This page documents our current thinking around use of generative AI tools (e.g. LLMs) in the lab (this excludes use of machine learning methods for data analysis, which we use routinely). This landscape is evolving very quickly, and we anticipate the need to keep these guidelines under review.

The first general principle is that – as with elsewhere in the lab – we are a broad church, and Steve does not seek to limit which tools we use or how we use them (see also the guidelines on scientific software use here).

The second general principle is that we prioritise understanding and trust in our science (again as documented elsewhere in this wiki).

Within these broad guidelines, we can make some general recommendations:

• Lab members should not be forced to use AI tools if they do not want to; equally lab members should feel free to explore workflows in which AI is useful.
• Understanding should be prioritised at every step of the research process. You should always be able to explain your thinking, your hypotheses, your analysis plans, and each line of your code.
• As we know from metacognition research – good confidence calibration (knowing what good looks like) is only possible when first-order skills (e.g. programming) are firmly in place. Be mindful of AI usage eroding these skills, or preventing you from gaining them in the first place. “The most precious resources of our profession are students and ideas, and these we nurture with great care.”
• Generative AI is not like deterministic software; it is probabilistic. The path it takes in coding, writing or idea generation will be stochastic. Always keep this limitation in mind when using these tools (see further guidance from Russ Poldrack in his advice on AI-assisted coding).
• Be careful about how (or if) you use AI to help with writing. Writing IS thinking – and if we use AI to help with this (e.g. early on in the process), we risk limiting the depth of our own thought process (we do not typically run the counterfactual of writing alone, and seeing which is better!)
• Be mindful of the social consequences of AI usage for writing and presentations. You might think that asking an LLM to help make slides is innocuous and saves you time, but it can create the impression in the audience that less care and attention has been given to the talk.
• Ensure substantive AI use is disclosed in papers and outputs from the lab, and discuss with Steve in advance if you are unsure whether AI usage is appropriate in part of the research workflow.
• Never give AI tools access to participant data or identifiable information.
• Avoid use of proprietary (closed) models as part of a research workflow (e.g. sentiment analysis), as these are not transparent or reproducible. It is relatively straightforward to set up local open weight models for this purpose (e.g. downloaded from Hugging Face).

The guidelines on this page were inspired by policies that have been developed by the labs of Nico Schuck, Liad Mudrik, Todd Gureckis and Molly Crockett, and draws on general principles for AI-assisted coding developed by Russ Poldrack.