By Yu-Tian Xiao

Back in medical school, I joined a journal club on clinical trials and biostatistics. The professors and statisticians used to say that plenty of doctors look like they read research papers, but many stop at the title and the abstract, where authors present the cleanest version of their story. Only a fraction of them read the full text, and even fewer pull up the trial protocol and the supplementary materials to read them side by side.

But the devil is in the details. I still remember the shock that hit me one afternoon in 2017, when I opened a trial protocol for the first time and found page after page blacked out for confidentiality, some sheets almost solid black ink. That quietly changed how I read medical literature. Later, when I started my training in computational biology and bioinformatics, the skepticism turned into a daily reflex. Whenever I scanned a new paper, I jumped straight to the Data/Code Availability section. If the raw data and scripts were sitting in an open repository, the paper was usually worth taking seriously; if not, it usually stopped being useful to me.

That was how I learned to dread a familiar six-word password, ‘Data are available upon reasonable request’. Whenever I saw that line, my interest evaporated. It sounded like an open door, but in practice it was usually a dead end. I found it disheartening that journals gave that password a pass, especially for resource-style papers whose job was supposed to be making a dataset or tool open to the scientific community.

So when journals started responding to generative AI (GenAI), I felt a familiar sinking feeling. Editors at the Journal of Medical Ethics have argued for a dedicated GenAI use declaration printed in the paper itself, rather than a casual tick-box in the submission portal. Even then, we risk repeating old mistakes if the policy relies on unverified self-reporting. Under the proposed options, authors can simply affirm that AI was used only for basic copyediting of human-drafted text. Under publish-or-perish pressure, ‘copyediting’ will stretch far past grammar. State-of-the-art models can already automate scientific processes from conception to drafting end to end, and I have no doubt that near-frontier models will soon do the same. An unverified declaration with no record to inspect is not transparency but another costless password.

This is what drove my commentary in JME Practical Bioethics, written for the call on death, authorship and GenAI. The claim there is practical. If GenAI rules rest only on unverified declarations, they will decay the way data sharing statements did. What we need is not an unreadable dump of every prompt, but a tangible record that leaves the computer before publication, tied to a named person who can still answer when a reader knocks. I call that answerability, the difference between signing at submission and still being there afterward.

I tried the experiment on the commentary itself by depositing my GenAI use record on GitHub and Zenodo. Doing it taught me something the declaration box never captures. The models and workflows I used for early drafts had already changed by the time I handled revisions, and neither setup looks anything like what I would reach for on a daily basis today. Claude Cowork never allowed a clean export from the very beginning. What’s more, on September 16, 2026, Anthropic announced that Claude Cowork and Chat are merging into one Claude, and therefore the workflow I had logged was already becoming a historical relic even before the paper appeared online.

This is a perfect example of why reproducibility is not out of reach only because GenAI models are stochastic in nature. Even with clear documentation and an open deposit of the kind I made, it is extremely difficult to retrace the path and arrive at anything close to the same process. Closed-source platforms and models often set up barriers for exporting a session cleanly, and the working harness and environment can vanish on any given day. So the point of depositing is not to guarantee a perfect rerun. It is simply the minimal baseline that makes a GenAI use declaration mean anything at all. Without a deposit, the section is basically empty. With one, at least there is a footprint to examine and someone named who can still answer.

Without a deposit, I think I already know the script. A month after acceptance, many authors will not remember how they used the tool or the model. Even if they do, they likely already have lost access to the accounts or session histories. Either way the black box wins. We spent decades watching one seemingly polite excuse hollow out open data sharing, and the GenAI use declaration shouldn’t be its heirloom.

Paper title:  Available upon reasonable request: the authorship function that was already dead

Author: Yu-Tian Xiao

Affiliation: Center for Cancer Research, Medical University of Vienna and Comprehensive Cancer Center, Vienna, Austria.

Conflicts of Interest: None to declare.

Social Media: @Bobby_XiaoYT; Linkedin; Bilibili: 2392617