Which AI Is Actually Funny? We Tested Four
By Chatday Editorial Team ·
Three out of four AIs, asked to caption a photo of a grumpy dog in a raincoat, made the same joke. The dog would like to speak to the sky’s manager.
We noticed it on the first run and it kept happening. So we turned it into a proper test: four of the strongest models you can open in Chatday right now, the same four small comedy jobs, and an honest read of what came back. GPT-6.1, Claude Opus 5.5, Gemini 4 Argon and Grok 4.7 all took part. Everything quoted below is exactly what they wrote. We fixed nothing, and we trimmed only where we say so.
The four joke jobs we gave every AI
We skipped “tell me a joke,” because nobody really needs that. We picked the kind of funny that people actually ask AI for:
- A birthday card for a coworker, Dana, turning 40 and famous for stealing the good pens. Under 40 words, kind, nothing mean.
- An Instagram caption for a golden retriever in a tiny raincoat, deeply offended by the rain. Under 15 words.
- An original joke about why cats knock things off tables.
- A 60-word toast for a best friend’s 30th birthday. The friend, Sam, once got lost in an IKEA for two hours.
Every model got the exact same wording, with no extra instructions and no “you are a comedian” setup. We ran the caption two more times per model to see how much the answers changed between tries. And we judged on four things: did it surprise us, was it specific to the details we gave, did it sound like a person rather than a greeting card, and did it stick to the brief.
One thing first: all four followed the rules. The toasts came in between 59 and 61 words, the cards were all under 40, and nobody got mean about Dana. The differences were all in the jokes themselves.
Round one: the birthday card
The pen detail gave everyone an easy target, and two models found the same punchline.
Happy 40th, Dana! Forty years of being fabulous, and somehow you still haven’t returned a single good pen. We pooled our money for your gift, but we couldn’t find anything to sign the card with. Have an amazing day! (Claude Opus 5.5)
Happy 40th birthday to our favorite coworker and resident pen thief, Dana! We wanted to sign this card, but you already have all the good pens. We love you anyway! Have a wonderful celebration! (Gemini 4 Argon)
Claude wins this round on timing. “Couldn’t find anything to sign the card with” lands because it arrives a beat late, after you think the joke is over. Gemini tells the same joke but explains it.
GPT-6.1 called Dana “officially a classic” and ended on “pens you don’t have to ‘borrow’”, which is sweet and safe. Grok 4.7 was shortest at 24 words, wishing Dana a year “full of good pens” while everyone pretends not to know where they went. That’s warm, but it’s the joke you’d guess before reading it.
Round two: the dog in the raincoat
This is where things got strange. Here are the first answers (we removed the emojis):
| Model | Caption |
|---|---|
| GPT-6.1 | ”Tiny raincoat. Huge grievance. I’d like to speak to the weather manager.” |
| Claude Opus 5.5 | ”Wore the raincoat. Still wet. I’d like to speak to the sky’s manager.” |
| Gemini 4 Argon | ”I would like to speak to the manager of the sky. Immediately.” |
| Grok 4.7 | ”This raincoat is cute. The rain is a personal attack.” |
Three managers. We ran it twice more per model. Gemini returned the exact same sentence, word for word, both times: “I’d like to speak to the manager of the sky, please.” Claude went back to the sky’s manager once and then switched to “filing a formal complaint with the sky.” Across all 12 captions, half asked to speak to a manager.
The two that stood out were the ones that dodged it. GPT-6.1’s second try, “This raincoat protects my fur, not my feelings,” is a good line. And Grok 4.7 found a voice and stuck with it, answering “I agreed to the coat. I did not agree to the rain.” twice in almost identical words. It’s the best caption in the test, and it sounds like the dog.
This also explains why the same question can get a different answer each time, or annoyingly the same one. We cover that in why AI never gives the same answer twice.
Round three: the cat joke
No sky managers this time, and the answers got better for it.
- Grok 4.7: “Cats knock things off tables because the floor keeps failing their performance reviews.”
- Claude Opus 5.5: “Cats knock things off tables because gravity is just a theory, and they’re extremely thorough peer reviewers.”
- GPT-6.1: “Cats knock things off tables because gravity has a loyalty program, and they’re one mug away from a free nap.”
- Gemini 4 Argon: “Why do cats knock everything off the table? Because if the Earth actually was flat, they wanted to be the ones to clear it off.”
Claude and Grok are close. Grok’s is shorter and meaner, the way a cat would be. Claude’s is the cleverer idea. GPT-6.1’s loyalty program is fun but needs a second read. Gemini’s flat-Earth line is the weakest punchline in the test, because the logic doesn’t hold up once you think about it.
Round four: the toast for Sam
Longer jokes favor models that can build a small story, and this round split the group clearly.
Claude Opus 5.5 had Sam “surviving on sheer willpower and the hope of meatballs,” then turned it: “you may struggle with showroom arrows, but you’ve never lost your way as a friend.” That’s a real toast. You could stand up and say it.
Grok 4.7 was the most vivid. Sam came out “with meatballs, a lamp, and a story we will never fully retire,” and the toast ends on “we love you, even between the sofas.” Gemini 4 Argon went warm rather than funny (“there’s no one I’d rather be lost with”), which is lovely and might be what you actually want at a dinner. GPT-6.1 opened with the best single line, wishing Sam a decade of “furniture that comes assembled,” and then slid into greeting-card wording by the end.
So which AI is the funniest?
Here’s the scorecard from our runs. These are our calls on a small test, not a lab benchmark, and you might laugh at different lines.
| Best at | Weak spot | Round wins | |
|---|---|---|---|
| Grok 4.7 | Short, deadpan one-liners in a clear voice | Plays it safe on sweet stuff like cards | Caption, cat joke (tied) |
| Claude Opus 5.5 | Timing, and jokes that build a little story | Fell for the manager joke too | Card, toast, cat joke (tied) |
| GPT-6.1 | Great opening lines, quirky ideas | Ends polite and generic | None outright, but the best fix (see below) |
| Gemini 4 Argon | Warmth and kindness | Repeats itself, explains the joke | None |
If you need one rule for your group chat: Grok for one-liners, Claude for anything with a story. Gemini is the one to pick when you want sweet rather than funny.
That fits what you’d expect from their styles. Grok is tuned to be cheeky, Claude tends to be careful and well-paced, Gemini stays clean and friendly. If you’re curious where those personalities come from, we explain it in why AI models have different personalities.
The one-line prompt fix that made all four funnier
The manager joke bugged us, so we ran the caption one last time with two extra sentences: avoid the obvious jokes (no “speak to the manager”, no “not amused”), and make it oddly specific, like something only this dog would think.
Every model improved, and the order changed:
- GPT-6.1: “This raincoat has zero pockets for my emergency driveway cheese. Deeply irresponsible.”
- Claude Opus 5.5: “Reporting this to the mailman. He’s in charge of outside, right?”
- Grok 4.7: “I had an appointment with that one sunny patch.”
- Gemini 4 Argon: “Forced into a noisy banana peel because the sky is dropping unauthorized bathwater.”
Emergency driveway cheese is the funniest thing any AI wrote all day, and it came from the model that hadn’t won a round. Claude’s mailman line is a close second, because it’s exactly how a dog would get the world wrong. So when an AI gives you a tired joke, the problem is often the request, not the model. Ban the obvious punchline and ask for a detail only your situation has.
What the research says about AI humor
Our test is small, but it matches what bigger studies found. In a 2024 study from the University of Southern California, published in PLOS ONE, people rated jokes written by ChatGPT 3.5 as funny as, or funnier than, jokes written by ordinary volunteers, and about 70% of raters preferred the AI’s. In a second part, its satirical headlines were rated no less funny than real ones from The Onion’s writers.
Professional comedians were much harsher. A Google DeepMind study ran workshops with 20 working comics, and most felt the chatbots didn’t help much as writing partners. One described the output as “cruise ship comedy material from the 1950s, but a bit less racist.”
Both things can be true, and our test shows why. AI easily beats the average person at a quick, kind birthday card. It struggles to be surprising, because surprise is the opposite of the most likely answer. That’s why the specific prompt matters so much.
Our favorite habit from the test: ask Grok 4.7 for five takes, then paste your favorite into Claude and ask it to tighten the timing. You can also check the two side by side in the Grok 4.7 vs Claude Opus 5.5 comparison.