ZEAL AI Policy

Motivation

Many university systems have made concerted efforts to promote the use of AI, including for purposes that academics generally disagree with, or for which there is not consensus. Given the ethical issues, environmental costs, and cognitive consequences that surround the use of AI tools, these very permissive policies for students and researchers shift responsibility from the university to the individual research groups. Regardless of your personal uses of AI and LLMs, any of your work with ZEAL must adhere to the policies outlined in this document.

Skill development is difficult, but ultimately better positions you to engage with the world around you in a critical, self-possessed way. These critical thinking skills, and specifically the ability to use them to navigate unknowns and for on-the-job learning, are one of the most important transferable skills you can develop, regardless of your career path of interest. While the effects of AI usage on skill development and cognitive performance are still being studied, there is a growing body of evidence that uncritical use of AI tools can lead to long-term erosion of skills, reductions in mental function, and weakening memory.

When you are reading papers, working on a draft for a manuscript, writing code for an analysis, or refining any of the other skills central to the work we do here in ZEAL, you may feel some degree of mental friction. This friction will lessen over time as you develop the necessary skills. AI tools provide instant gratification, bypassing this friction, but this “cognitive offloading” blocks your skill development. Further, without these skills, you lack the expertise needed to critically evaluate the quality of the output from the AI tool. It’s for these reasons (amongst others) that we are approaching the use of AI with such skepticism.

Policy Framework

For now, AI tools in our lab will essentially never be recommended, and in many instances, are still outright prohibited. Blanket policies like this are, however, likely too restrictive. We will instead use a framework based on task type and expertise, as outlined in Krishnan 2026. This framework breaks tasks into two categories, and describes recommendations for two types of users at different stages of skill development. Our AI policy starts with this framework because it centers skill development as a high priority goal, emphasizing that the expertise required to determine when, where, and how AI tools are useful is blocked by these same tools when they are used too early.

	Trainee (developing expertise)	Expert (established expertise)
Mechanical automation (executing established procedures)	With caution and verification	Acceptable with verification
Cognitive tasks (reasoning, synthesis, judgement)	Never	With caution

As a trainee (undergraduate and graduate students alike), your initial work must be your own - AI tools are prohibited until you have demonstrated (to Matt) the skill and expertise required to critically evaluate the output. You can discuss with Matt when / how this can be demonstrated. After this point, you will take more ownership of your workflow - if you choose to try using AI tools, you’re equipped with the expertise to evaluate the outputs. In all cases, you are responsible for the quality of the work.

Permission to use AI tools can be revoked if you do not maintain the ability to critically evaluate scientific work (this could be your own work, or the work of others).

Key Points

In short, you are here to learn - ZEAL provides compensation for your time under the assumption that your work reflects genuine effort to develop and refine skills. And while we value efficiency and productivity, forfeiting your work to an AI agent or LLM is a deviation from this agreement.

Below we summarize the main points, and the scenarios when LLMs and AI approaches can and cannot be used:

• Sensitive lab data should never be uploaded to an AI platform. This includes, but is not limited to, raw data, unpublished manuscripts, and grant applications.

• Writing: You cannot use AI or LLMs to draft manuscripts or reports - I expect your writing to reflect your own thinking. You may request permission from Matt to use AI to lightly edit text after (and only after) initial drafts have been reviewed.

• Images: Do not use AI image generation tools. This includes using AI for data visualizations, conceptual schematics, images for presentations, etc. As with your writing, the process of designing these elements is important for your own development.  Further, given that these tools were largely trained on the work of artists against their will and without their permission, there are particularly difficult ethical issues surrounding this use.

• Literature review: You can use AI agents or software to supplement literature searches, but this cannot be how you do most of your literature review. Further, I strongly recommend you completely avoid AI summaries of papers. You must read papers, and in turn, use your own initiative to ensure your coverage of a topic is complete.

• Coding assistants: You can use coding assistants for data analysis, but see below for important details on accountability.

• Use of AI in research: Any use in research (e.g. computer vision, video analysis, etc.) is only permitted after extensive human validation. The development of the workflow needs to be thoroughly documented.

• The lab will not pay for individual licenses.

Confidentiality and data protection

• If you choose to use AI tools, you should preferentially use AI tools that don’t train on your data.
• Privacy and data protection are extremely important for research groups. OpenAI, for example, has a terrible track record on security and privacy. Never ever upload sensitive material or any personal data. If you wouldn’t give a stranger access to the information, you should be hesitant to provide access to an AI tool or LLM.
• See here for another recent example of the risks associated with providing AI tools with access to personal information.

Writing

• Outside of your lab work, writing is one of the most essential skills for any researcher develop. Because of this, our policy on the use of AI for writing is the most restrictive. Learning to write is not just about producing reports and papers more quickly, but about developing how you think through analyses. Writing is thinking.
• AI in writing is inherently risky for your career.
• The general policy for writing is that you are better off doing the writing (and most importantly, the thinking) yourself. We prioritize quality over quantity, and ensuring your ‘voice’ is reflected in your writing is a major component of creating high quality scientific products.
• While AI writing is improving, people can still see through it. There is often a visceral reaction to AI writing in professional environments, and uncritical use of AI writing endangers not just your reputation, but the reputation of our entire research group.
• You should be able to point to every word and explain why you chose it - even (or especially) if you were only ‘editing’ your writing with AI.  

Purpose	Allowances
Concept outline	Allowed as a portion of the literature search. Not allowed for outline generation.
First draft	Not allowed.
Major revisions	Not allowed.
Light edits and readability	Allowed for light edits, so long as: 1) Matt has given you approval, and 2) all changes are personally verified.

References

• Never write or edit reference sections or BibTeX with AI tools. This is non-negotiable: it gets papers banned from preprint repositories and flagged by journals when submitted for peer review.
• So long as your Zotero library is well-curated, there should never be a need to use AI to generate reference sections.

Coding agents

• LLMs are getting to be quite good at generating code. In general, however, trainees should not use AI for coding until you are knowledgeable enough to interpret the code generated. And no matter what, do not upload raw data to AI tools!
• In all cases, you need to have a thorough understanding of your code. This means that you should be able to explain exactly what every line of code is doing, why certain decisions were made, how products from one section fit into other sections, etc. There will be consequences if you use AI tools to code but cannot then explain what the code is doing.
• If you can’t code, an AI tool is unlikely to help in the long-term. Instead, take advantage of opportunities to learn from your peers - developing your coding skills is another important aspect of working with ZEAL. Consider proposing code review sessions for the lab, or ask to cover topics during your one-on-one meetings with Matt.
• If you are going to use AI assistance with code, you should consider these tips.

Summarizing Literature

• AI is terrible at summarizing the kinds of papers that are typically useful for our work - don’t use it for that. Moreover, while the temptation may be to ‘boil down’ a study into easily digestible statements, learning to do this yourself is yet another critical skill for researchers.
• As you’re getting started, it might take you several days of dedicated study to pull out the main points from a single paper - after a few years, it will take you one hour. This is a skill that takes practice to improve at. Not only is AI terrible at summarizing the nuanced information we need to be getting out of papers, but using AI will effective prevent you from building up the skills to do this yourself.

Peer review

• In no case should you use AI for peer review. AI agents are not our peers and have no place in that process. The entire system of peer review is built on the utility of human perception. An AI agent easily spits out what sounds like an intelligent review, but as with other forms of writing, these reviews rarely hold up under scrutiny.
• I’m under the impression that much of the temptation to use AI for peer review has to do with the stigma around saying “I don’t know what’s meant here” or “I don’t understand this section”. Remember, peer review is about making writing more accessible to an audience. If you are the intended audience, and some point is confusing or isn’t clear, it’s an issue with the effectiveness of the writing. The most useful feedback tends to be about sections that could be clarified or better explained.

Disclosure and accountability

• Disclosure of AI usage needs to happen upfront and immediately.
• Internally, you are fully accountable for your use of AI and LLMs. This includes but is not limited to the information you transmit via code, slack messages, emails, etc. Don’t share anything within the lab that has not be thoroughly vetted.
• Externally, Matt and the lab as a unit will be perceived as responsible for your AI use. It is important that you always disclose when you’ve used AI.