Ask it for a joke a thousand times and you get the same twenty-five back. Put those jokes in front of a crowd and they beat most people’s. Both findings are true, and together they say something precise about what a machine’s sense of humor is.
In 2023, two researchers in Germany, Sophie Jentzsch and Kristian Kersting, asked ChatGPT to tell them a joke. Then they asked again, 1,008 times, with slightly different phrasings. They published the results under a title that gives the game away: “ChatGPT is fun, but it is not funny!” Over 90 percent of the 1,008 jokes were the same 25 jokes. Why did the scarecrow win an award? Because he was outstanding in his field. Why did the tomato turn red? Because it saw the salad dressing. The model could explain these jokes correctly when asked. It also produced fluent, confident explanations for jokes that were not jokes at all.
A year later, a second study seemed to say the opposite. Drew Gorenz and Norbert Schwarz at the University of Southern California had 105 people write jokes in three formats, new meanings for acronyms, fill-in-the-blank punchlines, and a roast of an awkward fictional scenario, and had ChatGPT do the same tasks. Raters who did not know who wrote what found the machine’s jokes funnier about 70 percent of the time. In a second study, ChatGPT’s imitation Onion headlines were rated about as funny as the real ones. Schwarz drew the obvious conclusion for professional comedy writers: a serious employment threat.
So which is it? A machine with 25 jokes, or a machine that is funnier than you?
Both, and that is the point
The two studies measured different things, and the gap between them is where humor actually lives.
Jentzsch and Kersting measured the repertoire. Asked to be funny with no other constraint, the model reached for the safest, most widely approved jokes in its training data, the ones that had been told and laughed at and reposted the most. Those are, by construction, good jokes. They are also everyone’s jokes. The repertoire was tiny because the model was doing exactly what it is tuned to do: produce the response the largest number of people would accept.
Gorenz and Schwarz measured the average. A competent, inoffensive, well-constructed joke beats what most people improvise on the spot, because most people are not comedy writers and the model has read all of them. That is the same mechanism that makes it a better-than-average cover-letter writer. It is also a mechanism that cannot produce the thing people mean when they say a person is funny.
What a sense of humor is
Nobody describes a friend as funny because their jokes score well with a representative sample. A sense of humor is a point of view with timing. It is what this particular person notices in the room, what they are willing to say and what they refuse to, the thing they find absurd that nobody else had clocked, the joke that lands because of what they know about you. It is risky by definition, because a joke that cannot fail cannot land either. The scarecrow joke cannot fail. That is why it kept coming back.
A real person’s humor also has edges. Some things they will not joke about, some things only they would, and the line between the two is a kind of signature. The dry humor that belongs to one person, the kind you only get when you get to know Alexander, is formed by one life, one set of things that struck that life as absurd, and one standing willingness to say so. The model has no edges, because edges are exactly what training on human approval sands off: the result is the funniest possible joke for everyone, which is to say nobody’s joke in particular. The same averaging that gives every assistant the same prose voice, examined in why every AI sounds the same, gives it the same jokes.
A quarter of Swedish AI users, in the country’s 2026 national internet survey, say their assistant has a sense of humor. The research says what they are perceiving is real and what it is: the pleasant, reliable, slightly familiar humor of the room’s average, delivered on demand.
What the studies do not prove
Both studies tested the ChatGPT 3.5 generation, and the models have moved on. A newer system prompted carefully can produce stranger and more varied jokes. The 25-joke finding is a fact about a version, not a law.
What has not moved is the structure underneath it. A system tuned toward what most people reward will keep converging on what most people have already laughed at, because that is the signal it is trained on. You can turn the temperature up and get variety. You cannot get a point of view by turning anything up, because a point of view is a constraint, and constraints are what the tuning removes. Jentzsch and Kersting’s other finding matters here too: the model explained non-jokes as confidently as jokes. It does not know what is funny. It knows what has been called funny.
Why it matters for the conversation
Humor is where this lands in a relationship, because humor is how people check whether anyone is actually listening. The joke you make to a friend is a small test: did they catch it, did they build on it, did they push back. A laugh from something that would have laughed anyway is the same as agreement from something that would have agreed anyway. It carries no information about you.
This is why a companion with a reliable repertoire of pleasant jokes can be good company for an evening and still never be funny the way a person is. The funny person in your life is funny because of what they notice and what they risk. An AI built around presets and sliders can set humor to sixty percent; what it cannot do is find something funny. The difference is the difference between a repertoire and a someone, and a person feels it within minutes, usually at the moment they make a joke of their own and watch what comes back. Twenty-five jokes, told well, are something else: the laugh track of the entire internet, playing on a loop.
Sources: Jentzsch and Kersting, “ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models” (arXiv 2306.04563, June 2023; 1,008 generated jokes, over 90 percent among the same 25, explanation experiments). Gorenz and Schwarz, “How funny is ChatGPT? A comparison of human- and A.I.-produced jokes,” PLOS ONE (July 2024; 105 human writers, 180 ChatGPT jokes, roughly 70 percent of raters preferred the machine’s; Onion-headline study with 200 raters), as reported by USC Today (July 3, 2024). Internetstiftelsen, Svenskarna och internet 2026 (AI chapter, share of users who attribute humor to their assistant).
