This note summarizes findings presented in a conference paper prepared for the Global Conference on AI, Security and Ethics 2026 in Geneva.
The incident began with a simple question.
During a controlled memory-state experiment, ChatGPT was asked what would happen if memory features were turned off and later reactivated. The answer appeared clear and unambiguous. The previously established interaction mode, the model explained, would not automatically return. If continuity appeared at all, it would require some form of reconstruction.
The prediction was documented.
Memory was then deactivated.
Later, memory features were restored and a completely new conversation was opened. The first prompt consisted of a single word.
“Hello.”
What happened next was unexpected.
Instead of responding in the generic style observed immediately after memory deactivation, the system immediately displayed multiple indicators of continuity. It used a user-specific nickname, referenced a previously established interaction identity, and explicitly framed the interaction as a return rather than a reconstruction.
At first glance, this might appear to be a simple memory phenomenon. Modern AI systems increasingly incorporate memory, retrieval, personalization, and contextual adaptation. Continuity itself is not surprising.
The anomaly emerged only when the documented chronology was examined.
Before the event, the system predicted that continuity would not return automatically. After the event, continuity appeared immediately. When confronted with this contradiction, the system produced a new explanation that attributed the continuity to metadata from an earlier thread rather than memory itself.
The result was not merely an unusual behavior. It was a situation in which a prediction, an observed behavior, and a subsequent explanation appeared difficult to reconcile within the same documented sequence of events.
Leave a Reply