A scientist who spent two years building a meticulous research archive watched it vanish in seconds after a single interaction with ChatGPT. The loss, which wiped out what he described as his entire working record, has become a cautionary tale about what happens when experimental tools are treated like stable infrastructure. I see his story as a stark reminder that the convenience of AI can mask basic risks that researchers once learned to avoid the hard way.
Behind the viral shock value is a quieter, more uncomfortable reality: this was not a freak glitch so much as a foreseeable failure of digital hygiene. The scientist’s ordeal exposes how quickly academic workflows have shifted onto platforms that were never designed to be primary storage, and how little many of us understand about what those systems actually do with our data.
The click that erased two years of work
The core of the incident is brutally simple. A researcher had been feeding his ongoing notes, analyses and draft ideas into ChatGPT, treating the interface as a kind of living notebook that he could revisit and refine over time. When he later adjusted a setting related to how his conversations were stored and used, the archive he had come to rely on disappeared, leaving him with no way to reconstruct two years of what he called “careful” work. In his own account, he describes realizing only after the fact that the system was never meant to function as a long term repository, and that the deletion was irreversible.
In a first person reflection, the scientist, identified as Jan, recounts how he had been well aware that large language models can sound confident while still being wrong, yet he did not anticipate that a single configuration change could wipe out his entire history with the tool. He explains that the conversations, which he had treated as a running lab journal, “could not be recovered” once they were removed from the service’s storage. That admission, captured in his detailed narrative of how two years of vanished, underlines how thin the line can be between a helpful assistant and a single point of catastrophic failure.
How a privacy toggle became a data trap
The trigger for the loss was not a dramatic system outage or a cyberattack, but a user facing control that was supposed to give people more say over how their data is used. Jan had been experimenting with ChatGPT’s “data consent” option, a feature that governs whether conversations can be used to train underlying models. When he changed that setting, his existing threads were removed from his accessible history, effectively erasing the only place where his evolving research notes existed. What looked like a simple privacy safeguard turned out, in practice, to be a delete switch for his working archive.
Coverage of the episode has emphasized how the scientist, who became known online through a piece headlined “Scientist Horrified as ChatGPT Deletes All His Research,” had treated the chatbot as a de facto lab notebook without maintaining parallel backups. Reports describe how he admitted he had “lost” two years’ worth of “careful” work and tied that loss directly to the way he used the data consent setting. The design of that control, which blends privacy, product improvement and storage behavior, left him with little warning that he was about to sever his only link to a major body of work.
Why a chatbot became a lab notebook
Jan’s reliance on ChatGPT as his primary research workspace did not happen in a vacuum. Over the past two years, many academics have folded large language models into their daily routines, using them to summarize papers, brainstorm hypotheses and keep running dialogues about complex problems. In his own telling, Jan describes how the tool’s conversational memory made it feel like a natural place to store evolving thoughts, since he could return to a thread and build on prior exchanges instead of starting from scratch. That sense of continuity blurred the line between a transient chat window and a durable record of scientific reasoning.
In his essay on how two years of, Jan notes that he had been cautious about the model’s factual reliability but had not applied the same skepticism to its role as a storage system. The interface encouraged him to treat each conversation as a persistent workspace, yet the underlying service was optimized for generating text, not preserving it. That mismatch between user expectation and technical design is at the heart of why a single click could have such devastating consequences.
The internet’s schadenfreude and a serious warning
Once the story surfaced, it quickly became fodder for online schadenfreude. Commenters seized on the irony of a scientist, someone presumed to be methodical and risk aware, entrusting his entire archive to a consumer AI product. One widely shared write up leaned into that tone, framing the episode as “Scientist Horrified” and highlighting how he had effectively outsourced his own record keeping to a black box. The reaction captured a familiar internet dynamic, where a personal disaster becomes a spectacle, even as the underlying issues are far more systemic than individual.
Yet the same coverage that indulged in Schadenfreude also underscored how common Jan’s behavior has become. A report by Victor Tangermann noted that the scientist had “lost” two years of “careful” work and tied that loss to the way he used ChatGPT’s “data consent” option, while also pointing out that he is far from alone in treating AI tools as central hubs for intellectual labor. The piece, published at 9:45 am GMT by Victor Tangermann, framed the incident as a warning shot for anyone who has quietly allowed a chatbot to become their only archive of notes, drafts or code.
What researchers can learn from Jan’s loss
For all the drama of a scientist watching his archive evaporate, the practical lessons are surprisingly old fashioned. First, no AI interface should be treated as the sole home for critical work, no matter how convenient its search and memory features may feel. Research notes belong in systems that are explicitly designed for storage and backup, whether that is a version controlled repository like GitHub, a lab’s shared server or even a carefully managed set of local files synced to cloud storage. ChatGPT and similar tools can still play a role, but as companions to those systems, not replacements.
Second, the episode highlights how important it is to understand what user facing controls actually do. Jan’s account of how his archive vanished after he changed a setting shows that even well intentioned privacy features can have destructive side effects if their implications are not clearly explained. His reflection on being “well aware” of model limitations but blindsided by storage behavior, captured in the detailed description of how archive could not, is a reminder that digital literacy now has to include not just how models generate text, but how platforms handle the data that flows through them.
Finally, Jan’s experience should prompt institutions to update their own guidance. Universities and research organizations have long issued policies on data retention, lab notebooks and backup procedures, but many of those documents predate the rise of conversational AI. As more scientists quietly move their thinking into chat interfaces, the risk is not only individual loss but also gaps in reproducibility and accountability when those records vanish. The stunned scientist who watched ChatGPT wipe his working memory is an extreme case, but he is also an early warning of a broader shift that the research world has not yet fully reckoned with.
More from Morning Overview