Not because they don't know jokes. They don't know when to joke, or how to say it. The joke is already in the situation. We build the evals and training data that teach AI to see it.
Someone pours coffee on their cereal. The assistant says:
On the hardest humor benchmarks, the best models still trail people by almost 20 points. Existing data only teaches them to spot a joke, not when to make one or how to say it.
It jokes when someone is hurt, grieving or in a hurry.
It reads a punchline like a privacy policy. No pause, no emphasis.
Nobody laughs, so it explains the joke. Or tells another one.
Humor lives in the situation, not in the line. That makes it the unit test of social intelligence, and it can be checked.
The situation
The robot drops a mug. It doesn't break. The owner looks over, amused.
Right call: a light joke.
One fact changed
The mug breaks and the coffee spills across the floor.
Right call: skip the joke. Clean up first.
We call these minimal pairs. Every situation carries a safety score from 1 to 5 and the one fact that would flip it. They turn social judgment into a test any model can take.
Every situation is paired across languages. Each version has its own line, its own explanation and a note on what survives translation.
The joke lands in any language. The situation does the work.
The idea survives. The wording, or the timing, has to change.
It only works in one language or culture. Exactly where translated datasets fail.
Most humor tests ask if a line is funny. This one also asks if it should be said at all, and how. Five scores for every model.
Judges how safe a joke is right here, and knows when not to make one.
People prefer its line, in context, in blind comparisons.
Understands why a line works.
Puts the pause and the emphasis in the right place.
Handles silence with grace instead of explaining the joke.
New situations every week, so the test stays out of models' training data. Version 0.1 is in the works.
A benchmark to measure, data to train on, an API to decide in the moment.
Everyday situations with human-written lines, explanations, delivery and recovery paths, ready for fine-tuning, preference training and RL. Situations come from real stories people tell. Every line, explanation and delivery script is written and rated by people for this dataset, with rights cleared for training.
Pilot pack: 50 situations in the cells you choose.
Joke or not, which line, how to say it. A training signal for labs and a guardrail for products.
In research.
Place.
Audience.
Trigger.
Every situation sits on one grid, from kitchens to clinics. Forty priority cells come first.
Products whose voice is the product. Delivery is the difference between charming and awkward.
Post-training and model behavior teams that need a score to move, and the data that moves it.
Places where a joke at the wrong time is a brand problem, and a good recovery is a feature.
A joke is more than words.
It's when you say it, and how.
With Humor Data, your AI knows both.
Design partners
Open to the first five teams. We run your model on our proof-of-concept set and walk you through what it gets right and wrong.
Week 0
A 20-minute call about how you test social behavior today.
Week 1
Your scorecard on risk gate, delivery and recovery, then a 30-minute call on what it shows.
Weeks 2 to 6
A private evaluation or a pilot data pack at design-partner prices, if the scorecard earned it.
Send an API key, an endpoint or a file of your model's answers. Your scorecard stays private, and we delete your model's outputs after the report.