Add AI_POLICY.md

[SteveBronder]

I'm very pro ai coding, but I think having something like polars AI_POLICY.md is a good idea. One thing not listed there that Bob brought up is that totally autonomous bots cannot assign copyright ownership (in the US at least). So accepting fully autonomous pull requests puts us in a very weird spot.

[drezap]

are you writing tests with a human, or are we expected to have models that we’ve evaluated on simulated data in the model section, or something? I haven’t really tried merging or testing AI generated code but wondering if anyone has? Best, Andre Zapico linkedin.com/in/andre-zapico gitub.com/drezap ME Information and Communication Engineering University of Electronic Science and Technology of China Consultant, Owner likely llc likelyllc.com Stan Developer mc-stan.org BS Mathematical Sciences: Probabilistic Methods BS Statistics University of Michigan, Ann Arbor 2017

…

On Thu, Feb 26, 2026 at 12:12 PM Steve Bronder ***@***.***> wrote: *SteveBronder* created an issue (stan-dev/math#3289) <#3289> I'm very pro ai coding, but I think having something like polars AI_POLICY.md <https://github.com/pola-rs/polars/blob/main/AI_POLICY.md> is a good idea. One thing not listed there that Bob brought up is that totally autonomous bots cannot assign copyright ownership (in the US at least). So accepting fully autonomous pull requests puts us in a very weird spot. — Reply to this email directly, view it on GitHub <#3289>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACY543B4MGTUCC52LEQOTXL4N4SPHAVCNFSM6AAAAACWAZXQ3GVHI2DSMVQWIX3LMV43ASLTON2WKOZTHE4TMNZUGI3DIMY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[drezap]

stupid response? i got AI slop from a student for code that I developed with no pre existing training data, so what she generated was trash but for things like bugfixes (or whatever) could be good, i am not sure how effective this is in practice Best, Andre Zapico linkedin.com/in/andre-zapico gitub.com/drezap ME Information and Communication Engineering University of Electronic Science and Technology of China Consultant, Owner likely llc likelyllc.com Stan Developer mc-stan.org BS Mathematical Sciences: Probabilistic Methods BS Statistics University of Michigan, Ann Arbor 2017

…

On Thu, Feb 26, 2026 at 2:18 PM Andre Zapico ***@***.***> wrote: are you writing tests with a human, or are we expected to have models that we’ve evaluated on simulated data in the model section, or something? I haven’t really tried merging or testing AI generated code but wondering if anyone has? Best, Andre Zapico linkedin.com/in/andre-zapico gitub.com/drezap ME Information and Communication Engineering University of Electronic Science and Technology of China Consultant, Owner likely llc likelyllc.com Stan Developer mc-stan.org BS Mathematical Sciences: Probabilistic Methods BS Statistics University of Michigan, Ann Arbor 2017 On Thu, Feb 26, 2026 at 12:12 PM Steve Bronder ***@***.***> wrote: > *SteveBronder* created an issue (stan-dev/math#3289) > <#3289> > > I'm very pro ai coding, but I think having something like polars > AI_POLICY.md <https://github.com/pola-rs/polars/blob/main/AI_POLICY.md> > is a good idea. One thing not listed there that Bob brought up is that > totally autonomous bots cannot assign copyright ownership (in the US at > least). So accepting fully autonomous pull requests puts us in a very weird > spot. > > — > Reply to this email directly, view it on GitHub > <#3289>, or unsubscribe > <https://github.com/notifications/unsubscribe-auth/ACY543B4MGTUCC52LEQOTXL4N4SPHAVCNFSM6AAAAACWAZXQ3GVHI2DSMVQWIX3LMV43ASLTON2WKOZTHE4TMNZUGI3DIMY> > . > You are receiving this because you are subscribed to this thread.Message > ID: ***@***.***> >

[spinkney]

What happens if someone does not disclose they used AI? Or is it just that we hope people disclose. It's going to be hard to enforce except in the most obvious cases. I'm ok if it's aspirational to have disclosure or if we are clear about the aspiration and the consequence of not disclosing. I assume it will be, at least, a rejection of the PR until it is updated but let's be clear about that.

[SteveBronder]

Yeah I think this may have been a bit aspirational / a response to current events. Lets keep this open and if we start getting AI PRs then we will deal with it then.

[drezap]

I think the AI generated code thing may be a bit of a bubble. There was a GP model question that was generated and it was totally incorrect. I think they AI generated. And the code bases are so large, for example in finance, and when I was at SAS, there was no stack overflow answers. I had to ask some senior devs (30 years experience in C), and there was no solution. I was passing pointersback and forth between C/C++. I had to trial and error it out. And I had worked on a finance code-base as an intern back in the day, and there's no way an LLM, at least as of now, could have built this. We're talking millions of lines of code. I think this is a bit of an intellectual tech bubble... we're over estimating it's capabilities. In C++, one small error ruins it, so I'm not sure. I think Andrew made a blog post years ago about AI generated Stan code, but you still have to have an understanding of probability theory and Bayesian statistics and Stan modeling language to fix it. The post about GP's was total nonsense... to simulate from a multivariate gaussian distribution, you need to use cholesky decomposition, and clearly, this person did not understand this. I'm not sure, it's interesting but may be an "intellectual bubble". You still need to understand the math. I was just curious. Like, Steve worked in finance, hedge fund code bases are huge, you're piping lot's of data around, and it's done off the cuff extremely quickly, so sometimes it's a mess. It's essentially random, but it compiles. So I'm not sure how useful LLM technologies are on production code-bases.

[drezap]

The capabilities are a bit over-stated, I think. "build me a computational fluid dynamics software with a slick GUI," prompt. Ha, no.

[jachymb]

Maybe a simple thing to do in the meantime could be to edit the github pull request template and add a requirement to simply at least disclose AI use?

[spinkney]

The Linux kernel has a good template https://github.com/torvalds/linux/blob/master/Documentation/process/coding-assistants.rst. I like the license call-out, following the specific development process, and having a specific form for attributing the AI assistant.

[SteveBronder]

We have stan-dev/design-docs#59 and I think we are just waiting for the SGB to do the final approval before adding an AI_POLICY.md file here

[florence-bockting]

FYI: Stan's AI policy has been added to this repo through PR #3320