Guidance on the use of AI in the Kartzinel LabTyler KartzinelLast updated January 2026.
Artificial intelligence is increasingly useful as a tool to improve our research and learning. We use it to troubleshoot code, polish writing, get good ideas about how to visualize data, create document templates that save time on busywork… But at the same time, we must be cognizant of legitimate concerns about the accuracy of information it can provide, its ability to reuse confidential information that we disclosed in chats, and the risk of short-circuiting our own creative uses of the scientific method.
This post summarizes rules that lab members should follow when using AI in their work. I do not want to regurgitate the types of dry, legalese we are provided by our employer--rather I will attempt to illustrate the fine-line we have to walk to ensure we are using the tool appropriately while minimizing the risk of unintended harm. I will summarize reasons for concern using language familiar to biologists and conservationists broadly. Some of the details are specific to researchers at Brown, but I believe the information is readily transferable and I welcome others to use this document as a template for their own policies. Please read on... Rules from the lab
1. Never upload or paste your data, code, or scientific writing into a commercial AI platform. Doing so would risk violating legal requirements concerning confidentiality and the use of original data, while compromising our collective efforts to advance the bounds of knowledge.
2. Always use Brown University’s internal LibreChat system if you want to use AI to check your work or save you time. This system provides access to many of the same AI tools that researchers often access online—ChatGPT, Claude, Canva AI, etc.—but it does not share our data with external tech companies. 3. If you use commercial AI platforms—including Google Search with AI—do so with extreme care to ensure you do not violate Rule #1. Only do so if you can be certain that you are entering completely generic prompts that cannot be connected to your work in the lab—if there is any doubt, use LibreChat instead. 4. Be able to independently verify all insights and information you obtain from these platforms when asked. The baseline quality of information these platforms can provide is limited based on the quality of information sources they are using—they make mistakes and have the potential to mislead with some frequency. It is worth remembering that our job is to generate information that no one has ever had before and thus you cannot rely on these tools for information that you are not independently verifying yourself. Illustrating the risk of making inappropriate disclosures without intending to
Read each of the four pairs examples and consider differences between each query is constructed:
How would you characterize the difference in risk?
In each of the first cases, you see that there are ways to create prompts that enable you to use these kinds of algorithms as time-saving tools without compromising the confidentiality of your work in the lab.
In each of the second cases, you see that it is all too easy to provide third-party commercial platforms with information that you may not be allowed to knowingly disclose--even if that is not your intent. This is the difference between using AI as a tool to accelerate our research and get past roadblocks versus using AI to do things for us. The second case clearly has the potential to cause frustrating experiences and lead to errors—reason enough not to outsource your responsibility to a commercial chatbot—but that is just about the user's short-term experience. The risk of making unintended disclosures can be far more significant and long-lasting. Below, I will illustrate reasons for concern. Awareness will help ensure we are able to use the tools appropriately and are quick to recognized inappropriate types of interactions. Reasons for concern
Disclosing confidential, protected information. We work with protected plants and animals; ethical considerations, and sometimes legal requirements, prevent us from disclosing sensitive information that could jeopardize them, their habitats, or the people who work to protect them. For example, we redact or coarsen information about where we sample protected wildlife populations when we publish results in order to avoid inadvertently aiding poachers. However, AI platforms store information you provide in order to provide information to others--this is a significant concern. Brown also has categories of data risk that may apply: See the "Data Risk Classifications" link below.
Getting scooped, plagiarism, missing out on credit. We have ethical obligations to one another, and to our funders, to ensure the utmost integrity in our research. We strive to share all scientific information quickly and responsibly—and especially with respect to tax-payer funded work we are required to do so—but we undermine our own best efforts when our work is shared with others before we can fully verify results with our careful quality control procedures. This can happen, even unknowingly, if another user enters a prompt and the response they get is based on unpublished information originating within our group. Other individuals may use that information as if it were original or their own, with no opportunity to verify it using the peer-reviewed literature or to properly credit us for the idea. In some cases, AI bots can search and find data or code and package it in ways that make it all too easy for someone else to claim credit. See the related link below about "IP" (Intellectual Property) at Brown. Failing to verify results, perpetuating errors. Each of us bears significant responsibility for ensuring the accuracy of our results—inclusive of how our data are recorded, how samples are handled, and how data are analyzed, visualized, interpreted and communicated. Failure to do so can result in irreparable harm to our reputation, careers, and field writ large. Mistakes happen in research, but we can only correct them if we are attentive to detail—and the practice of attending to detail requires time. Failing to verify code, graphics, citations, or translations of original text are just a few of the ways that overreliance on these tools introduces the risk of perpetuating damaging errors. Social and environmental harms. The explosion of AI data servers is taxing local electric systems, water supplies, and the social fabric of surrounding communities all over the world. Many of these facilities are being constructed in regions with few environmental protections, exacerbating the associated environmental harms. Life is not without impact—we should use all the tools available to maximize the quality of our science and progress in our careers—but we should be cognizant of impacts and strive for moderation. Failure to learn, metacognition. In science and academia, your brain is your greatest asset. It is what provides you with genuine intelligence. Evidence shows that we learn best by recalling information and applying it in new situations. What may feel like a time-saving opportunity that advances your work could ultimately undermine your goals if it becomes a tactic to avoid investing sufficient effort in the difficult task of learning and improving with time. Therefore, I recommend learning about the “Desirable Difficulty” concept—it will make you a better student and teacher, who is prepared to use AI effectively while managing its associated risks. It is ultimately our job to recognize what advice we get from AI is based on useful information, and learning to do this efficiently and effectively is going to become an increasingly important part of academic training. To illustrate and then solve the problem...
Links to policies and resources at Brown
Lab members should use LibreChat by default. Learn about this and appreciate what it does for you—every time you use it!
LibreChat: https://librechat.ccv.brown.edu/login
Additional links relevant to the use of AI at Brown University:
0 Comments
Your comment will be posted after it is approved.
Leave a Reply. |
RSS Feed