I.D.E.A. Lab Generative AI Usage Guiding Principles & Policies
(modified from The Polldrack Lab @ Stanford, The Music Lab @ U Aukland, and the Carlson Lab @ Yale)
links: https://www.poldracklab.org/labguide/research/ai_policy.html ; https://bsky.app/profile/mehr.nz/post/3mq3kq244pk2a ; https://www.carlsonlab.bio/thoughts ; see also https://brettcodes.com/im-done-using-ai/
Generative AI usage guiding principles:
Generative AI tools, capabilities, and implications are rapidly changing, and so may our policies.
We must be vigilant to the AI hype cycle, and we should consider AI’s environmental impact, its exploitation/theft of scholarly and artistic work, its potential security/privacy risks, and the extent to which its closed, proprietary source code and overall opacity conflicts with open-science principles, before engaging with it. AI’s historical and present harms must be weighed against its potential benefits.
The purpose of basic science is to advance human’s understanding of the universe; the goal of the I.D.E.A. lab is to advance our own understanding, and other human scientists’ understanding, of cognition. Said more simply, in basic science it is critical that humans do some things ourselves.
Thinking is an essential aspect of writing, and the process of writing both illuminates and facilitates our thinking. If we offload our writing to generative AI tools, we thus offload our thinking. This conflicts fundamentally with sound scholarship.
In basic science, generative AI tools should (at most) supplement and accelerate human expertise, not replace it. It’s especially important for graduate student trainees to limit their reliance on AI tools, as they’re just beginning to build their expertise and are at risk of being de-skilled.
Doing science can be hard, and completing graduate school can be hard; nonetheless, learning to do challenging things can be deeply satisfying and pleasurable, and so scientists and students should be careful not to rob themselves of such motivating experiences by offloading challenging tasks to generative AI tools.
To the extent that the writing produced by generative AI models reflects the ideas and words of previous publications that the models were trained on, users must weigh the risks of career-damaging accusations of plagiarism. As Colin Carlson notes, “Plagiarism is subjective. Plagiarism is a continuum. Not all plagiarism accusations are made in good faith.” Because most scientific publications result from collaborative efforts, one author’s use of generative AI in writing could put many co-authors’ careers at risk.
Students should not assume that generative AI tools are necessarily the best ones available for the specific task at hand, especially if verifying their results is especially time-consuming.
In our scientific work, accuracy is always much more important than speed. It is better to be careful—and thus more likely to be right—than it is to be fast or first.
A human must ultimately understand and take responsibility for all outputs that go into a scientific publication. Any generative AI outputs should be validated and verified just as one would for an output from a research assistant or a novel R package or shinyapp.
Generative AI usage for any aspect of a research project is a critical decision that should be discussed beforehand with the full research team, including Mike.
Generative AI usage policies:
Generative AI tools may not be used to draft narrative text for manuscripts or student milestone projects. Lab members are expected to manually draft all the text that they put into a paper; no copying and pasting of text from an LLM into a manuscript is allowed without explicit discussion and agreement.
Although generative AI tools might have utility for copyediting text that a human has created, graduate student trainees need to beware the slippery slope from having AI “clean up” their writing to having it change their writing style to having it produce their writing. Note also the potential risk of plagiarism accusations if generative AI text too closely resembles other authors’ work.
Generative AI tools may not be used to draft revised text for the lab’s manuscripts or student milestone projects.
If the lab approves of an LLM to be used to review a lab’s manuscript prior to journal submission, then it should be prompted with an admonition that the LLM should not rewrite the text; rather, the author must rewrite any relevant sections manually based on LLM feedback.
Generative AI tools may not be used to create reference lists for lab manuscripts or student milestone projects
All reference lists (regardless of how they are generated) should be double-checked to ensure that the bibliographic information is correct, since hallucinated references may infiltrate non-AI reference databases, and penalties for including hallucinated references in articles and grants are becoming steep.
Because many journals are moving to require disclosure of AI usage, it is essential to record all the ways in which generative AI is used in a research project and to save all prompts and output for posting to the project’s OSF page.
Do not provide any lab dataset to a generative AI tool, as this is a security risk.
AI tools should not be used to directly generate data tables/figures (vs. generating code to produce tables/figures in a reproducible and auditable way)
AI-generated code should be tested to ensure that it accurately solves the problem at hand; see Better Code, Better Science for more on this.
Do not let generative AI tools summarize scientific articles for you in your lab work or your milestone projects, especially those you have not yet read.
AI systems cannot replace human participants in empirical research.