If ChatGPT Ever Tells You A Joke… DON'T LAUGH
MirrorLume
0:00 / 0:00
If ChatGPT Ever Tells You A Joke… DON'T LAUGH
10 просмотров · 23 ч назад
MirrorLume
22 подписчика
10 просмотров · 23 ч назад
Germany, 2023. Two researchers asked ChatGPT for a joke more than a thousand times. Out of 1,008 jokes, over 90% were the same 25 jokes, over and over. The favourite came up 140 times: "Why did the scarecrow win an award? Because he was outstanding in his field."
You probably didn't laugh. So let's find out why, because figuring out why a machine isn't funny is a good way to see what's going on in your own head when you laugh.
We go one level at a time. What a joke is: the old idea of incongruity, a path in your head that the punchline flips onto another track, so for a second both meanings are true. Why the machine plays it safe: a chatbot guesses the next likely word, and the most likely joke is the one it has seen the most, so it is, in a way, trained against surprise. Why you can't tickle yourself: your cerebellum predicts your own touch and turns the feeling down; jokes and tickles both need you not to see it coming, and "benign violation" says funny lives where a little wrongness meets safety. What laughing is for: Robert Provine found most laughs follow ordinary lines, not jokes, and people laugh about 30 times more with others; even tickled rats chirp. And is AI really not funny? On a page, judged by strangers, it holds its own; in a room full of professional comedians, it sounded like cruise ship comedy from the 1950s.
Sources
Jentzsch and Kersting (2023), "ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models" (1,008 jokes, 909 identical to one of the top 25, the scarecrow joke 140 times, all top 25 found online, made-up explanations for non-jokes): https://arxiv.org/abs/2306.04563
Incongruity theory and Kant's Critique of Judgment (1790): https://plato.stanford.edu/entries/hu...
Blakemore, Wolpert and Frith (1998), Nature Neuroscience, why self-produced touch tickles less: https://web.math.princeton.edu/~sswan...
Scientific American, why you can't tickle yourself: https://www.scientificamerican.com/ar...
McGraw and Warren (2010), Psychological Science, benign violation theory: https://leeds-faculty.colorado.edu/mc...
Robert Provine's laughter research: https://www.webmd.com/men/features/wh...
Panksepp, tickled rats chirp at about 50 kHz: https://www.scientificamerican.com/ar...
Ishiyama and Brecht (2016), Science, tickled rats jump for joy: https://www.npr.org/sections/health-s...
Gorenz and Schwarz (2024), PLOS ONE: https://www.psypost.org/ai-outshines-...
Mirowski et al. (2024), Google DeepMind, "A Robot Walks into a Bar": https://deepmind.google/research/publ...
Heise on the comedian study: https://heise.de/en/news/Stand-up-com...
The joke piles use the real counts of the top 25 jokes from the paper's appendix; only the scarecrow joke is shown. The "next word" bars are our own math on those counts (how often each word followed "Why did the" in its jokes), not ChatGPT's internal numbers. The ChatGPT windows, web pages and newspapers are our recreations; no other joke or headline is shown. The slider, the tickle waves and the Venn diagram are simplified sketches. Painted scenes are illustrations; no real person is painted.
Photos (Wikimedia Commons)
Immanuel Kant, c. 1790: public domain
Critique of Judgment, 1790 title page: public domain
Kristian Kersting: Stephanie Fröba, CC BY 4.0
Sarah-Jayne Blakemore: Shamus O'Reilly, CC BY 2.0
MRI scanner: Jan Ainali, CC BY 3.0
Cerebellum (DK atlas): Razvan V. Marinescu, CC BY 4.0
Peter McGraw: LeedsAlumni, CC BY-SA 4.0
Lab rat: Janet Stephens, public domain
Jaak Panksepp: Andres Tennus / University of Tartu, CC BY-SA 3.0
Edinburgh Festival Fringe, 2014: Brian McNeil, CC BY 3.0
Liner Queen of Bermuda, 1952-53: Tupelo the typo fixer, CC BY-SA 3.0
Logos: OpenAI and Google DeepMind, used to name the companies. Icons: Material Design Icons (Apache 2.0).
Chapters
0:00 1,008 jokes
0:17 The same 25 jokes
0:37 Why didn't you laugh?
1:02 Level 1: what is a joke?
1:21 The flip
1:57 Level 2: why the machine plays it safe
2:12 The joke it has seen the most
2:40 Trained against surprise
2:57 Level 3: why you can't tickle yourself
3:07 The 1998 brain scans
3:38 Benign violation
4:08 Level 4: what is laughing for?
4:33 Thirty times more with friends
4:45 Rats laugh too
5:11 Level 5: is AI really not funny?
5:52 Twenty comedians
6:16 Back to you