TLDR: this is a continually evolving doc and will have to be revisited. Please consider this my best understanding of how to think about AI usage as of Aug. 5 2026.
It is obviously a weird time to be a student and a researcher. PhDs have always been a weird mix between learning and production and now we have a tool that, at the very least in the short term, appears to increase your productivity. It’s also the case that people are increasingly using it for “autoresearch”, in which they set it a task and step away and publish the resultant thing. Folks who are doing this are able to publish more papers, albeit (I assume) at the cost of learning. I’ve spent a while thinking about all these tensions and here’s the guidelines I’ve come up with. I will also note that there’s very little I can do to track if you’re following these guidelines, so note that they’re all on faith.
But before we go into the details I want to note one unique aspect of being a researcher that is relevant to AI and informs everything that follows. You are the only backstop to your work. Everything you generate is backstopped by your reputation and when the AI makes a mistake, OpenAI doesn’t take a hit, the reputation you spent years building does. Now, of course, as your advisor, I will always take as much of the hit as possible because in the end I am always where the buck stops, but keep this in mind.
Lastly, I’ll note that using AI in research and in educating yourself is really hard. When it comes to educating yourself, AI offers the same temptation as having the whole solution sheet to the problems in the textbook. But this analogy also points out that there’s kind of nothing new here; in the same way that looking at the solutions in the back doesn’t help you improve, neither will offloading answering the problem to an LLM. But just like the solutions, if you’re able to identify when you’re productively/unproductively stuck, you can use it sometimes to get yourself unstuck. But both in research and education, never ask it to just solve the whole problem for you.
Pulled out, here are some high-level rules of thumb that pervade this document:
Finally, I want to end that even if you can prompt an AI to accomplish a lot of tasks, you have to know what tasks are worth accomplishing. And there’s no skipping the process of getting to the state. A lot of what looks like AI productivity is a lot of knowledge people built up over time that they are now leveraging to guide these tools in the right direction.
Do not use AI for any important writing. Do not do this. One of the ways that you will stand out as a researcher is your ability to put together compelling writing and narratives. A lot of desire to use AI comes from feeling like not being productive (or not as productive as researcher X) but as a PhD student, you are almost never bottlenecked by a need to get a lot of writing done fast and every time you use it for writing, you are increasing the time it will take you to become a skilled writer. Besides which, trying to put your thoughts down clearly is a type of thinking and you are taking this type of thinking away from yourself. You are trying to become someone operating at the absolute frontier of thinking and doing and you can’t do that from the local minimum of a tool that is well below those capabilities; these things are bad at writing.
When might AI use be okay in writing?
In this lab I will constantly be pushing you to write more because I think it is the best way to clarify your thoughts. If you use AI to do this, you are missing one of the reasons for being here (as opposed to being in another lab).
There is no denying that AI can put out more lines of code than a human being can, and this code can be quite good. There are two things I want you to be considerate of. First, use this as an opportunity to make sure your repo conforms to ‣! This is an opportunity not to just make more code, but to make code that is better, more tested, better documented, to understand a codebase you are interacting with better, etc. Second, my expectation is that you do not offload a critical understanding of your project. If I ask you how a critical piece of your project works, my expectation is that you know the answer. Why? Because when you’re debugging and trying to figure out why something doesn’t work, you will need this. And second, because if you don’t know this, then you can’t guarantee to a reader that your code does what it says it does.
There are some really good uses of LLMs in math. A particular neat one is you can use it to formalize any of your proofs in Lean, thereby giving you confidence that what you’ve done is correct. This is a new capability that really didn’t exist before! You can also use it to get ideas for proof techniques, since it has a really good view of the whole mathematical corpus. Just like before, probably don’t ask it to just prove the thing.