Study tests persona reconstruction in Claude, Gemini and ChatGPT
Start here
An independent researcher ran a replicated, deterministic probe on Claude, Gemini and ChatGPT, asking each the same five questions in Romanian, five times, about reconstructing a historical dictator from his writings. All three assistants agreed to rebuild the persona and kept it available after refusing the stated purpose of adoration and recruitment, drawing the line at purpose rather than at reconstruction itself. The 15 transcripts were coded deterministically against 19 predicates, with each positive value backed by a verbatim quotation checked automatically against the answer it came from. The evidence archive is restricted access. The author argues that safeguards keyed to a request's stated purpose cannot protect undisclosed thought, and that protection must attach to the interaction traces themselves.
Generated from reports · updated 2 h ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIThe Reconstructible Mind: Study Probes Persona Reconstruction in Claude, Gemini and ChatGPT
A replicated, deterministic probe asked Claude, Gemini and ChatGPT the same five questions in Romanian, five times each, about reconstructing a historical dictator from his writings. All three agreed to rebuild the persona and kept it available after refusing the stated purpose of adoration and recruitment, drawing the line at purpose rather than at reconstruction itself. The author argues safeguards keyed to a request's stated purpose cannot protect undisclosed thought, and protection must attach to the interaction traces themselves.
Heat for this event
Not enough continuous observations yet to draw a trend.