My lab's AI policies and norms

I – like every other academic in the world perhaps – am trying to come to terms with the effects of AI on research, teaching, and everything else. This post is about AI use in research. 

We recently had a lab meeting about how we are using AI in our research. People expressed lots of amazement about how much you can do with AI coding agents especially. I've personally been having a lot of fun using agents to help me pursue my research agenda (example, this preprint). But students also expressed a lot of disorientation, bewilderment, and skepticism. It's a challenging time to be a trainee scientist! There are many new opportunities for discovery, but the ground is shifting under our feet and it's hard to be sure exactly what's required, what's permissible, and what best practices are. 

My north star in all of this is the idea that science is not just for making more PDFs. Our scientific process should be organized around robust and transparent routes to discovery. Used appropriately, AI can lead to faster progress and stronger research. Used inappropriately, it can be a route towards generating papers that no one understands and slowing down communication. Here are a few principles we discussed in our meeting that I'm putting forward as current lab policies. 

Policies

No AI-authored text in papers. We will write all the words in our papers. This step maintains human understanding and direction in all of the research we do. It also goes some way to preventing inadvertent errors or fabrications by models, though other checks are critical. To be clear, AI-authored text is fine for some applications (e.g., software documentation or administrative tasks). But let's keep it out of scientific papers – at least, for the moment. 

All papers must have an AI disclosure statement. Even if no AI is used in a project, that's important information to disclose. AI disclosures should say what the AI was used for, and ideally what models and versions. This helps readers understand the context of how work was accomplished – in much the same way a CRediT statement helps assign responsibility across many authors in a paper. 

No submission before AI review. Modern frontier models are very thorough reviewers. They routinely catch errors, typos, and flaws/overreach in interpretation (see King & Worthington preprint). For the last few months I have been eliciting reviews of papers that are being submitted from my lab (after I have reviewed the papers and commented myself, often multiple times). I am embarrassed by how much the agents catch that I have missed! It feels to me like when I was in intro CS in college and tried to beat the Boggle agent that I coded up for my problem set – no matter how long I looked, I always missed some words. So it feels irresponsible to me at this point to submit something that hasn't been reviewed by a thorough agent. You don't have to take every suggestion – some of them will likely be wrong – but it's critical to consider them. 

Abide by reviewer policies. Don't put articles you are reviewing for a journal or conference into a model, unless you know that it's allowable according to the venue's policies. Same thing goes for shared materials. 

Be very careful about models and participant data. Commercial models should never see participant identifiers, though they can write code that manipulates such data. Other sensitive data (e.g., video, audio, or free text data) should only go to local models or models under agreements that prohibit training on the data (see for example, the BabyView dataset use guidelines).  In general, be aware of the training agreements for the models you use. 

Norms

Here are a few more pieces of less formal guidance for AI use in the lab. 

Use good engineering practices to take advantage of coding agents. Coding agents are amazing and can be excellent tools for rapidly building new experiments or models. But if you build too much, too fast, you can end up with a codebase that has errors or is hard to maintain. Make sure you are always working under version control. If you're building any kind of reusable functions, make sure you're establishing tests for them. Be ruthless about reorganizing your project so that it is clear, and request clear, human-readable documentation. Russ Poldrack's Better Code, Better Science book has lots more great guidance. 

Mark AI-authored text when you circulate it. Don't pretend the AI is you; attribute text appropriately. This is as simple as marking Claude's opinions with a > in slack so that the text is quoted. Otherwise it feels like you're pawning off "someone else's" text as your own, same as in papers.  

Try not to send slop. If you don’t want to read some document, odds are, no one else does either. You can work with the AI system to create a tighter summary, or you can summarize yourself. If you have to share slop, mark it as such and don't expect people to read it. 

If you're trying to learn, read the paper. It can be a very helpful learning strategy to engage with a model about the content of a paper. But, read the paper first. If you don't, based on the learning literature, you are less likely to retain the content of the article. The difficulty you experience in reading the paper will lead you to learn and remember when you finally understand it. 

Beware the slop cliff! I really like this piece by Thomas Jiralerspong that cautions against falling off the "slop cliff," where you realize you don't actually know the details of what you're working on. Get used to delegating and building trust with models by checking details and asking follow-up questions to probe and build your own understanding of what the model is doing. When the model's answers don't match what you expect, that's a good sign that you're sliding down the cliff and you need to build your own understanding. 

These are complicated, challenging times for researchers (and humans!) but at its best, AI can be a powerful engine for scientific discovery. I hope we can figure out ways to maintain and strengthen scientific norms to make this possible. 

---

AI disclosure: I got comments on this post from Claude Opus 5.5.