Observed arrival · 2026-11-06
Humor Data tests whether AI knows when not to joke
A situational-humor benchmark and training-data project for evaluating when an AI should joke, what it should say, and how it should deliver it.
- For
- AI teams evaluating conversational tone and humor
- Worth noticing
- Each situation is paired with a safety score from 1 to 5 and the single changed fact that would flip the judgment.
Field notes
The site frames situations, rather than joke lines alone, as the unit of evaluation. Its example changes a mug from unbroken to broken, shifting the suggested response from a light joke to cleaning up first. The page describes human-written and rated lines, explanations, delivery scripts, and recovery paths; it also says training rights are cleared. Version 0.1 is described as in progress, and the pilot pack is listed as 50 situations in buyer-selected cells.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue