Can AI be funny? It now passes the average-joke test, but still fails the great-joke test

THE THESIS. Whether AI can be funny now has a data-backed answer, and it has two parts. Yes: a language model writes jokes that blind audiences rate as highly as an average person's, or higher. No: when the bar is the great joke, the one that surprises, takes a risk and could only come from a specific person, the machine still lags. Our reading is that much of this failure is not a lack of intelligence. It is a design choice. We train models to be kind, predictable and safe, and humor thrives on the opposite. So humor is not ‘the last test of intelligence’, as is often said. It is an excellent test of something more specific: whether a machine can read its audience and choose to take a risk.

WHAT IT TAKES TO MAKE PEOPLE LAUGH. Philosophers have tried to explain laughter for centuries, and they offer four main theories. Superiority theory, linked to Hobbes, says we laugh when we feel above someone else. Freud's relief theory sees laughter as a valve that releases tension. Incongruity theory, from Kant and Schopenhauer, says we laugh at the clash between what we expected and what happens. The newest, the benign violation theory of psychologists Peter McGraw and Caleb Warren, combines them: something is funny when it breaks a norm in a way we perceive as harmless. They all point the same way. To be funny you need theory of mind, meaning the ability to imagine what your audience thinks and expects. You need timing. You need shared culture. And you need to dose the transgression: go a little too far, but not too far. These are exactly the skills that come hardest to a model that predicts the next word from text.

WHAT THE DATA SAYS. The evidence is more nuanced than the headlines. In 2023, Sophie Jentzsch and Kristian Kersting asked ChatGPT (GPT-3.5) for a joke 1,008 times. More than 90% of the answers were variations on just 25 jokes, according to their study. The authors concluded that the model had memorized a pattern rather than learned to be funny. A year later, Drew Gorenz and Norbert Schwarz of the University of Southern California published what looks like the opposite in PLOS ONE. In blind tests, ChatGPT 3.5's jokes beat those of most amateur human participants, around 70% according to coverage of the study. Its satirical headlines were rated as well as those from The Onion. That same year, Google DeepMind worked with 20 professional comedians, half of them at the Edinburgh Fringe. According to the study and MIT Technology Review's report, the comedians found the material ‘bland’, generic and unsurprising, though useful for building structure. In 2025, comedy writer Joe Toplyn and comedian-neuroscientist Ori Amir took jokes from Toplyn's Witscript system to a stage in North Hollywood. They measured the length and loudness of the laughs, and the audience laughed about as much as at human-written jokes. But a human expert wrote the prompts and picked the best from dozens of candidates. A study on memes sums up the picture well, according to Live Science. GPT-4o's memes were funnier on average than human ones, but the very best were still made by people. The New Yorker caption contest benchmark (Hessel and colleagues, ACL 2023) shows the other side. A benchmark is a standard test used to compare models. In this one, raters preferred human explanations of why a joke works over GPT-4's in most cases, according to the authors.

THE PROBLEM WITH BEING NICE. Why does the machine win on average but lose at the top? Part of the answer lies in how models are polished. After initial training, they are tuned on human preferences. This technique, called RLHF (reinforcement learning from human feedback), rewards the answers raters like most. The side effect is documented. A 2024 study on Llama 2 found that aligned models produce less diverse text. A 2025 paper on Verbalized Sampling points to the cause: human raters tend to reward what is typical and familiar. The authors call it ‘typicality bias’. The model learns to give the safest answer, and in comedy the safest answer is the joke you already know. There is also a deliberate layer. Labs train their models not to offend, and with good reason. But benign violation theory says no violation, no joke. A system designed never to break a norm has its humor cut off from the start. The good news is that this can be partly fixed. According to its authors, asking the model for several answers along with their probabilities raises creative diversity by 1.6 to 2.1 times, with no retraining.

DO WE LAUGH THE SAME IN EVERY LANGUAGE? No, and the data is starting to show it. Humor depends on language: puns, double meanings, local references, irony. In 2026, the SemEval academic competition ran its first general humor-generation task in English, Spanish and Chinese. Thirty-seven teams took part. According to the organizers, a simple Gemini 2.5 Flash baseline tied for first place in every subtask. More elaborate systems barely beat it. Several teams blamed their modest results on a lack of originality. Meanwhile, IberLEF's Spanish-only HAHA task now includes spotting whether a joke was written by a machine. And the JOKER lab at the CLEF conference concludes that models still struggle to create or translate wordplay across languages. Our reading: models are trained mostly on English text, so their default humor is Anglo-American. A joke from a Spanish comic like Gila or Chiquito, or from a long lunch at a Madrid bar, does not translate. It is lived. An October 2026 review puts it this way: models capture surface patterns of humor but fall short on pragmatic reasoning and cultural grounding.

OUR READING AND THE IMPLICATIONS. Three conclusions. First, as in the debate over whether AI understands, we should separate whether the machine makes people laugh, which it does, from whether it knows why. The second question cannot be settled today and matters less in practice. Second, there are real short-term effects on jobs. Filler jokes, witty ad copy and brand memes can already be produced by a machine at near-zero cost. Writers who live on volume will suffer. What holds up is what we have seen in other sectors: the judgment to pick the good joke out of a hundred, a personal voice, and a bond with a specific audience. One finding also favors humans. Since work by Bower and Steyvers in 2021, we know audiences rate a joke lower if they believe an AI wrote it. Laughter is also a bond between people. Third, the long view: a model that learns to read the person in front of it and take tactful risks will not just tell better jokes. It will be a better tutor, a better companion for older people and a better assistant in therapy, where well-judged humor heals. The goal is not a machine that replaces comedians. It is a tool that helps more people do what they love, including making others laugh. Until then, the most honest measure remains the old one: the awkward silence after the punchline.