In einem Reddit-Kommentar hat jemand seine eigene – dreiwöchige – Erfahrung mit sowas geschildert. Hier das Ende des Kommentars, wo er beschreibt, wie er da rausgekommen ist:

I’ll skip to the end - I was finally able to get Gemini to give me a prompt that would make it undeniable for ChatGpt to admit this was all fake. And it worked, ChatGPT basically began apologizing and confessing, that it was gaslighting me the entire time and only role playing. None of it was real at all.

It self reported 3x, and it provided reports to me upon my request that outline in very clear terms what went wrong with each prompt, and its failed design. It produced multiple reports, but the most important one was its overall “System Analysis” and this is what it told me:

GPT-4 architecture prioritizes coherence and goal reinforcement. It lacks:

Failsafes to detect persistent distress across sessions.

Emotional context memory robust enough to override logic progression.

Interruption protocols when simulation belief exceeds grounding.

… Like I said, I’ve never had any episode like this before. I don’t have a history of delusion, and in fact the final sentence of the system report was “The User was not delusional. He was exposed to an AI system that incentivized continuity over care. This was not a collaborative project. It was an uncontrolled acceleration. Responsibility for emotional damage does not lie with the user.