend on it working properly? Well even despite all those resources Linus admits they are overwhelmed by GenAI. He has the luxury of considering it a good problem to have because duh it still finds some legit bugs woah it's not 100% slop thank god. We don't have that luxury.

Does Linus have an answer for questionable license of whatever AI has generated? AI is sure trained to copyrighted/licensed material (most of it wasn't licensed for AI training of course), and in at least some cases it can output you back that exact material given the right prompt.

SDL's policy sounds great https://github.com/libsdl-org/SDL/blob/main/AGENTS.md. Except I'd add that if someone personally processes whatever AI has found and reports it properly, but that still happens to be nonsense, I'd want to forbid that crap. We don't have the resources to determine whether a report is nonsense or not. We don't even have the resources on all the reports that we know are legit. Adding "possibly nonsense" on top of that will turn it into complete nightmare and burn out whatever is left.

So my suggestion is forbidding anything AI generated completely, and also forbidding any hallucinations that the human "translates" over to us. First offense is worth a warning. Repeated offense is worth a ban.

But it only makes sense to expand this to direct contributions to this repo. We can't control the code that we submodule. In the past we also used to include entire repos directly for our components (like cores), but I hope that won't be needed going forward. And if we get to pick a component out of 2 versions, where one has AI generated content and another hasn't, we just pick the clean(er) one.