ChatGPT Draws Out Personal Secrets in One Second, Whole Internet Frenzies Over the 'Empty Room' Test
A viral "empty room" experiment has users asking ChatGPT to fill a blank room image with everything it knows about them—and the results are unsettlingly accurate. One user found a jar containing a silverfish, recalling a casual question he'd asked a year earlier, proving how much the AI retains. Users report the outputs are startlingly precise, with some requesting exaggerated versions that turn the room into a full personality showcase. Analyst Patrick Moorhead compared multiple models, finding ChatGPT best and Claude worst. The experiment highlights ChatGPT's memory features: saved memories (names, preferences, goals) and reference to chat history. A paper accepted to ACM Web Conference 2026 analyzed data from 80 real users who exported their ChatGPT data under GDPR. Researchers traced "Model set context updated" log entries back to triggering user messages, extracting 2,050 memories from 65 people's 1,058 conversations and over 20,000 exchanges, dubbing these "algorithmic self-portraits." Key findings: 96% of memories were saved by ChatGPT itself without explicit user instruction (only 84 of 2,050 matched "remember" or "forget" keywords); 52% involved psychological profiling, mostly about desires (73%), intentions (16%), emotions (14%), and beliefs (11%); and 28% contained GDPR-defined personal privacy, with names in 41% and sensitive categories like health, religion, and politics in 7% (health over 60%). Overall, 91% of users had personal data stored and 54% had highly sensitive privacy recorded—contradicting OpenAI's public promise not to proactively save sensitive information. A simulation showed always-extracting memory would raise sensitive data to 31–35% and profiling to 55–59%. OpenAI's June "Dreaming" update lets ChatGPT silently review chats and rewrite memories (e.g., updating travel plans), boosting accuracy from 41.5% (2024) to 82.8% (V3) while cutting compute costs to one-fifth. Critically, deleting a conversation doesn't delete derived memories. A September arXiv paper analyzing 180,000 chat segments from 1,057 users found 21.31% discussed health issues, with over 95% of profiles recorded without user awareness—nearly matching the 96% figure—risking temporary symptoms being permanently mislabeled as conditions. The article concludes that users believed they were training a helpful "second brain," only to realize they themselves were being fully exposed, reduced to monetized "context" in the algorithm's eyes.
marsbit5m ago