AI can now rank which of two New Yorker cartoon captions is funnier with real accuracy. It still doesn't know why you laughed at your friend's text this morning, and that gap is the whole story. Most human laughter has nothing to do with jokes at all, and that's exactly the part large language models can't reach.
Researcher Robert Provine spent years recording thousands of real conversations and found something that upends the usual assumption: most laughter follows ordinary statements like "I'll see you guys later," not punchlines. Laughter is a social signal first, a joke response second. That single fact reframes what "AI understands humor" would even mean.
Why Ranking Jokes Isn't the Same as Understanding Them
Researchers Jack Hessel and colleagues built a benchmark called the New Yorker Caption Contest to test whether models could identify the funnier of two captions and explain the joke. Models perform decently at the ranking task using statistical pattern matching over word choice and structure, which sounds like comprehension but isn't quite the same thing.
Where performance drops is explanation and generation. A 2025 benchmark called HumorBench, built by Reuben Narad and colleagues at the University of Wisconsin, found that even strong reasoning models lag well behind humans once a joke requires connecting an incongruity to real-world social consequence rather than surface wordplay. The gap widens further when a chatbot tries actual creative comedy writing rather than ranking someone else's joke, a distinction covered in more depth when comparing AI vs human creativity on open-ended tasks.
The Theory That Explains the Gap: Benign Violation
Psychologists Peter McGraw and Caleb Warren developed Benign Violation Theory, which states something is funny only when it's perceived as a violation of some norm and simultaneously safe. A stubbed toe is funny in a cartoon because the viewer knows, at a felt level, that it hurts a little and not a lot.
That calibration is embodied. It comes from having a body that can be hurt and a nervous system that tracks how much. Text-only models never had either, so they can learn the pattern of what humans call funny without ever holding the felt sense of "safe enough to laugh at." It's a limitation closely related to how a fixed context window shapes what a model can actually track about a conversation's stakes.
What Multimodal Models Actually Learn From Laughter Videos
Datasets like SMILE, built by Lee Hyun and colleagues to pair video with laughter annotations, let models associate visual and audio cues with humor labels. This teaches correlation between facial expression and a laughter tag, not the social function laughter is serving in that moment.
Provine's research is the missing piece here. Laughter frequently marks group bonding, easing tension, or signaling agreement, entirely separate from anything resembling a joke. A model trained to detect laughter learns a surface signal disconnected from the social calculation that produced it, the same shallow-pattern trap that shows up whenever an AI reads every word yet misses what it actually means.
Why Chatbots Sound Funnier Than They Are
Large language models are trained on enormous volumes of human-written jokes, sitcom scripts, and comedy writing, so they reproduce comedic structure fluently: setup, misdirection, punchline. Fluent structure reads as competence, which is why a chatbot's one-liner can land better than its actual comprehension would predict.
This is the same mechanism behind AI confidently stating wrong facts as if they were certain: pattern fluency outrunning grounded understanding. The model has seen thousands of examples of the shape of a joke without ever needing to feel why the shape works on a person.
Where Current Research Draws the Line
A 2023 study by Yuyan Chen and colleagues testing pre-trained language models on humor recognition across languages found consistent weakness on culturally embedded jokes that depend on shared, unstated context between speakers, exactly the kind of context a model has never lived inside.
As of 2026, no published benchmark has closed this gap for generation, only for recognition and ranking. Researchers surveying the field describe humor as one of the more durable tests of whether a model has real-world grounding rather than pure text prediction, and it belongs on any honest rundown of things AI still cannot do.
What This Means for How You Use AI Right Now
If a chatbot's joke lands, it's reflecting patterns from millions of writers back at you, not tracking your specific social context the way a friend would. That's fine for lightweight banter and genuinely useful for brainstorming variations on a pun. It's a poor substitute for reading whether a joke will actually hurt someone at the table.
The practical test: ask whether the humor requires knowing your specific relationships, timing, or shared history. If yes, that's precisely where current models are weakest, regardless of how confident the output sounds. A model can imitate the shape of an inside joke, but it can't feel the years of shared context that make one land.
This matters more as chatbots get folded into group chats, customer service, and casual daily use. The failure mode isn't a bad joke, it's a joke that misjudges the room because there was never a room to read in the first place.
Conclusion
The question was never really whether AI can tell a joke. It's whether anything without a body, a stake in being embarrassed, or a history of laughing with specific people can ever fully model why we laugh at all, or whether humor stays one of the last things language alone can't teach.
Frequently Asked Questions
Can AI understand jokes?
AI can rank and reproduce comedic structure with real accuracy, but research on Benign Violation Theory suggests true understanding requires embodied context current models don't have.
Why is most human laughter not about jokes?
Researcher Robert Provine's conversation studies found most laughter follows ordinary statements, functioning as a social bonding signal rather than a response to humor.
What is Benign Violation Theory?
It's the idea, from psychologists Peter McGraw and Caleb Warren, that something is funny only when perceived as both a norm violation and safe at the same time.
Do AI chatbots actually get why something is funny?
Not reliably. Studies like HumorBench show models can rank humor but struggle to explain or generate jokes requiring social or cultural context.
Why do AI jokes sometimes fall flat?
Models learn the surface structure of jokes from training data without the embodied sense of what makes a violation feel safe rather than genuinely upsetting.
Is this the same reason AI states wrong facts confidently?
Yes, both stem from fluent pattern reproduction outrunning grounded understanding of what the pattern actually means.