Statement on AI and Scientific Research
How this lab uses, and declines to use, AI tools in scientific work. It applies to our research, to what we publish, and to the code and data we release.
01 Principles
The people in this lab believe in the following principles
-
Tools amplify, they do not replace
In the hands of someone who already understands the problem, a tool is a force multiplier. It sharpens and speeds up work, but it cannot substitute for the knowledge required to use it well. For the same reason, a tool used without responsible direction can just as readily amplify bias, discrimination, and harm.
-
Tools do not make decisions
However capable a tool is, the judgment calls remain ours. A tool can inform a decision; it cannot stand in for the process of making one.
-
Some tedious work is part of the process, not an obstacle to it
Even where a tool could reliably automate a manual task, we do not concede that step by default. Certain work teaches us something precisely because it is slow and manual, and we protect the right to keep doing it by hand.
02 Statement
On the use of AI in our scientific work
Based on these principles, and on the work and thoughts of other researchers [C], we have developed the following statement.
“AI” refers to a diverse set of automated and semiautomated technologies including (but not limited to) predictive machine learning systems, automated transcription, generative models that produce images and text, chatbots built on large language models (LLMs), and coding agents that automate programming tasks. Some forms of this technology offer substantial benefits for our research, while other forms threaten our core values. [C]
The people in the lab consider two priorities equally important: producing reliable knowledge and supporting the development of knowers, namely students, teachers, researchers, and practitioners. Neither one comes first. There is no knowledge without knowers who exist in community with one another, and we are committed to sustaining that community so the practice of science can continue.
Papers and reports document our work so that others can learn from it and build on it. Software and data give that work a working, reproducible form, so others can run it, check it, and extend it. Neither is the goal. The goal is to improve our understanding of the world, to build our knowledge and skills as scientists, and to pass that knowledge and those skills on, so that the world benefits from them. Reliability and trust are the guiding principle behind both priorities: we need knowledge that can be trusted, and we need the knowers who produce it to be trustworthy in turn. We are concerned that the widespread and uncritical use of AI tools in science erodes the personal relationships and collective learning on which both priorities depend.
We are also concerned about the widespread use of proprietary models for producing scientific findings, which threatens transparency and reproducibility, cornerstones of robust science. As a rule, we do not rely on the conclusions of proprietary models as findings in their own right; any result we report must be reproducible on inspectable, fixed machinery, not on an opaque system whose workings we cannot inspect or rerun identically.
We recognize that many of the companies developing AI tools engage in unethical business practices, e.g., by training their models on stolen data and exploiting the labor of vulnerable people. We are alarmed by industrial development of power-hungry data centers during a climate emergency, and by the deployment of chatbot products that harm mental health. When we use AI tools, particularly those developed by profit-seeking companies, we are complicit in these harms. Open, local models already exist, and we expect them to become practical to run alongside proprietary ones. We do not build new models ourselves; our research is aimed at building applications that work on top of these open alternatives, so that dependence on proprietary products is a choice, not a necessity.
These values and acknowledgments inform our approach to incorporating AI in our scientific work. We agree with the position taken elsewhere [C] that many of the potential benefits and threats of this technology cannot be summarized in a generic lab policy; rather, the tradeoffs depend on the specifics of the project, the AI tool in question, and the training stage of the lab member.
Every lab member develops a personal AI policy in consultation with their mentors, revisited each semester. We support conscientious refusal as a valid choice, and collaborative projects must respect and communicate each member’s policy from the outset.
03 Guidelines
Example guidelines
How we currently handle the situations that come up most often. We use these tools; what we do not hand over is the decision, the confidential material, or the responsibility for the result. The examples are a personal policy to start from, not a rule that settles every case.
- Peer review
- The assessment is ours to make and to sign, and a manuscript or grant under review is somebody else’s confidential work: it does not go into an AI tool we do not control. Once the review is written, a tool may help with language and clarity, and may raise further points to consider, which the reviewer weighs and takes or leaves. Nothing private or de-anonymizing is shared in the process.
- Presentations
- Building a talk is how the presenter works out the argument, so the drafting is kept manual on purpose. A tool could produce the slides; doing the work by hand is what produces the understanding. Once the message is written, a tool may template it, summarize it, or translate it.
- Coding
- We use coding agents to turn precise, already-specified requirements into code, and to search for bugs, accepting that such a search is not exhaustive and checking what it finds. The technology choice and the algorithm are decided by the researcher first: both are the design work, and an algorithm we did not design is one we cannot claim authorship of. Tests are written first, whatever writes the code.
- Diagrams
- Drawing a diagram is part of understanding what it shows, so the first pass is made by hand even where a tool could produce one. Once the person making it has understood and described the information, a tool may turn that description into a diagram. The result stays manually editable: a diagram nobody can correct by hand is one nobody will correct.
- Finding citations
- A tool may suggest candidate citations. Every citation in our work is then read and assessed by the author: we confirm that the paper exists, that it says what we claim it says, that it supports the point it is attached to, and that the bibliographic data is correct.
- Accountability
- Whatever tool was involved, the result carries our names and we answer for it. A tool is never the reason a number, a citation or a claim was wrong.
We will discuss and revisit this statement annually and revise it in light of our experiences and new information.
04 Sources
Acknowledgement and references
This statement adapts the original text of the Crockett Lab statement, and is informed by our own work on AI and scientific research. Both are below.
-
Crockett Lab Statement on AI and Scientific Research
. Zenodo, . doi:10.5281/zenodo.22144397
-
Towards Reliable Conversational Data Analytics
. Proceedings of the 28th International Conference on Extending Database Technology, EDBT 2025. doi:10.48786/edbt.2025.78
Last update: .