Key Takeaways
- Agentic AI tools such as Claude Code can dramatically accelerate neuroscience research, turning tasks that once took days into minutes‑long prototypes.
- Rapid adoption sparked immediate concerns among trainees about the erosion of deep skill development and the pressure to keep pace with AI‑generated output.
- The lab responded by drafting a formal AI policy that foregrounds transparency, skill‑preserving practices, verification, data safety, and proper attribution.
- Core principles include: manually completing intellectually formative tasks before delegating to AI; validating model output with independent tests; restricting data access; investing in prompt‑engineering competence; and treating AI as a tool, not a co‑author.
- Beyond rules, the policy’s greatest impact has been cultivating an open lab culture where AI use is discussed rather than hidden, boosting meta‑confidence in results.
- The author urges the neuroscience community to follow fields like mathematics in establishing collective norm‑setting statements to guard against over‑reliance on corporately owned AI technologies.
The Barbados Demonstration: A Glimpse of Agentic Speed
In early March, while attending a meeting on the Caribbean island of Barbados, Konrad Kording stood before roughly thirty principal investigators in swimming trunks and live‑demoed Claude Code. Using a projector barely visible in the broad daylight, he asked the model to build a web app, and within minutes the application 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,” the author recalls. The scene crystallized not only the scope of change but its breathtaking speed, turning what would have been days of work into minutes of iteration.
Rapid Adoption and Early Experiments in the Lab
Back at the neuroscience lab, the impact hit almost immediately. Three weeks later, during an annual retreat focused on building analytical pipelines, several team members—already accustomed to agentic coding—prototyped a complex new decoding pipeline for their own data in a single day. This feat “unintentionally shocked the rest of the lab,” prompting lunchtime and dinner conversations that shifted from project updates to debates about how AI would reshape scientific practice. The urgency of these discussions convinced the lab that a formal AI policy was needed rather than leaving individuals to navigate the terrain alone.
Emerging Concerns Among Trainees: Skill Development vs. Speed
One of the most vocal worries came from Ph.D. students, who feared that AI would shrink the space for deep, time‑intensive skill building. As the author notes, “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?” Trainees also questioned how much understanding of AI outputs is necessary to verify results reliably. Simultaneously, many expressed genuine amazement at how well AI could handle tedious chores, creating a tension between efficiency and the preservation of foundational expertise.
Formulating the Lab AI Policy: Transparency as a Cornerstone
In response, the lab instituted its first rule around the trade‑off between human knowledge gain and AI use. The policy explicitly asks everyone to “manually complete tasks that build core intellectual skills, such as developing questions, building models and writing arguments,” while allowing trainees to offload work they are less interested in mastering or already proficient at. Reflecting a common sentiment, one lab member’s blog warned, “The machines are fine. I am worried about us.” The policy further mandates that any text must be drafted by the researcher before being handed to an AI writing assistant, with careful monitoring for AI‑induced shifts in content or argument.
Core Principles: Verification, Data Safety, and Prompt Competence
Additional principles followed naturally. Because “AI output sounds confident even when wrong,” the lab requires verification and validation: researchers must write scripts that falsify model outputs rather than relying on the model to self‑check. Citing methodological work by Russ Poldrack and others, the policy stresses that such tests, while non‑trivial, are essential. To mitigate risk, participant data may not be shared with AI tools, and agents receive access only to the folders they strictly need, guarding against hidden instructions embedded in innocuous‑looking documents. Finally, the lab agrees that investing time in learning to scope and prompt effectively is crucial, since output quality hinges largely on these skills.
Authorship, Responsibility, and the Limits of AI Attribution
The policy aligns with emerging journal and conference standards regarding authorship: AI is treated strictly as a tool, not a co‑author, and the excuse “the model said so” is inadmissible. As the author puts it, “You own everything you make public.” This clause reinforces personal accountability and discourages the diffusion of responsibility that can arise when opaque model outputs are presented as indisputable facts. By anchoring responsibility in the human researcher, the lab safeguards scholarly integrity while still leveraging AI’s productivity gains.
Cultural Shift and Meta‑confidence: Discussing Rather Than Hiding AI Use
For the principal investigator, the most valuable outcome has been the initiation of open, regular conversations about when and how AI contributed to a model or piece of text. This transparency has helped the team identify which AI‑assisted results merit additional scrutiny before being trusted. Moreover, because the PI now spends less time directly with data and code, judging trustworthiness has become more challenging; yet the heightened lab discourse bolsters “meta‑confidence, or the confidence about my confidence in a result.” The author reflects, “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.” This cultural shift aims to let members employ AI for rapid prototyping while maintaining a clear map of what is understood and what remains provisional.
Call for Field‑wide Norm‑Setting: Learning from Mathematics
Looking beyond the single lab, the author warns that AI use among Ph.D. students is already near universal, and the accompanying anxieties are widespread. Drawing on sociological insights, they note that “the mere existence of a circumstance over time normalizes it, making it seem legitimate and just.” Without deliberate intervention, quiet adoption will become the de facto norm. The author points to mathematics, where the Leiden Declaration exemplifies a collective stance against excessive reliance on corporately owned technologies, and urges neuroscience to follow suit with its own norm‑setting statement. Such a statement would help the field deliberate on shifts in funding, priorities, and research agendas, ensuring that AI serves scientific inquiry rather than subverting it.
How, when and why to use agentic AI in our neuroscience labs