> home / perspectives / scientific ai policy
scientific ai policy
Last updated 25 Sept 2026
|
It’s my job to conduct science well and transparently. In this age of large language models and generative AI, I’ve started realizing that there is no going back to a pre-AI time. (Here I use “AI” as a catch-all, which does not include Machine Learning or Deep Learning methods, standard AI tools especially for the scientific work I do). And so, as scientists, we must look forward, creating structures that will help us use this new tool in an ethical yet practical way, without ceding our authority, expertise, agency, morality, or public trust.
Much like forming a group or starting a project, if we don't have structure, humans easily fall into the heuristics that flourish under inattention. Growing weeds in a garden is easy, but growing the crops and flowers you want takes a lot of hard work and intention. Therefore, I am providing here my own personal Scientific AI Policy (SAP) to guide how I use AI in my work, and to share that with others. Generative AI tools were not used, and will never be used, in the writing of this document. Useful further reading is provided in the References section at the end of my SAP. This living and iterative document is currently organized into five sections: (I) Thinking, (II) Coding, (III) Data, (IV) Writing, (V) Communication. Please note -- I do not address the ethics of the AI industry here, nor the land conversion, fossil fuel generation, or other environmental concerns with the outbreak of data centers (maybe in a future post). Using a problematic tool in productive ways to improve our planet, while still having moral reservations about its proliferation and plagiarism problems, is a complicated ethical question that I'm still wrestling with. See panel below. 
I think of this Matt Bors comic almost every day of my life now. https://thatmattborscomic.com
I. THinkingAI will not replace my own thinking. I don’t turn to colleagues and ask them to “give me a research idea”. I don’t ask anyone else but me to draft my papers, make connections between ideas, form opinions on complicated issues, read books for me, have difficult conversations, or decide what is right and wrong. In the same way, I don’t want to use AI to outsource my own reasoning, connections, and relationships. It takes time to do these things. It’s supposed to. Friction is where learning happens, and I need time to ruminate, percolate, and muse my way through things. I need time to examine my principles and beliefs, to become a better scientist and a better person. That’s life. II. CodingA lot of my work involves computer programming; knowing how my code works is very important to me. I will not use “vibe coding” in my scientific analyses. To me, this replaces my thinking brain in a way that I’m not comfortable with. I do want to leverage vibe coding for interactive apps to display my data or results, in cases where I don't know or care to learn the language. Instead, I will use AI tools as pair-programming and troubleshooting resources, akin to a deep search on Stack Overflow or turning to a friend for assistance. Both of these use cases involve providing reproducible, granular examples of the exact problem I’m having, not feeding my entire codebase into a machine. In some cases (e.g., unfamiliar compiled languages underlying R or Python code), I may instead provide a link to a specific file on GitHub that is throwing an error if I’m really struggling. This method ensures that I only provide AI tools with my code if it’s already publicly searchable on the Internet. AI will remain a tool in my toolbox; it will not replace real conversations, trial runs, error handling, or online research. I want to learn and grow as a programmer, and this cannot happen if I offload the hard work of coding. Often, I learn more from the search, trial, and error than I ever would had the answer been offered to me with no obstacles. I’d like to keep it that way. III. DataI will not provide AI with any form of non-public, real data; only public datasets or simulated examples that I have crafted to receive troubleshooting help with syntax and errors. This is especially important in my line of work, where elephant location data is heavily protected to prevent poaching. This policy may change at a later date depending on my institution’s AI policies, and whether I’m sure that institution-approved local AI instances will not steal my data or be used to train other AI models. IV. WritingAs of this edition of my SAP, I will not be using AI tools to assist in scientific publication writing. All words are my own, and manuscript text will not be uploaded to AI tools to protect my intellectual property, privacy, and morals. This last policy may change at a later date depending on my institution’s AI policies, and whether I’m sure that institution-approved local AI instances will not steal my writing or be used to train other AI models. I will use AI tools to help with literature searches, compiling relevant papers that I will personally sift through and, where appropriate, read thoroughly. The scientific literature in many fields has gotten so broad that there is no way to be aware of every useful paper; AI is only the latest in a line of tools, like ResearchRabbit and Scopus, that have curated literature searches for years now. All prompts used to assist in literature searches for official publications will be provided as an additional document for review. As of this edition of my SAP, I am not comfortable having AI summarize papers for me that I should be reading; that’s ceding too much of my own learning. V. CommunicationI will not use AI tools to generate photographs, graphics, or images of any kind. It is important to me that I use real (credited) photographs of the animals I study; that I include my own figures that I have generated; and that, especially when engaging with Skype A Scientist classrooms, I not provide a bad example for children and young adults on the proper use of AI. Finally, I will not use AI when there is a real human whom I could connect with first. If I have trouble understanding a topic, come across a thorny statistical problem, or want to bounce ideas around, I’d better do it with a real person first before turning to AI. Human connection is one of the greatest resources we have, and I don’t want to cede those relationships for one with a statistical model.
|
|
