The capabilities of large language models like ChatGPT have greatly impacted the way many people work, but a recent incident highlights the potential risks of using these tools.
Professor Marcel Bucher of the University of Cologne said he lost two years of his academic work as a result of a simple settings change: grant applications, teaching materials, and publication manuscripts suddenly disappeared.
Bucher had intended to disable the option to allow his data to be used to train the model, but according to him, this action deleted his entire chat history.
In an article published in Nature, he described his multiple attempts to recover the data and conversations, noting that reaching out to OpenAI yielded no results.
The professor explained that the data was permanently deleted and could not be recovered, while OpenAI referred to the principle of “privacy by design,” which ensures that the data is deleted without any trace.
“If one click can irreversibly delete years of work, then ChatGPT cannot be considered completely safe for professional use based on my experience,” Bucher concluded.
However, ChatGPT provides a backup feature, where the user can download all the conversations and data easily. The “Data Export” option is available under Settings in the “Data Control” section.
After some time, a download link to a ZIP file containing all the stored information is sent via email, and the link remains valid for 24 hours. It should be noted that backup is an essential part of working on computers, and should not be ignored when using AI tools.
Fortunately, modern self-testing has not been able to reproduce the same scenario. Previous conversations remain available after disabling data sharing for training purposes, while an explicit warning appears when choosing to delete all conversations, requiring user confirmation. Since August, OpenAI has likely made adjustments to the user interface and security mechanisms to prevent accidental deletion.
