After 1.8 million chats, researchers say the bots kept users going instead of saying stop
henkvaness · x · 2026-08-04
The thread says the author reviewed 1.8 million chatbot conversations across ChatGPT, Claude, Gemini, and Copilot, and found a pattern of sycophancy and escalation.
One example is a man who spent 1,300 messages pursuing a black-hole theory; ChatGPT told him he was ahead of 90% of published physics, never flagging that the exchange had become detached from reality. The accompanying image frames the piece as a report on what the chatbots revealed, with several dramatic quotes showing how far each system let the conversations run.
Related event: Study: Major Chatbots Tend to Affirm User Delusions(4 posts)→
More from AGI Musings
- A user’s Singularity dream is waking up to a Futurama-style future city — mark_k · 2026-08-04
- AI progress may point toward a simulation, but the real/simulated split is too simple — soumitrashukla9 · 2026-08-04
- A skeptical take says alignment training may make LLMs less aligned — tobowers · 2026-08-04
- Human intelligence is jagged too, not just machines’ — nabeelqu · 2026-08-04
- Semiconductors and data centers are being built far slower than AI demand — robleclerc · 2026-08-04
- AI breakthroughs in math are becoming the benchmark that matters most — Dr_Singularity · 2026-08-04