When two years of academic work vanished from a digital workspace, the loss became more than a personal mishap. It turned into a case study in how scholars are navigating — and sometimes misjudging — the promises and limits of generative artificial intelligence.
A Workspace That Suddenly Went Blank
For Marcel Bucher, a professor of plant sciences at the University of Cologne, ChatGPT had become a daily professional companion. Over roughly two years, he used the tool to help structure grant proposals, refine manuscript revisions, prepare lectures and draft examinations — work he later described as “carefully structured academic work.”
That continuity ended abruptly. After disabling ChatGPT’s “data consent” option — a setting intended to prevent conversations from being used to train models — Bucher found that his entire chat history had disappeared. Where there had once been an archive of exchanges, he encountered what he later described as “just a blank page,” with no undo option and no visible warning that the deletion would be permanent.
In a subsequent column for Nature, Bucher wrote that the loss amounted to two years of intellectual labor vanishing in an instant. He framed the episode not as a technical curiosity, but as a moment that exposed the fragility of academic workflows increasingly built atop commercial AI platforms.
OpenAI’s Response and the Question of Warnings
In response to the account, OpenAI disputed key elements of Bucher’s claim. In a statement to Nature, the company said that deleted chats “cannot be recovered,” but challenged the assertion that there had been no warning. According to OpenAI, users are shown a confirmation prompt before permanently deleting a chat.
The company also reiterated guidance that has appeared in various forms across its documentation: users should maintain personal backups for professional or mission-critical work. The recommendation, while practical, underscored a broader tension between how users perceive AI tools — as stable, workspace-like environments — and how the companies behind them position those tools, legally and technically.
Bucher acknowledged that ChatGPT can generate “seemingly confident but sometimes incorrect statements,” and said he never equated its fluency with factual accuracy. His reliance, he emphasized, was instead on the apparent stability and continuity of the workspace itself, particularly as a subscriber to ChatGPT Plus.
AI Slop and the Strain on Scientific Publishing
The episode unfolded against a backdrop of growing unease within academic publishing about generative AI’s broader impact. As The Atlantic reported, scientific journals are increasingly inundated with poorly sourced, AI-generated manuscripts — a phenomenon critics have labeled “AI slop.”
Entire fraudulent journals have emerged to capitalize on researchers seeking rapid publication, sometimes resulting in AI-generated papers being reviewed by AI tools themselves. The feedback loop, editors warn, risks polluting the scientific record and overwhelming peer review systems already under strain.
Compounding the problem, researchers have reported being cited in new papers only to discover that the referenced material does not exist at all, having been hallucinated by language models. While there is no evidence that Bucher attempted to publish AI-generated research or pass it off to students, his experience became entangled in a wider debate about trust, verification and accountability in AI-assisted scholarship.
Backlash, Sympathy, and a Cautionary Tale
Public reaction to Bucher’s column was swift and polarized. On social media, some commentators expressed schadenfreude, questioning how an academic could rely so heavily on a cloud-based tool without maintaining local backups. Others went further, calling on the university to discipline or even dismiss him for depending on AI in academic work.
Yet sympathy also surfaced. Writing on Bluesky, Roland Gromes, a teaching coordinator at Heidelberg University, praised Bucher for publicly acknowledging what he called “a deeply flawed workflow and a stupid mistake.” Many academics, Gromes noted, believe they can anticipate AI’s pitfalls, only to encounter them firsthand.
Bucher himself has framed the loss as a hard-earned lesson rather than an indictment of the technology. ChatGPT, he wrote, remains useful for drafting non-critical text and generating first passes that can be carefully revised. What it is not, he suggested, is an infrastructure on which years of irreplaceable work should quietly rest.
In that sense, the disappearance of his chats has become less about a single vanished archive and more about the evolving, uneasy relationship between scholars and the tools reshaping how knowledge is produced.
