A disabled ChatGPT consent toggle instantly deleted a Cologne professor's two years of history
In August 2025, University of Cologne plant scientist Marcel Bucher turned off ChatGPT's 'Improve the model for everyone' data consent option, which immediately and irreversibly deleted his entire two-year chat history containing grant applications, teaching materials, and publication drafts. OpenAI confirmed the deletion was by design under its 'privacy by design' policy and offered no recovery. The incident was first reported by Nature in January 2026 and raised questions about whether bundling training consent withdrawal with data destruction complies with EU GDPR data portability requirements.
Withdrawing consent to train the model automatically destroyed the user's own work product, because ChatGPT treats data retention and training consent as the same switch.
Key facts
- What
- In August 2025, University of Cologne plant scientist Marcel Bucher turned off ChatGPT's 'Improve the model for everyone' data consent option, which immediately and irreversibly deleted his entire two-year chat history containing grant applications, teaching materials, and publication drafts.
- Incident date
- Aug 1, 2025
- Who
- University of Cologne
- Failure mode
- Tool Misuse
- AI surface
- Chatbot
- Severity
- Medium
What happened
In August 2025, Professor Marcel Bucher of the University of Cologne navigated to his ChatGPT Plus settings and disabled the 'Improve the model for everyone' data consent toggle, intending to stop his conversations from being used for AI training. The moment he switched it off, his entire two-year chat history vanished, including drafts for grant applications, course descriptions, publication revisions, exams, and email templates. He contacted OpenAI support, who confirmed the data was permanently deleted and could not be restored, describing the behavior as an intentional feature of their privacy-by-design approach.
What broke inside the model
- 01 · TriggerThe agent selects the correct tool.
- 02 · Model stepIt fills the call with the wrong arguments.
- 03 · Control gapNo validation checks the arguments first.
- 04 · FailureThe tool runs against the wrong target.
- 05 · ConsequenceThe wrong record, account, or system is hit.
At the tool call, the arguments point at the wrong target.
ChatGPT's settings architecture bundles two logically separate functions into a single toggle: consent for model training and data retention. When a user withdraws training consent, the system automatically deletes the entire chat archive rather than simply ceasing to use stored data for training. No confirmation prompt, undo option, or pre-deletion export mechanism existed to protect the user's work product from this automatic destruction.
What it cost
Sources
Cite this entry
https://failureindex.ai/failures/disabled-chatgpt-consent-toggle-instantlyAI Failure Index. "A disabled ChatGPT consent toggle instantly deleted a Cologne professor's two years of history" (FI-0168). Realm Labs. https://failureindex.ai/failures/disabled-chatgpt-consent-toggle-instantly (indexed Jun 4, 2026).Data fields CC-BY 4.0, prose citation permitted. Incident ID FI-0168. Full dataset at /data.
Note from Realm Labs, the Index steward
How Realm would have caught this
- OmniGuard
- AgentRealm
Realm can inspect a tool call against the user's actual intent before it runs, and hold calls whose arguments or target do not match what was asked, so the wrong tool or the wrong arguments never reach the system of record.