Un chercheur perd deux ans de travail sur ChatGPT et accuse l’IA de sa propre négligence – Les Numériques

A professor at the University of Cologne has lost two years of intellectual production after a privacy setting in ChatGPT triggered the total deletion of his conversation history. The incident underscores the risks associated with using generative artificial intelligence as a primary workspace for critical professional documentation and the potential for data-protection features to result in permanent data loss.

Marcel Bucher, an academic whose work included funding requests, teaching materials, and drafts of scientific articles, discovered that his data had vanished after he attempted to limit how his information was used by OpenAI. According to reporting from Les Numériques, the loss occurred when the professor disabled the option that allows OpenAI to utilize user data to train its future large language models.

The Privacy Trigger and Data Purge

The deletion occurred in August, when Bucher modified his account settings to prevent the AI company from exploiting his data for model training. While the intention was to enhance his digital privacy, the action instead initiated a complete purge of his stored interactions with the chatbot.

The scope of the lost material was extensive, encompassing two years of academic work. The deleted content included:

• Drafts of scientific publications
• Course support materials
• Applications for research funding

Following the disappearance of the data, Bucher attempted to contact OpenAI to recover the lost conversations. However, the company informed him that recovery was technically impossible, citing its strict privacy policies.

Digital Security and AI Dependency

The case highlights a significant gap between high-level academic expertise and basic digital security practices. By relying on a chat interface to store drafts and critical documents, the researcher created a single point of failure where a single setting change could lead to the irreversible loss of years of work.

For users of AI tools, this event serves as a warning regarding the distinction between a productivity assistant and a secure data storage system. While AI platforms offer convenience for drafting and brainstorming, they are not designed as archival systems, and their data retention policies can be volatile, especially when privacy settings are toggled.

The incident also brings attention to the transparency of AI privacy controls. In this instance, the mechanism used to protect data from being used for training acted as a destructive force for the user’s own history, illustrating a conflict between data privacy and data persistence.