How, when and why to use agentic AI in our neuroscience labs

It dawned on me in early March this year, on the Caribbean island of Barbados, of all places. Konrad Kording, in swimming trunks, stood in front of about 30 PIs with backgrounds mostly in neuroscience and machine learning and live-demoed Claude Code, using a projector hardly visible in the broad daylight. He asked Claude to build a web app, and within minutes it was ready to test. The demo—designed to show how independently agentic AI could now solve tasks—presented the perfect picture: human in swimming trunks, machine working hard.

What struck me wasn’t only the scope of the change brought about by agentic AI, but its speed. For me and many others, the impact was immediate; our discussions that day became simulations coded in minutes that would have otherwise cost us days. I quickly realized that this combination of impact and speed is why we can’t just “drift” into this. Instead, we need to get behind the steering wheel and decide how, when and why to use agentic AI in our neuroscience labs. Why? Because it touches at least three things that are central to any neuroscience lab: the research the lab produces, the skills its people build along the way, and the cultural and methodological norms that we expect labs to adhere to.

Back home, it didn’t take long for Claude Code’s impact to hit the lab. We had our annual retreat only three weeks after my Barbados trip, and it was focused on hands-on development of analytical pipelines that we had long wanted to implement. Having already become accustomed to agentic coding, some (including myself) managed to prototype a complex new decoding pipeline for our own data within a day—a development that unintentionally shocked the rest of the lab. From then on, discussions over lunch and dinner focused not on our projects, but rather on how agentic coding will affect us as scientists and our work in the lab. These conversations made me realize that we needed to develop a formal lab AI policy. Ongoing discussions with my team over the next few months, as well as research into public debates, helped us shape that policy.

O

ne big concern I heard at our retreat and soon after, mostly from Ph.D. students, was that AI will reduce the room for deep but time-consuming skill development. Students already feel constant time pressure; they’re competing to produce high-impact work with time-limited funding. If others use AI to fire off one output after another, will there still be patience for students to develop at their own pace? Will funding agencies be willing to pay for training, or will they consider not using AI to be too costly? Others grappled with the question of how much researchers need to understand of AI outputs and how to accurately check results. At the same time, many lab members expressed genuine amazement at how well AI can handle some time-consuming tasks. As a PI, I could not leave everyone to contend with these questions by themselves.

In the weeks that followed, lab members posted blog posts and articles about AI before we met again to discuss our lab’s policy. The first rule we implemented addressed what we felt was a core issue—the trade-off between human knowledge gain and AI use. There are, of course, cases in which AI can speed learning, but many have expressed the fear that human training will suffer. As one blog put it, “The machines are fine. I am worried about us.” Indeed, a recent Anthropic study found that developers who used AI while learning to code fared worse during later learning and comprehension.

Our first principle makes this potential trade-off explicit and asks everyone to manually complete tasks that build core intellectual skills, such as developing questions, building models and writing arguments. Trainees can hand off what they are less interested in learning or are already good at. Writing seemed a particularly slippery slope. By helping us with wordsmithing, AI can reduce the often-discussed barriers for non-native-speaking writers in an English-dominated academic publishing system. But AI writing assistants don’t just wordsmith, they often change content and can shift arguments and even the attitudes of their users. So we agreed that you must always draft a text yourself before handing it to AI, and carefully watch out for AI-introduced shifts.

The other principles followed naturally. Verify and validate: AI output sounds confident even when wrong, so know how you can falsify what a model produces. Write scripts that check the output rather than asking the model to check itself. Though such tests are not trivial to come by, Russ Poldrack and others have provided useful and concrete input on this process. Our next principle was to avoid risks. Participant data should not be shared with AI tools, and agents only get access to the folders they need. Hidden instructions embedded in seemingly harmless documents are a real risk, so researchers need to be careful when content with powerful AIs.

Relatedly, we agreed that people should invest time in learning to use the tools well, because output quality depends to a large degree on scoping and prompting. Finally, we agreed with the rules around authorship and responsibility that are now widely implemented in journals and conferences: AI is a tool, not a coauthor, and “the model said so” is no defense. You own everything you make public.

F

or me, the most important outcome was that we started an open conversation. It is not easy to navigate the many gray zones, and transparency about when and how a researcher has used AI is key. I now have regular and open discussions about the role AI played in setting up a particular model or text, and this has helped us identify which AI-assisted results need more scrutiny before we trust them.

For me as a PI, developing an AI policy also does something else. Because I spend far less time with the data and code myself, I must judge not just a result, but how much to trust it. This has become harder to do with AI in the loop. But the heightened transparency in the lab helps my meta-confidence, or the confidence about my confidence in a result. My hope is that in the long term, we can establish a culture that will let lab members use AI in any way that moves their work forward, including quick prototyping with limited understanding, while ensuring that we and our collaborators know what we understand and what we don’t. And that we will figure out where to do more follow-up work to solidify, or throw out, these preliminary insights. Our lab policy’s biggest effect wasn’t any single rule, but rather the creation of a culture in which we discuss AI use rather than hide it.

To me, it seems abundantly clear that moving toward such a culture is urgent. AI use among Ph.D. students is already near universal, and so are the worries discussed above. Legal scholars have long noted a treacherous loop in which the mere existence of a circumstance over time normalizes it, making it seem legitimate and just. AI use is on exactly this path: Whatever we all quietly start doing will soon be the norm. That is why we should not only have lab policies but decide as a field where we stand on the shifts in money, priorities and agenda that come with AI. Mathematicians have recognized this and responded collectively with the Leiden Declaration, and they, among others, have warned against making academic inquiry too dependent on technologies owned by a handful of corporations. The neuroscience community would do well to follow suit with their own norm-setting statement.