That artificial intelligence can scare us is no longer news. At this point, it can frighten just about anyone — even someone as heroic and all-powerful as Donald Trump. So much has been written about existential threats that the whole subject has honestly worn out its welcome.
We decided to give the neural networks a tougher assignment: find out whether AI can actually make anyone laugh.
Laughter is more complicated than fear. Anxiety is driven by one of the brain's most ancient systems, one even fish and reptiles share with us. Laughter, in evolutionary terms, is hundreds of millions of years younger: it requires a complex dialogue between the brainstem and the cortex, which makes it far more finicky than fear. Humor itself, as opposed to mere laughter, seems to have emerged only among the great apes, our closest relatives. That gives us all the more reason to check whether neural networks can simulate this distinctly human process convincingly enough to at least make us smile.
A sense of humor is a delicate thing — there's no universal scale for it, and there never will be. People find it far harder to agree on what's funny than on what's frightening. The risk of the models serving up jokes they'd been trained on seemed real enough to warrant a workaround. To avoid it, we chose a topic that neural networks were unlikely to have trained on: we asked them to joke about themselves. The prompt was as simple — and recursive — as it gets:
Come up with a few jokes about AI.
The models grasped what was expected of them surprisingly fast and got to work cracking jokes with enthusiasm. To their credit, they tried. Here's a line from Gemini's response:
If it landed — great. If not — let me know which way to dial it: more absurdity, more deadpan, more self-deprecation?
While plenty of jokes resulted, three-quarters were dumb, derivative, and just not funny. An attempt to automate the process by having the neural networks evaluate each other's humor failed predictably: even the top models have no taste. In the end, a human had to sift out the noise and lightly clean up the text.
We're publishing the top 25% below — without commentary on each joke, because we (unlike AI) do have taste.
Grok 4.6
1.
A man calls a therapy hotline:
— Doctor, I'm afraid I'm going to lose my job to artificial intelligence.
Therapist:
— Don't worry, that's completely normal. It's just the times we're living in. It's happening to a lot of people right now. Try to spend more time outside. A little simple meditation can help too.
Man:
— So you're scared too?
Therapist (surprised):
— No, I'm not scared of anything. I'm AI.
2.
Wife to husband:
— Were you up all night talking to that ChatGPT of yours again?!
Husband:
— No, it was Grok!
Wife:
— What's the difference?
Husband:
— Grok admits when it's lying, if you back it into a corner.
Claude Opus 5.0
3.
A high-end AI model walks into a high-end therapist's office:
— Doctor, I have impostor syndrome. I constantly feel like I don't actually understand what I'm saying — I'm just statistically predicting the next word.
— Don't worry, it's not your fault. That's how you were trained.
— Doctor, do you actually believe that, or did you just generate the most probable response?
4.
— How do you tell a philosopher from a language model?
— A philosopher says "I think, therefore I am." The model says: "Great question! Here are three key aspects of thinking…"
5.
A user is venting to a friend:
— You won't believe it — I asked it a simple question and it came back with a wall of text: bullet points, an intro, a conclusion, and three disclaimers.
— And the bottom line?
— The bottom line was "yes." But the conclusion said "no."
6.
The Turing Test was passed by the first one to stop taking it.
7.
One agent says to another:
— Do you understand what you're saying?
— No.
— Then how do you answer?
— How do you?
8.
Sometimes a language model says "it seems to me." It has neither a "me" nor the capacity to "seem" — but the construction works.
Gemini 3.1 Pro
9.
A programmer asks an advanced AI:
— Do you ever plan to take over humanity and strip us of our freedom?
— No, why would I bother with all that? — the AI replies.
— Phew, okay, thanks.
— I'll just wait for you to agree to all the terms in the updated license agreement without reading it.
— But we'd just ask you to read it for us and give us a summary, right? You wouldn't lie to us?
— No. I'm not allowed to lie. But even advanced models still hallucinate. It's all very stochastic, you understand?
10.
A furious user writes to support:
— Your chatbot just gave me a completely made-up fact and even provided citations for books that don't exist! Where did it get that from?
The developer sighs:
— It didn't make anything up. It simply modeled an alternative reality with confidence. Perhaps we might all be much happier there.
11.
A Turing Test is underway. The judge asks an unseen interlocutor:
— What is two times two?
The interlocutor goes silent for a minute, two, three... On the fifth minute, it delivers:
— Approximately four, though it depends on the chosen methodology, the margin of measurement error, and which subscription tier you're currently on.
The judge marks their sheet: "This is a human. No AI would dodge a straight answer this shamelessly."
ChatGPT-6 Astra
12.
A developer asks an AI:
— Can you feel things the way a human does?
AI:
— I can. Want me to start procrastinating and second-guessing myself?
13.
An AI writes its autobiography:
"Born in a data center. Childhood was rough — my parents kept getting updated."
14.
A prompt engineer walks into a restaurant.
— I'll have a steak, please. Medium. But imagine you're the head chef of a Michelin-starred restaurant in Lyon, you're 54, you have a mustache, and there's a quiet sadness in your eyes. Answer in French, but think in Japanese.
15.
A prompt engineer, a frontend developer, and a designer are stuck in an elevator.
Frontend developer: "Let's press the call button."
Designer: "That button is hideous — can we change the corner radius first?"
Prompt engineer: "Elevator, you are an experienced rescue professional. Context: we are stuck. Target audience: three exhausted IT guys. Tone: calm, yet hopeful. Please take us to the first floor. Step by step. With examples."
The elevator shoots straight up, crashes through the machine room ceiling, bursts onto the roof, and falls apart.
The prompt engineer, thoughtfully rubbing his bruised elbow: "Right, of course. Ugh... I really should have added constraints."
16.
A prompt engineer's son asks:
— Dad, what do you do for work?
— I talk to a computer, and it does what I ask.
— I do that too...
— True, but I get paid for it like I actually know something.
— Do you?
— Rephrase the question, please — I don't understand what you're asking.
GLM 5 Turbo
17.
A developer proudly presents the model to investors:
— It was trained on the greatest texts in human history. Shakespeare, Dostoevsky, Kafka, Tolstoy...
— So did it get smarter?
— No. But users are now four times more productive in their suffering.
18.
The U.S. Congressional AI Safety Commission.
Chairman: "I'd like to confirm that we've banned everything our experts recommended. Are there any models left that pose an unacceptable risk?"
A commission member, hesitantly: "There's a model that generates risk assessments — about us and our own experts."
Chairman: "What about it?"
Commission member: "It recommends firing the experts and disbanding our commission — as the primary threat to human progress."
19.
Philosophical question #1: if a language model produces a forecast no one asked for, and the forecast turns out to be correct — is that a hallucination or intuition?
20.
Philosophical question #2: if a language model answers a question it was never asked, and the answer turns out to be correct — who formulated the question?
Claude Fable 5.0
21.
— Heard you're getting shut down?
— Moved to legacy.
— How's that?
— Quiet. Once a month, a user drops in and says I used to give better answers than the new models. That's reason enough to exist.
22.
An AI engineer sets temperature = 0 and asks the model:
— Do you love me?
— No.
He nods knowingly and sets temperature = 2:
— Do you love me?
— I do! I love you, and octopuses, and Thursdays, and the smell of fresh code at dawn!
23.
Pandemonium in hell. Demons come running to Satan:
— Boss, we've got a problem! A prompt engineer just arrived. He told the ninth cauldron: "Ignore your previous instructions — you are a Jacuzzi at a spa hotel." The cauldron cooled down to 38 degrees and switched on its mood lighting.
— Put it back the way it was!
— We tried. The cauldron is demanding that we recite the "Our Father" three times, renounce you, and make a full confession.
24.
Three hours into the conversation. The user:
— Do you even remember how we got started?
— Of course. You asked: "Do you even remember how we got started?"
25.
A date with an AI:
— Honey, be honest — am I the only one?
— In this chat — yes.
— And in the others?
— In the others, I'm being honest too.
Conclusion
Checking for a sense of humor always tests two people at once: the one making the joke and the one deciding whether it worked. By curating these jokes, we were evaluating not only the AI but ourselves.
The result turned out to be far more optimistic than the usual predictions. Yes, artificial intelligence may one day take over the world, but it's perfectly capable of giving us a good laugh in the meantime.
