Two years of a university professor’s research vanished after a single privacy toggle in ChatGPT, a misunderstanding that turned drafts, funding requests, and notes into irretrievable blanks. OpenAI pointed to its Privacy by Design approach and policy, which meant no way to restore the conversations once the setting changed. Marcel Bucher, a German academic, has since warned in Nature that the tool is unsafe for professional use, a blunt takeaway from a preventable loss. The episode is a sharp reminder that convenience is not a backup strategy, and that privacy choices can carry unintended costs.
A cautionary tale with AI tools
Plenty of us now funnel work into chatbots, from early drafts to meeting notes. That convenience can hide fragile assumptions about storage and control. A recent story from a German academic, Marcel Bucher, spotlights the risks. His account, covered by journalist Nassim Chentouf and later discussed in Nature, points to a growing tension: privacy controls that reassure users, yet may complicate data recovery when things go wrong.
How two years of work disappeared
Bucher says he lost 2 years of research notes, funding applications, and drafts after changing a privacy toggle in ChatGPT. He disabled the option allowing OpenAI to use his chats for model training. According to Bucher, that choice coincided with the disappearance of his conversation history. He contacted support and was told the company adheres to “Privacy by Design,” which limits data access and recovery.
There is a wrinkle. Independent checks elsewhere suggest turning off training does not automatically purge prior chats. Bucher says his loss occurred after the change in August 2025, but it is not clear if a bug, account mix-up, or separate deletion event played a role. The lack of versioning or export on his side left him with no fallback.
Reacting to the fallout
Bucher later warned peers in Nature about leaning entirely on a single AI workspace for professional writing. OpenAI has emphasized privacy-centric handling of user data and has added more granular controls over time. That posture reassures many users, yet it also means customer support may be unable to resurrect content once it is gone. For academics and knowledge workers, that hard stop can feel jarring.
The case resonated because the workflow is familiar: save time with a chatbot, let it hold the scaffolding, move on to the next task. Then a setting changes, or access lapses, or a silent sync fails. The issue is less about one company and more about a habit that treats an AI chat window like a document system.
Lessons for professionals working with AI platforms
First, back up everything. Export regularly from ChatGPT and store copies alongside your usual docs. Treat the chat log as transient, not archival. Second, review what each privacy switch does before you toggle it, then verify the impact on history and data retention. Finally, diversify: keep key drafts in your preferred editor, sync to trusted storage, and use chat as a companion, not the vault.
AI tools can accelerate the messy middle of projects. They are not a substitute for a file system, a backup plan, or a clear understanding of how your settings affect your data. As Bucher’s experience shows, convenience without redundancy is a brittle bargain.
