About Me Teaching Mentoring Professional Activities Publications Talks Links Bio AI Policy CV (May, 2026)
AI Policy

Version 1: September 5, 2026

The following is a concrete set of policies and less concrete guidelines for my research group based on our current understanding of the role of AI in theoretical computer science (TCS). It is worth emphasizing what this page is not: it is not about my personal opinion1 on what AI can do for TCS, what it means for the future of TCS, any excitement or fear about AI, whether we should use AI to the fullest extent or entirely shun it, or any ethical argument about "right" or "wrong".

This document will be updated regularly based on discussions in the group and what has worked for us and what did not. I do not think anyone has all the answers (I certainly do not have any of them), so we are all learning here. Given this, I reserve the right to scrap this whole page at some point and instead say: "No AI policy, go crazy!". Until that time comes however, we will be using this AI policy or its subsequent updated versions.

One last comment before we start: if you are currently my student, I am happy to pay for your subscription to use top AI models and I encourage you to be curious and keep an open mind about AI and examine different ways of incorporating it as a tool in your research and learning. I found this important to say because the rest of this policy may sound like it is putting too many restrictions on how to use AI in your research, and I want to make sure the takeaway is not that AI is "bad".2 The goal of this policy is to keep us within the standards of TCS research while making sure you get the most out of your training and/or graduate school as part of our group.

AI Disclosure: All texts on this page are my own, but I have used Claude for preparing the webpage and polishing the writing.

Specific Policies

  1. Documenting AI use: Document all AI use for each of your projects meticulously and separately. This means keeping track of all prompts you use in your research and their responses, all data you share with AI, and detailed (AI-generated) summary logs with token counts and number of hours if you are using a harness.3 We can discuss how to implement each of these steps in more detail.4
    Many of our TCS conferences require AI disclosures currently, but there is still no well-established standard for what these disclosures should include. For reference, here is an example AI disclosure from the FOCS 2026 call for papers:
    "AI Disclosure: We used [Tool Name] to assist with [Brief Description of Use]. The tool materially affected [Sections X and Y]. More details can be found in [Section Z]. The authors verified the correctness and originality of all content including references."
    The AI documentation we collect internally is meant to be considerably more expansive than the one-paragraph or half-page disclosure itself.
  2. Writing: All final write-ups---our papers together, your own separate papers, thesis, or technical reports---must be written by humans from scratch with no AI copy/pasting. We should be able to do such writings after internalizing any AI answers we used in the project and with no further access to the AI outputs.5 Receiving proofreading and light editorial feedback from AI is certainly fine and encouraged (to the extent you do not lose your own voice in the writing). But please do not give AI write-access to the files or folders used for the final write-up of any paper you write (ask for a PDF of feedback and comments to polish and implement them yourself instead, so you can verify the suggestions and learn from them for the future).
    I understand this policy is among the more controversial ones so while we are here I am just going to ask you for something even more controversial. Do not use AI as a shortcut for writing those pesky details or calculations that we all know how to do. Experience has made me a firm believer that nine times out of ten you will indeed only “waste” your time writing something obvious—but on the tenth, you will hit a roadblock you would otherwise have missed, and working through it will teach you enough to be worth all ten tries.6 Therefore, delegating this part to AI is going to rob you of an excellent learning opportunity.
  3. Exercises and practice: Until very recently, we could use simple open problems as exercises---for learning a new topic, learning how to do research, or simply for practice---and most of the time, we could also publish the solutions. It is entirely possible that already now or in the near future, the second step, namely, publishing the results, is no longer possible (because these questions can be solved in a very short span of time with the help of AI or some AI pipeline has already claimed them).7 This does not mean you cannot still practice---after all, the main goal was never publishing but learning and practicing, and AI cannot replace your understanding.
    In summary, never use AI for a practice problem or exercise no matter how tempting it is (we will be clear in our discussions when we are practicing).
  4. Privacy: Please turn off any data sharing settings on the AI tools you are using for research (e.g., in ChatGPT, turn off “Improve the model for everyone”). Even after having done so, please never share any document or half-baked ideas of someone else with your AI tool unless you have explicit permission from them.
  5. Sharing/claiming AI-generated ideas: If you would like to include an AI-generated idea in a paper or even present it in our meetings, keep the following in mind: you should first acknowledge this part was AI generated but then be able to present all the ideas entirely from your own notes and in your own language in a way that if all AI chat/content was deleted, you could still reproduce the whole argument. Until you reach that level of understanding about a result, we will be treating the AI-generated idea as an unverified manuscript and even a possible hallucination (this is regardless of the fact that the AI might be able to generate ten ways of Lean-verifying this statement).
    Despite the high bar of fully understanding the result first, you are still required to disclose if an idea was generated by or with the help of an AI, regardless of the fact that in hindsight, you understand it well enough that you could have certainly proved it yourself also.
  6. No AI back-and-forth between us: A reminder that if all you do is to prompt some AI between our meetings and report the answer to me, I can do that well on my own. Therefore, I ask you to entirely refrain from playing mediator between me and some AI and in return I promise to not do the same to you; this means, unless you specifically ask me to run something for you on AI, any comments on your drafts or emails you receive from me are entirely my own.8
  7. Standard practices in TCS: We will be following the policies set by our TCS community, primarily communicated through our conferences like STOC/FOCS/SODA.9 You can check the call for papers for recent iterations of these conferences for this information (e.g., the current practice that all listed authors on a paper must be human and take full responsibility for all content). I would also like us all to follow the Leiden Declaration on Artificial Intelligence and Mathematics, in all policies that do not go against existing TCS guidelines.10

More General Guidelines

More generally, I would like to ask you to consider the following in your interaction with AI:

  1. Our main goal is doing research for the sake of TCS: understanding something new and contributing to the general understanding of our research community, and, at the risk of sounding too full of ourselves, the understanding of humanity as a whole. Keep that in mind in all your interactions with AI. This also means that you should not reduce your job, or let others reduce it for you, to being merely a problem solver who jumps from one question to the next (especially to publish more and more papers).
  2. The following quote by Tao seems like a clear yardstick for whether or not a result is ready for publication:
    "if the authors cannot convincingly demonstrate that they are able to give a clear, expert-level talk on their results, one that is correct and properly attributed, then the result should not be published."
  3. A general principle to keep in mind about using AI is to use it to do better work, not more work: for instance, use AI to solve a much harder problem, to further polish the presentation of your work, to find more connections to the existing literature, to learn about an entirely new topic, to explore alternative ways of presenting your result, to write a blog post about an interesting research idea (yours or someone else’s), or to run experiments to see your work in action---instead of using it to pick even more low-hanging fruits and publish them.
  4. Try to build a general discipline about when to use AI and when not to. It is tempting to use AI to move past a hurdle in research, or to just run it to solve 10 problems while you sleep, or to finalize the last annoying part of an argument that you think you should know how to do but do not seem to be able to do just at the moment. In summary, do not use AI for cutting corners and watch out for becoming too dependent on AI.
  5. If you are using AI as part of your problem solving and TCS research, only use it on problems that you cared about a priori, whose relevance and importance you can communicate without having to rely on the AI, and whose solutions you can fully verify yourself (possibly after having gone through multiple rounds of back and forth with the AI to learn the new ideas in the solution).11
  6. Remember that if all we do is rely on some AI to do our work for us, then we ourselves are really not needed.
  7. Finally, I encourage you to learn how AI can enhance your learning and research: be curious about how it works and how to prompt it well, pick up best practices for using AI in TCS and math, form your own opinion about the role of AI in our field (and our world), and come up with your suggestions and comments about how to improve our AI policy in a way that serves you better.12

Resources and References

In preparing this document, I consulted the following resources:

  1. Use of AI in mathematical research: A guide for young mathematicians, by Pavel Etingof;
  2. Mathematics in the age of AI, by Terence Tao;
  3. Leiden Declaration on Artificial Intelligence and Mathematics;
  4. Mathematica: A Secret World of Intuition and Curiosity, by David Bessis, translated by Kevin Frey;
  5. Proofs and Prompts blog and specifically the posts by Mark Wildon, Tian Lan, and Martin Hairer;
  6. AI-Native PhD Students, by Eytan Adar;
  7. Can a junior mathematician remain an AI-vegetarian?, by Marcel Goh.
  8. The Crisis of AI-Generated Mathematics, by Max Weinreich.

Footnotes

  1. I generally do not see the point of sharing my opinion publicly because no one cares (*) and because there is already too much opinion online about everything (**). Of course, both (*) and (**) are opinions (and not just about myself) and the irony is not lost on me; you see, there is good reason for me to not share opinions. ↩
  2. Although I do understand the argument of those who would rather shun it entirely; if you belong to this camp, I respect your decision and promise not to ask you to use AI in your and our joint research, as long as we are both clear about the implications. ↩
  3. If this sounds like too much work, just look at how lab-based fields keep their lab notebooks. ↩
  4. As a simple example, do not have a single chat session for all your projects and personal questions and ask everything there, nor ask each question in some random chat that you will not be able to recover later. Create a proper project and folder for your interactions and keep those chats focused on the related project only. ↩
  5. I understand this is going to get very tricky in cases when we collaborate externally and we may need to be more flexible in those cases. It is best to discuss these views at the beginning of the collaboration and revisit them throughout to make sure all parties have a consistent view of AI policy in place. ↩
  6. For me, most of the time, if I feel something is true or even trivial but "annoying" to write, it is because I have not found the right way to express it and there is a gap in my understanding. ↩
  7. Although I also believe (or at least strongly hope) that our community will come up with a solution to how to navigate these types of results very soon, given the many encouraging discussions happening. ↩
  8. I would say that if I ended up finding way too many typos or silly mistakes in something you send me, I may have to send it back to you and ask you to run it by some AI on your own to catch the typos and mistakes before I read it again. ↩
  9. Until better sources emerge, e.g., EATCS or SIGACT publishing a community guide. ↩
  10. To my knowledge, currently there are no policies with conflicting views between the two. ↩
  11. There can be valid exceptions to this rule, for instance, if you are instead doing AI research, or research on benchmarking the power of AI tools in solving different types of problems. ↩
  12. I am very much open to experimenting with new ideas, as long as they put TCS research at the center. ↩

Created by Sepehr Assadi -- Last Major Update: September 2026