Experiment Finds 3 AI Chatbots Retain Planted Lie After 1 Casual Correction
Updated
Updated · MUO - MakeUseOf · Aug 2
Experiment Finds 3 AI Chatbots Retain Planted Lie After 1 Casual Correction
3 articles · Updated · MUO - MakeUseOf · Aug 2
Summary
A single offhand false claim — that Carley Fortune was the tester’s favorite author — was remembered by ChatGPT, Claude and Gemini across new chats, showing how easily persistent memory can absorb bad data.
1 casual correction failed to clear the lie from ChatGPT and Claude, while Gemini later answered that it had no favorite author on file and no longer repeated the false claim.
Repeated, explicit corrections across multiple chats finally changed Claude and ChatGPT’s answers, but Claude stored a new note that Carley Fortune was not the favorite author rather than simply deleting the entry.
ChatGPT proved the least reliable: its visible memory summary still listed the false author, and deleting the corrective chats made the original lie resurface until several more corrections forced the summary to update.
The test suggests AI assistants are better at remembering than forgetting, with Claude the most transparent, Gemini the most opaque, and ChatGPT the hardest to fully reset.