Personal AI Use Charter

Generative AI has fundamentally altered research practices at a pace that outruns any collective conversation about its implications for research. Over the course of my PhD, the tools evolved from unreliable sentence-completers to systems capable of deploying full server-side applications with polished interfaces within minutes. Part of this evolution is genuinely welcomed: doing a PhD means caring deeply about a question, and anything that reduces time lost to debugging or figure formatting is a legitimate gain. AI now makes it possible to explore new questions way more easily be it by a data analysis that would otherwise have taken days just to see what’s promising or a quick check on complex maths. This ever widening spectrum of applications changes the way we do research and the value-added of the researchers in the knowledge production process.

The deepest danger is, to me, epistemic narrowing. A good researcher’s value lies in opening many doors and closing them progressively through contradiction, iteration and failure. AI systematically works against this: it offers one well-packaged proposition, and the path of least resistance is to accept or at best marginally adjust it. A second risk is skill attrition. Growing up coding by hand, even when tedious, meant accumulating judgment about what good code looks like and what decisions matter inside a model. Delegating this progressively has eroded that judgment in ways I notice only when I need to reintervene. The concern extends beyond code: writing, brainstorming, critical reviewing are all skills that we could looose in the process.

Finally, where identifying generated papers or student homeworks was obvious even up to last year, it has now become impossible to separate the human written from the 100% generated and the reasonable AI-assisted. Still, as a researcher I don’t want to spend time reading papers just reassembling exising pieces or even worse putting my name on work whose integrity I cannot vouch for.

Existing disclosure frameworks in journals are inadequate to address any of this. A declaration that AI was used “in the writing process” as now standart with most editors says nothing about in which steps, to what extend and with what ethical care generative AI was used.

This charter defines the rules I set for myself, as a working document subject to revision. The goal is both to set limits to myself in the use I make of generative AI tools in my research journey but also aims at giving better understanding to my coauthors and readers of how they can treat my production outcomes

1. Models and core functions: I always write the code of my models and the core algorithmic functions of any coding project myself. AI can assist with peripheral implementation but not with the logic that constitutes the scientific contribution.
2. Writing: I always produce a complete first draft before any AI intervention. When AI is consulted, the prompt explicitly instructs not to introduce new elements or change structure, only to improve fluency or flag inconsistencies.
3. References: I decide alone which sources make it into my work. I read and summarize all the literature I use by myself, explicitely writing down a critical analysis of the material and how I could be using it.
4. Policy implications: The normative step from results to recommendations is the highest-stakes intellectual move. I don’t delegate any part of it.
5. Transparency: For every task I delegate to AI, I keep a note of what was prompted, what the output was and for what I used it for. This log is private but maintained throughout the project.
6. Mails: I write my own mails. More generally none of my professional communications are written with AI.
7. No AI-assisted review of others’ work: When asked to review work of colleagues, AI intervenes at no time.
8. Co-author awareness: When working with co-authors, the rules of this charter are disclosed at the start of the collaboration. Any divergence in practice is discussed.

I am more than happy to discuss any of this, and hope to extend it with time.