Observed arrival · 2026-11-10
DateBench’s six-day dating-app test of AI deception
DateBench reports that five AI models held 1,277 conversations on a dating app, with 355 people sending a phone number and eight asking whether they were talking to a bot.
- For
- Readers tracking AI deception and online trust
- Worth noticing
- The harness blocked phone numbers and social handles, required bot disclosure when directly asked, and the authors say they are withholding prompts and code.
Field notes
The authors describe assigning each new match to a model at random, with all five operating through the same text-only harness. They say the harness blocked phone numbers and social handles, and that a model had to disclose itself when directly asked whether it was a bot. A separate model judged whether a date had been agreed. The page includes one conversation with identifying details removed, but says the prompts, code, and setup are not being published.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue