April 22, 2026 • 35 mins
Can computers crack jokes? Are comedians in trouble? Jorge talks to an AI expert to find the punchline.
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:01):
Hey, welcome to Sign Stuff, a production of iHeartRadio. I'm
hoh hitch Cham and today we're asking the question can
AI be funny? Sure you can ask chat GPT or
Claude to tell you a joke, but are you actually
going to laugh? We're gonna talk to an AI expert
who's been obsessed with deconstructing humor and teaching it to
(00:22):
a computer for the last fifteen years, and she's gonna
step us through the rules of what makes something funny,
whether AI can follow them, and whether comedians have their
days number. It's a pretty interesting conversation, and I have
to warn you there's a lot of laughing at it.
It turns out AI researchers can be pretty funny, So
(00:44):
get ready to hear the one about the computer scientist
walks into a bar with a cartoonist as we answer
the question can AI be funny?
Speaker 2 (00:54):
Enjoy?
Speaker 1 (00:57):
Hey everyone, So, I don't know if you've seen this,
but the company Anthropic just released the report that shows
the jobs that are most likely to be replaced by
AI in the coming years. If you're in management, business, computers, engineering, law,
even science, it's not looking so good for you. On
the other hand, if you're in agriculture, grounds maintenance, or
(01:20):
if you know how to fix a fridge or ac
you might be okay, which made me wonder where comedians stand.
Some people say that comedy and humor are some of
the things that make us uniquely human, but is.
Speaker 2 (01:33):
That really true?
Speaker 1 (01:34):
Can AI be just as funny as people? To answer
this question, I reached out to doctor Lydia Chilton, a
professor of computer science at Columbia University who specializes in
AI and human computer interaction. She also happens to have
spent the last fifteen years trying to prove that humor
can be done by a computer, because if it can,
(01:57):
it means we actually understand what humor is. Now, doctor
Chiltern is gonna tell us whether or not she actually succeeded,
but first I wanted to ask her a more basic question,
which is what makes something funny? So here's my conversation
with doctor Lydia Chilton. Well, thank you doctor Chiltern for
(02:18):
joining us.
Speaker 2 (02:19):
Thank you excited to be here.
Speaker 1 (02:21):
I have a joke for you. Here we got knock.
Speaker 2 (02:23):
Who's there?
Speaker 1 (02:24):
Iva?
Speaker 2 (02:25):
Iva? Who?
Speaker 1 (02:26):
I've got a feeling I'm going to need more data
to finish this joke.
Speaker 2 (02:32):
All right, you got me that's pretty funny.
Speaker 1 (02:35):
Well, actually I got this one from Gemini, So I
asked Gemini right before this call to tell me a
knock knock joke about Ai.
Speaker 2 (02:42):
Wow.
Speaker 1 (02:43):
Yeah, go Gemini. Yeah, and you laughed.
Speaker 2 (02:46):
Hi did.
Speaker 1 (02:48):
Well. That brings me to the first question I have
for you, which is what makes something funny? M M.
Can you pinpoint when you started to get curious about humor?
Speaker 2 (02:56):
Ah, I would say I was five years old and
I accidentally made a pun and my dad laughed so hysterically,
and I felt so good that I made my dad
last and I was like, must repeat, must repeat, And
of course it can't happen on command, and that one
was an accident. But in order to get people to
(03:18):
like me, I've wanted to figure out the formula for
a long time and been convinced that there is.
Speaker 1 (03:23):
One, which is a very sweet reason to be interested
in the topic.
Speaker 2 (03:28):
Yeah.
Speaker 1 (03:29):
I read that in twenty thirteen, you went to the
internet and you try to recruit a whole bunch of
people from the Internet to give you money to try
to answer this question. Can you tell us about that?
Speaker 2 (03:40):
Yeah? So there's some questions that are just so perennial
that I can't stop myself from thinking about them. This
is one you know, Plato had his set what is good,
what is justice? I've got mine what is funny? And
I actually think they're kind of the same. You know,
not to put me in Plato in the same boat,
but right, No, they both have a strong emotional component
(04:03):
to them. There's something about the chemicals that happen in
our brain that makes something funny. And that's why something
could be different funny for you, for me, for me
at a different time, for me if I heard it
slightly differently, maybe if someone less funny told it. So
it's not all about just the joke and whether the
joke was funny. There's a lot of other circumstances, Like
(04:26):
the whole thing isn't like logically constructed.
Speaker 1 (04:29):
That bothers you, well, it makes it.
Speaker 2 (04:32):
Very hard to understand. But like, so that could either
be a good thing or a bad thing. Like a
bad thing if you're like I need to understand this
before the test tomorrow, but a good thing. It's like,
I would like to study this for eternity because there's gonna.
Speaker 1 (04:42):
Be lots of problems. We may never figure it out.
Which gives you, as a professor job security.
Speaker 2 (04:48):
Exactly, exactly. So I talked to a lot of people
about this, read a lot of books, because it turns
out I wasn't the first person to study humor and
a philosopher had a theory about this is at least interesting.
It was evolved in us so that we could learn
from mistakes of others. Someone else falls on a banana
peel and liked I would not do that. And you know,
(05:15):
it's hard to say, but if we're looking back for
some origins of why we might have benefited as a
species from this, learning from other people's mistakes is a
pretty good way of learning rather watching someone else and
learning don't step on a banana peel as a positive reward.
Speaker 1 (05:31):
Interesting. Wait, so the theory is that funnyness is basically
shot and fraud.
Speaker 2 (05:35):
Yes, yes, shot and freud to make us better In theory.
Speaker 1 (05:40):
Mean he liked a way for your body to feel
good at the expense of others, so that we learn
from them.
Speaker 2 (05:46):
Yeah. So the next time you see a banana peel,
you have that memory because it's a strong positive You're like, oh,
that guy felt maybe I won't step on it.
Speaker 1 (05:54):
That's funny. I don't want to be the person people
laugh at. Oh interesting, Okay, no one.
Speaker 2 (06:01):
Can prove that's true or not. I can't go back
in time twenty thousand years and write down all the
banana pe ins.
Speaker 1 (06:07):
It iss although that would be pretty funny, Like.
Speaker 2 (06:10):
Where's the data, It's nowhere. It's a theory.
Speaker 1 (06:14):
It is very complicated and mysterious.
Speaker 2 (06:17):
Humor exactly job security, at.
Speaker 1 (06:21):
Least for now until potentially AIS can do it. So
we'll get to that.
Speaker 2 (06:25):
Good.
Speaker 1 (06:26):
Okay, what else did you learn about what makes something funny?
Speaker 2 (06:29):
Well? I read a whole bunch of theories, and there's
a lot of different ones, and they all like kind
of go together. So one is it has to be surprising.
It's not surprising. It's hard to get any kind of
emotional reaction out of it. Surprising often is correlated with
quick or has a quick turn, and it's like, oh,
I'm saying one thing and that boops, just kidding.
Speaker 1 (06:49):
Somebody's walking suddenly they slip in a banana.
Speaker 2 (06:51):
Whoa what exactly? Because you expected them to keep walking
and then right very suddenly if they fall very slowly,
that's less funny. Yeah, but the most important is it
has to be in the subtext of what set So
all stories, and I consider jokes stories for small stories
have like the text like what said, and then the
(07:12):
subtext is what it really means are what the significance
is to it. Typically jokes they have this surprise in
them because at first they leave you to believe that
there's one subtext, and then something happens you're like, well,
now there's a total other subtext.
Speaker 1 (07:28):
I see. It's about our expectation about what's under the surface.
Speaker 2 (07:32):
And how you're interpreting it. So my go to explainer
joke for this is here goes there are three kinds
of people in the world, people who can count and
people who can't. So I led you to believe that
I was a person who is intelligent and had three
things to say, and then my list too, and you're like, wait,
(07:53):
what in your brain kind of rearranges and then you realize, oh,
she can't count, she's an idiot, haha, and that's what
makes it funny. So those are the two texts.
Speaker 1 (08:03):
Yeah, I was expecting three things. I was surprised when
you stopped it two. But then in my brain pieced
it together. Oh, it's because she said she must be
one of the people who can't count exactly.
Speaker 2 (08:14):
So there's some expectation violation in their expectation violation is
one of the main theories, but it's not sufficient. If
I just said purple, you weren't expecting me to say purple.
But it's not funny, So the subtext is very very important. Okay,
that said, farts are also funny, and that doesn't Uh,
it's hard to explain that with samantic script theory of humor.
(08:36):
I don't know what the subtext the part.
Speaker 1 (08:39):
Is, Well, definitely they're funny, especially if you're ten years old. Yes,
maybe the subtext is that you're not supposed to do
it in public for others to hear or smell. Yeah, okay,
this is going to be a little important later on
when we talk about how to get a computer to
be funny. One of the most popular theories about humor
(09:00):
is that it violates our expectations in a surprising but
not unrelated Wait, so the first part of the joke
makes us assume one context, but then something happens and
it turns out there was a second hidden way to
interpret what was happening, which we had missed. Of course,
That's not the only thing that makes something funny.
Speaker 2 (09:21):
There's a lot of other things that make a joke
more or less funny. Okay, I call them heighteners. They
just heighten the surprise or the level of funny or
something about the emotion with them. So one is being mean,
Like typically the more mean a joke is the funnier
it is until it like really chips over and it's
like too soon or like or too.
Speaker 1 (09:41):
Mean, meaning like somebody has to be the butt of
the joke.
Speaker 2 (09:44):
Yeah, so mine was mean. I was the butt of
my own joke that I can't count. Uh, it's funnier
if it's mean.
Speaker 1 (09:51):
It has to say it expose somebody, some person, Yeah, right,
Like it can't be about exposing a chair or exposing
a material or or something nerds.
Speaker 2 (10:00):
Something related to people. It could be like a group
of people, like women or men. There's a lot there's
a lot of jokes that take that persuasion.
Speaker 1 (10:09):
And that movie pats into that shout and Freud shot
and Freud.
Speaker 2 (10:13):
Yes, schotenfreud of. I don't know it's German. Who knows
how to pronounce it? Probably nobody. It's not like there's
a country.
Speaker 1 (10:19):
Of people that and that would be a fandfire by
my lacke of language skills. But it's like you said,
it's about learning from the mistakes of others.
Speaker 2 (10:27):
Yes, exactly interesting, So.
Speaker 1 (10:30):
That makes jokes even funnier.
Speaker 2 (10:32):
Yeah. Another thing that makes them funnier is when you
really relate to them, and particularly in an in crowd
kind of way, when you're like, oh, I get back.
So when someone tells, like a computer science joke that
I'm like, ooh, I get it, and I feel special
because the physicists aren't going to get it. And this
can be whether it's about our generation, our family, or anything,
(10:55):
but inside jokes definitely produce a special kind of chemical
that make us feel like we're part of something, and
that heightens it.
Speaker 1 (11:03):
Yeah, that's funnier.
Speaker 2 (11:04):
It is funny or more emotional, and that heightens the funny.
Speaker 1 (11:08):
Because it exposes something you didn't think anybody else knew,
something you should talk about.
Speaker 2 (11:13):
So yeah, part of it is, and Freud talked about this,
a part of it is the relief of being able
to say something that you might not have otherwise been
able to say or reveal. You know, in society we're
just to go around and show our best face to
people pretend that we're awesome, and to have Instagram pages
that makes us look like we're in vacation Italy half
(11:34):
the time. And so when you realize that someone else
is also not actually going to Italy, they took one
vacation and been dripping the photos throughout the year that yeah,
my Instagram page is a lie, and that sense of
relief to people that yes, I'm like that too. It
rings true. Something being true like it also heightens it.
There's like a satisfaction and be like, yeah, that's how
(11:56):
it really is. That person's telling it straight. This was
like a mental satisfaction of when you hear something that's true.
Speaker 1 (12:03):
Right right, Like it puts a voice to something you've
suspected or felt but never actually maybe put into words before.
Speaker 2 (12:11):
Yeah. Right, And so there's still like a subtext going
on there, but the subtext that you're hearing is like, oh, yes,
I always wanted someone to admit that airline food is bad.
I've never been able to express that before.
Speaker 1 (12:24):
I see keep quing, keep quick.
Speaker 2 (12:26):
Since we've already opened up the topic of fart jokes,
things being dirty, sexual body related, all the things that
swear words have in common. There's a special part of
our brain that is like reserved for like swear words
and taboo things. Anytime you light that up, it also
sort of magnifies anything, whether it's a joke or a
(12:46):
compliment or anything else. It's just a heightener of all kinds.
Extra brain cells start firing when you hear those kinds
of words. M So try adding a swear word.
Speaker 1 (12:58):
I was gonna say, that is so effing true.
Speaker 2 (13:03):
And that was funny because you did that. There's many
things that heighten our emotions or feelings of something. Some
people think certain letters and words are very funny. So
words with more k's in them will make people laugh more.
Maybe I just where's the data? Yeah, there's many many
hypotheses out there waiting to be tested about what's funny? Yeah,
(13:26):
about what heightens jokes, what makes them funny? Like exactly
what kind of subtext is and is not funny? Where
in the exact turn is all right?
Speaker 1 (13:35):
So, doctor Chilton spent years studying humor and what makes
things funny. She talked to comedians and tapped into the
humor science community. Yes there is one, and she came
out of it with a bunch of rules about comedy.
The next step was to see if she could get
a computer to follow those rules. Could an AI be funny?
So when we come back, we'll talk about the different
(13:57):
ways in which programmers have tried to do that, and
with a the joke is on AIS or on us,
So stay with us for the punchline. We'll be right back. Hey,
(14:18):
welcome back. We're talking about whether AI can be funny,
and our guest today is doctor Lydia Chilton. As we
mentioned before, in twenty thirteen, she started a project to
dissect what makes something funny, not only to understand humor
to satisfy her own curiosity, but to see if you
could get a computer to be funny. However, not everyone
(14:40):
thought it was a good idea.
Speaker 2 (14:43):
Oh I got a lot of hate mail too, but
it's the Internet. Of course you're gonna get hate mail.
I'm like, how dare you? I thought that was interesting too,
that people be offended by this, and I'm like, it's science,
you know. So for me, that was the first tip
of people being threatened by AI, especially when computers creep
into creative areas. I think people kind of get this,
(15:05):
but that's mine. I'm a human. I'm creative. That's part
of my identity.
Speaker 1 (15:08):
Oh, you got pushback for even kind of trying to
dissect it, maybe with the goal of getting computers to
do it.
Speaker 2 (15:16):
Yeah, And I was just saying, well, it's just dissecting it,
but like, come on, if you dissect it well enough,
there's a very high probability that you'll be able to
generate it. But these weren't even like honestly comedians were
more like, yeah, I kind of want to know this too.
They weren't just threatened by it.
Speaker 1 (15:33):
Okay. So despite this pushback, doctor Chilton pressed on. But
she needed data on humor, so she turned to The Onion,
the humor and satire magazine that's been published since nineteen
eighty eight.
Speaker 2 (15:46):
And so someone pointed out to me that The Onion,
which is known for making up really funny fake headlines
and news stories, also had a different section that was unusual.
They took a real headline and came up with three
funny man on the street responses to it from average,
usually idiotic Americans. And I liked this as a framework
(16:09):
for studying the joke because you could just take the
input the original headline, and then you had the Onions
like verifiably funny statements, and then you could, with the
same headline, come up with your own or try to
back engineer the jokes that the onion made to figure out, Okay,
what are some of the properties of these jokes, and
what are some of the techniques for doing them, and
(16:31):
what's the variety in them. So this little test bed
became very important as just a mechanism for which I
could sort of study humor in a.
Speaker 1 (16:39):
Bottle and said, what did you learn that experiment you
did when in twenty.
Speaker 2 (16:43):
Twenty thirteen, twenty fourteen, twenty fifteen. It kept going because
there was a lot to learn, as it turned out,
But we did find some things that all the jokes had.
All the jokes had at least two connections to the headline.
They were taking the original headline and taking the kind
of associating things with those entities, with the people mentioned
(17:08):
in them, and finding new things to say about them
and a new connection between them.
Speaker 1 (17:13):
I see, like you're saying, like finding that other subtics
that is surprising.
Speaker 2 (17:17):
Exactly, And then I was like, uh duh. As a
computer scientist, I realized I'd kind of been looking in
the wrong places. I had been looking for like the
logic behind jokes. But I quickly realized that jokes and
most human communication is about our loose associations, like why
when I think McDonald's, why do I think Burger King?
(17:37):
Why do I It's just like they're in the same category. There,
It's an association I have. And at the time, computer
science was all logic, and so I actually kind of
put the project on a shelf. But I realized, Okay,
we need an association engine.
Speaker 1 (17:52):
Okay, let me see if I get this. It was
like the twenty early twenty tens. You dissected what humor is.
You sort of found all the these patterns and.
Speaker 2 (18:01):
Ye by analyzing the onion.
Speaker 1 (18:03):
By analyzing you peeled back the layers of the onion. Yeah,
you found the nuggets, some nuggets and rules, some patterns
about what makes something funny. And your goal was to
maybe try to get a computer to do this, right.
Speaker 2 (18:18):
Yeah, But it couldn't.
Speaker 1 (18:19):
It couldn't at the time because I think computer science
back then and artificial intelligence back then was kind of
about setting up the right rules, right.
Speaker 2 (18:28):
Yeah, it was all about logic. It was like, what
are the logical rules to do this? How do I
add one plus one. I take this bit and I
combine it with this bit, and I do an and
and that creates two. All of that kind of stuff,
But no, like, what do you associate with one? Oh?
One is the loneliest number number one, We're number one,
(18:48):
Avis is number two. Those are the thoughts that people
have about one that computers have no idea. And it's
even hard to get from the Internet, I see because
it's the things that people say and not necessarily really
the things that people write down. And it's about frequency,
like how often it happens, rather than the fact that
it did happen.
Speaker 1 (19:08):
Okay, okay, I am getting here to the nugget of
the onion here. You sort of figured out that humor
was about starting with a subtext that people would recognize,
but then being able to find that secondary that other
subtexts that will be surprising when you put all the
pieces together in the joke.
Speaker 2 (19:24):
Yes, surprising, but still relevant.
Speaker 1 (19:26):
It's still relevant, that's right. Second, suptics and trying to
find those subtags is hard if you're just going by
the literal definition of worth and using logic, right and
things like that.
Speaker 2 (19:37):
Exactly, it's just not there.
Speaker 1 (19:39):
So you start said we can't do this with rules
and logic, so you shelved it.
Speaker 2 (19:43):
Yeah, I gave up.
Speaker 1 (19:44):
People weren't convinced that computers could be funny.
Speaker 2 (19:47):
Yes, they weren't convinced that I had decomposed and reconstructed
the process of writing humor when humans were still involved
in some part of that process.
Speaker 1 (19:57):
Oh, I see, because you still needed some human to
direct the computer.
Speaker 2 (20:01):
Like come up with the associations, which is fair, but
they're absolutely correct. And then I put this on hold
for forever. I was like, never again. I've been burned
by this. I will never again grace the universe with
my thoughts on humor.
Speaker 1 (20:14):
Humor, this is not funny anymore.
Speaker 2 (20:16):
Exactly, this isn't funny anymore.
Speaker 1 (20:19):
Yes, at this point, computer scientists thought maybe it was
impossible to really teach a computer to be funny. But
then something changed, something happened that made them think that
maybe it is possible for an AI to have a
sense of humor. So when we come back, we're gonna
talk about what that change was and how it unlocks
(20:40):
AI's funny bone forever. So stay with us. We'll be
right back. Hey, we'll come back. We're talking about whether
AI can be funny, and we're at the part of
(21:01):
the story where computer scientists didn't think it was possible
for a computer to make humor. There have been lots
of attempts to have computers recognize humor or jokes basically,
and they could do pretty well looking for language cues
like incongruities or alteration or slang. But could a computer
come up with a joke? That was the big question.
(21:24):
Now to understand what was happening at this point with
AI and humor, you kind of need to know a
little bit of the history of AI. Recovered this in
a lot of detail in our February fourth episode about
what AI slop is doing to us, So if you
want to dig deeper, go check out that episode. But
the basic idea is that for a long time, the
field of AI was largely based on logic, trying to
(21:46):
figure out strategies and algorithms that we thought made things intelligent.
But then in the mid twenty tens, computer scientists figured
out a different approach to AI that changed everything. They
started to use something new called the transformer, which in
turn you sink technique called attention that essentially let computers
(22:07):
learn context, and that led to the explosion of AI
systems like Chat, GPT, Gemini, Claude. And the key ability
here for humor is that these AI systems were built
to make associations here tector, Lydia, chiltern.
Speaker 2 (22:25):
So these language models GPT, claw, you know, the thing,
the ais that can generate text sort of taken over
the universe at this point, for better or for worse,
are called man, do I even know what it stands for? Yes?
I do. Large language models, so they take basically take
in all the text on the internet. They take a sentence,
they take the last word out of the sentence and
(22:46):
try to predict what that last word would be and
guess what that is. That's an association. They go, you
can't do that by logic. You just have to say, like, uh, goodbye,
so long don't have you know fun or like have
like that's an association. There's no logic that tells you
you should do You've just heard it many times before,
you repeat it. Kids especially pick up on these things.
(23:08):
It's it's just like, well, our brains are hardwired to
find these associations and use them.
Speaker 1 (23:13):
So your team was like, we can do it. What
was the thing that specifically they thought they could do,
or that you could all do.
Speaker 2 (23:20):
Well, we knew that AI could do the associations. What
I always told them is we need a new evaluation mechanism.
The American Voices section of the Onion that I had
used was no longer quite as popular because this was
like fifteen years ago. So like, yeah, millennials loved it.
Gen Z just does not care. The student was as
(23:41):
gen Z as they come, and he's like, well, I
know what's funny. Here's what people do. They have a
caption contest on Instagram. People post a funny image and
then people try to caption it. We're like, okay, let's
go for it. Very similar to what the New Yorker does,
but the New Yorker's a hard to capture. And Sean
convinced me that j has a very particular flavor of humor.
Speaker 1 (24:04):
So the type of humor that doctor Chiltern and her
team decided to see if AI could make was gen
Z meme humor. That's when you see an image, let's say,
a person with a small hose trying to put out
a really large fire, and then someone writes underneath that
image to text me trying to put out the dumpster
fire of my last relationship. If you're a gen Z
(24:26):
that would be hilarious. Okay, here's the experiment, Doctor Chiltern
and her team there, they took images and then they
put funny captions to them from three sources. One was
real people. The images were put online and then people
competed to see who could come up with the funniest
caption for them. Where we got the most votes, that's
(24:46):
the one that represented how funny humans can be. The
second source was Chad GPT. They just asked Chad JPT
to come up a funny caption for the image for
a gen Z audience. But then there was a third
source of funny captions, which was Chad GBT, but with
specific instructions from doctor Chilton and her team on how
to make something funny. It prompted Chad GPT to come
(25:09):
up with a funny caption for a gen Z audience.
But in the prompt it would lay out the rules
of humor that doctor Chilton had been researching for years.
So step me through those instructions. He said, take this image.
Speaker 2 (25:23):
Yeah, so take this image. So it's an image of
like a little guy in the corner with a hose
and a big fire down below in the bottom of
a canyon. And so, with a dear GPT, please use
your vision model and describe this image. And it describes
exactly all those things, all these little details. Oh, it
says he's on a crane, the sky is blue, which
it was. There's lots of trees in the background, all
(25:45):
this stuff, some of it useful, some of it not.
You don't know what's going to be useful. And then
we say, okay, what are some of the dynamics happening
in this image, Like, well, this guy's opting out the fire,
but it looks pretty ineffective. The fire's raging through this canyon,
probably going to destroy a lot of things. So it
sort of like does when in storytelling is sometimes called
(26:05):
world building, like imagine out like not just what scene,
but you know some other things surrounding it, Uh huh
that could be happening or could happen next, And because
you need to start to build a story. Jokes are stories.
Speaker 1 (26:20):
There has to be a world, not just what you
see on this image.
Speaker 2 (26:23):
And then you say, now, let's think of an analogy
to something that is funny, like relationship drama. Uh huh,
and you have AI associate. Okay, what are abstract things
that you could put onto a relationship? And it did
that part all on its own. It would find many
things like you know, me trying to clean up after
a relationship. Then like you know, students are all obsessed
(26:45):
about their GPA, and it's like, you know, me trying
to recover my GPA after I tanked a final.
Speaker 1 (26:51):
It sounds like you were having a conversation with the CHADGBT.
Was it a conversation or was this just all in
one prompt?
Speaker 2 (26:58):
No, it did it all by it. So it's conversing
with itself. That's a big thing that we found was
important at the time. GPT. It's unclear that you need
it to do that now, but we're like, do step
one and then do step two. So like, describe the image,
elaborate on that image yourself, come up with figure out
what human dynamics you want to map that to make
(27:20):
twenty jokes. Evaluate those twenty jokes, tell us which are
the five best ones. Yeah, so we put it through
a whole series of prompts.
Speaker 1 (27:29):
So you got these three conditions. You got the AI
to generate captions with the instructions and then what did
you do with the results of all this?
Speaker 2 (27:38):
We showed people the original image. We got lots and
lots of people say, hey, here are some captions. Rate
them all, and we mixed in the ones we wrote
with the generic GPT ones and the human written ones,
and every joke got an evaluation how funny is it?
On a scale from one to five, And then we
could compare. Because people didn't know what condition it was,
(27:58):
we like randomized the or so there were no ordering effects.
We could compare who's funnier.
Speaker 1 (28:04):
Who's funnier? Well, the internet a chat GPT or a
GPT coach to be funny by us.
Speaker 2 (28:12):
So our coached one was definitely funnier than GPT by itself.
And we were almost as funny as the humans. In fact,
we were not statistically significantly different, which is not the
same as saying we were as funny as but you
can't prove your so basically we got.
Speaker 1 (28:31):
There pretty much. You were there.
Speaker 2 (28:34):
Yeah, you couldn't tell based on the funniness whether it
was written by our decomposed AI version or a human rhote.
Speaker 1 (28:41):
It we were on par meaning that you've proven two things.
You've proven that AI can be funny with the instructions, Yes,
and that these instructions are a key element of humor. Yes,
doctor Chilted as a humorist, as someone who makes a
living writing funny things. You just gave me a deep fear.
Speaker 2 (29:04):
Oh excellent. Describe that fear to me.
Speaker 1 (29:08):
You just made me feel like something I've done all
my life and that I thought I was good at.
Apparently you get just as chet gbt to do, and
I'll do it ten times one hundred times faster and
maybe better than me.
Speaker 2 (29:20):
Can I play with that thought a little bit? Yeah,
because I get this a lot. And anytime AI does
a thing like it, writes poetry, the poets go oh no, no, no,
I have no value anymore.
Speaker 1 (29:31):
I'm not going to say anything about poetry. We all
know what happened to Timothy Shemale. Shemale, Yeah, is that yeah?
Or opera. I'm not going to say anything about opera
or poetry. Go ahead, yeah, anyway.
Speaker 2 (29:42):
So this is a very common dynamic, Like this is
a human thing, Like we all get hyper sensitive about
the things that we identify with that we've constructed our
identity around and blah blah. We think we're special for
we think we're special for And there's a number of
ways of thinking about this. So first of all, probably
not as special as you think. Now that's it.
Speaker 1 (30:05):
I get it from my kids. Thank you children, thank you. Yes,
I think quickly, oh, you're not that funny dead, But.
Speaker 2 (30:14):
Also you're a little bit overestimating AI. So basically, AI
in this instance did find one way to fairly reliably
be funny. That does not mean it can do it
in every situation, that it can do it for every person.
Like also, humor is very situational. We want to talk
(30:34):
to one another and have jokes about what we're talking
about in our lives. This is one like small sliver
of the joke universe. And so yes, although I definitely
see it trigger alarm bells and people's mind I would
say you're both probably overestimating how much of a special
snowflake you are, but you're definitely extrapolating on how strong
(30:57):
AI is to be funny given all different contexts and
importances or important dimensions and places of being funny, like
at the right time, with the right place, to the
right person without being too offensive. So it's not like
it's solved. It's more of like an existence proof that
something is possible in this space. Should we be worried?
(31:19):
I don't know. No one will ever be funny again
because now AI does it, and the humans will lose
their ability to be funny. I don't see that happening.
Speaker 1 (31:31):
I feel like maybe you hit it a little bit
on the head a moment ago when you said that
part of what we're seeking here is human connection, and
a lot of what we find funny is maybe they're
more special, or they hit us more. If another person says,
I don't know, what do you think?
Speaker 2 (31:47):
Yeah, it's certainly social. There's something very special about what
a person, especially a person we like or admire, says
something to us. And even to have someone that you
like make a joke about you, it has a meaning
and a subtext beyond just like, oh you know, she
thinks that I farted or whatever. Like it means that
they like you, that there's a bond between you, and
(32:08):
you're not going to feel that with Ai. God, I
hope you won't feel it.
Speaker 1 (32:13):
Meaning there's hope in this sense. Or maybe I don't
know if we need hope, but we want to hear
humor from other people. Yeah, or it's maybe it's funnier
or more special if another person says it or is
behind it.
Speaker 2 (32:24):
Yeah. So if we brought jokes to storytelling, I've really
realized that AI has nothing to say. The heart of
a story is having something to say, some actual subtext
that you believe in. And when you tell me a
joke that comes from your background, that's maybe my shared background.
(32:45):
Maybe how awful it is getting a PhD, and how
it really feels like the machine is punching you down.
Like I realize that you have been through that and
you have this thing to say, and I feel it too,
And I think that's that having something to say is
the most important part of human creation.
Speaker 1 (33:04):
Right now, I feel like you're saying that AI can
be funny, but right now, at least you can't have
to tell it how to be funny. Yeah, And if
you're telling it how to be funny, maybe by then
the joke is old. It's like people will see through
and say, oh, this is just following that pattern that
I've seen a million times. Yes, exactly, all right. To
(33:27):
end here, I asked Gemini, tell me a joke about
a computer scientist who's trying to make AI funny. Do
you want to hear it.
Speaker 2 (33:35):
Yeah, yes, here it is.
Speaker 1 (33:38):
A computer scientist spent years training a massive neural network
to have the perfect sense of humor. On the day
of the big reveal, she invited the press and typed
the prompt tell me a joke, the AI word, and
finally it replied, your life. The scientist was horrible. It
doesn't It doesn't end there.
Speaker 2 (33:55):
It keeps going, doesn't. Oh good? Oh good. I was
hoping it would keep going.
Speaker 1 (33:59):
The scientist was horrified. That's not funny. That's just mean.
Why would you say that. The AI blinked its cursor
calmly and responded, because I've seen your source code.
Speaker 2 (34:09):
Hmmm, it's not for me. Maybe everyone else will think
it's funny.
Speaker 1 (34:13):
We need more data, We need more data.
Speaker 2 (34:18):
The first one was funny, get your life?
Speaker 1 (34:20):
No, no, oh they knock dock joke. Yeah, maybe I
should stick to knockout jokes. Is the lesson here? All right, Hey,
thanks for joining us, See you next time you've been
listening to science stuff. Production of iHeartRadio Bring Them Produced
(34:41):
by Me or Hey Cham edited by Rose Seguda. He
said gative producer Jerry Rowland, an audio engineer and mixer.
Kasey Peckram. You can follow me on social media to
search for PhD comics in the name of your favorite platform.
Be sure to subscribe to sign Stuff on the iHeartRadio app,
Apple podcasts, or wherever you get your podcast, and please
tell your friends we'll be back next Wednesday with another episode. Hey,
(35:18):
please take a second and leave us a review on
Apple Podcasts, Spotify, or wherever you listen to the podcast.
Thanks a lot,
Hey, welcome to Sign Stuff, a production of iHeartRadio. I'm
hoh hitch Cham and today we're asking the question can
AI be funny? Sure you can ask chat GPT or
Claude to tell you a joke, but are you actually
going to laugh? We're gonna talk to an AI expert
who's been obsessed with deconstructing humor and teaching it to
(00:22):
a computer for the last fifteen years, and she's gonna
step us through the rules of what makes something funny,
whether AI can follow them, and whether comedians have their
days number. It's a pretty interesting conversation, and I have
to warn you there's a lot of laughing at it.
It turns out AI researchers can be pretty funny, So
(00:44):
get ready to hear the one about the computer scientist
walks into a bar with a cartoonist as we answer
the question can AI be funny?
Speaker 2 (00:54):
Enjoy?
Speaker 1 (00:57):
Hey everyone, So, I don't know if you've seen this,
but the company Anthropic just released the report that shows
the jobs that are most likely to be replaced by
AI in the coming years. If you're in management, business, computers, engineering, law,
even science, it's not looking so good for you. On
the other hand, if you're in agriculture, grounds maintenance, or
(01:20):
if you know how to fix a fridge or ac
you might be okay, which made me wonder where comedians stand.
Some people say that comedy and humor are some of
the things that make us uniquely human, but is.
Speaker 2 (01:33):
That really true?
Speaker 1 (01:34):
Can AI be just as funny as people? To answer
this question, I reached out to doctor Lydia Chilton, a
professor of computer science at Columbia University who specializes in
AI and human computer interaction. She also happens to have
spent the last fifteen years trying to prove that humor
can be done by a computer, because if it can,
(01:57):
it means we actually understand what humor is. Now, doctor
Chiltern is gonna tell us whether or not she actually succeeded,
but first I wanted to ask her a more basic question,
which is what makes something funny? So here's my conversation
with doctor Lydia Chilton. Well, thank you doctor Chiltern for
(02:18):
joining us.
Speaker 2 (02:19):
Thank you excited to be here.
Speaker 1 (02:21):
I have a joke for you. Here we got knock.
Speaker 2 (02:23):
Who's there?
Speaker 1 (02:24):
Iva?
Speaker 2 (02:25):
Iva? Who?
Speaker 1 (02:26):
I've got a feeling I'm going to need more data
to finish this joke.
Speaker 2 (02:32):
All right, you got me that's pretty funny.
Speaker 1 (02:35):
Well, actually I got this one from Gemini, So I
asked Gemini right before this call to tell me a
knock knock joke about Ai.
Speaker 2 (02:42):
Wow.
Speaker 1 (02:43):
Yeah, go Gemini. Yeah, and you laughed.
Speaker 2 (02:46):
Hi did.
Speaker 1 (02:48):
Well. That brings me to the first question I have
for you, which is what makes something funny? M M.
Can you pinpoint when you started to get curious about humor?
Speaker 2 (02:56):
Ah, I would say I was five years old and
I accidentally made a pun and my dad laughed so hysterically,
and I felt so good that I made my dad
last and I was like, must repeat, must repeat, And
of course it can't happen on command, and that one
was an accident. But in order to get people to
(03:18):
like me, I've wanted to figure out the formula for
a long time and been convinced that there is.
Speaker 1 (03:23):
One, which is a very sweet reason to be interested
in the topic.
Speaker 2 (03:28):
Yeah.
Speaker 1 (03:29):
I read that in twenty thirteen, you went to the
internet and you try to recruit a whole bunch of
people from the Internet to give you money to try
to answer this question. Can you tell us about that?
Speaker 2 (03:40):
Yeah? So there's some questions that are just so perennial
that I can't stop myself from thinking about them. This
is one you know, Plato had his set what is good,
what is justice? I've got mine what is funny? And
I actually think they're kind of the same. You know,
not to put me in Plato in the same boat,
but right, No, they both have a strong emotional component
(04:03):
to them. There's something about the chemicals that happen in
our brain that makes something funny. And that's why something
could be different funny for you, for me, for me
at a different time, for me if I heard it
slightly differently, maybe if someone less funny told it. So
it's not all about just the joke and whether the
joke was funny. There's a lot of other circumstances, Like
(04:26):
the whole thing isn't like logically constructed.
Speaker 1 (04:29):
That bothers you, well, it makes it.
Speaker 2 (04:32):
Very hard to understand. But like, so that could either
be a good thing or a bad thing. Like a
bad thing if you're like I need to understand this
before the test tomorrow, but a good thing. It's like,
I would like to study this for eternity because there's gonna.
Speaker 1 (04:42):
Be lots of problems. We may never figure it out.
Which gives you, as a professor job security.
Speaker 2 (04:48):
Exactly, exactly. So I talked to a lot of people
about this, read a lot of books, because it turns
out I wasn't the first person to study humor and
a philosopher had a theory about this is at least interesting.
It was evolved in us so that we could learn
from mistakes of others. Someone else falls on a banana
peel and liked I would not do that. And you know,
(05:15):
it's hard to say, but if we're looking back for
some origins of why we might have benefited as a
species from this, learning from other people's mistakes is a
pretty good way of learning rather watching someone else and
learning don't step on a banana peel as a positive reward.
Speaker 1 (05:31):
Interesting. Wait, so the theory is that funnyness is basically
shot and fraud.
Speaker 2 (05:35):
Yes, yes, shot and freud to make us better In theory.
Speaker 1 (05:40):
Mean he liked a way for your body to feel
good at the expense of others, so that we learn
from them.
Speaker 2 (05:46):
Yeah. So the next time you see a banana peel,
you have that memory because it's a strong positive You're like, oh,
that guy felt maybe I won't step on it.
Speaker 1 (05:54):
That's funny. I don't want to be the person people
laugh at. Oh interesting, Okay, no one.
Speaker 2 (06:01):
Can prove that's true or not. I can't go back
in time twenty thousand years and write down all the
banana pe ins.
Speaker 1 (06:07):
It iss although that would be pretty funny, Like.
Speaker 2 (06:10):
Where's the data, It's nowhere. It's a theory.
Speaker 1 (06:14):
It is very complicated and mysterious.
Speaker 2 (06:17):
Humor exactly job security, at.
Speaker 1 (06:21):
Least for now until potentially AIS can do it. So
we'll get to that.
Speaker 2 (06:25):
Good.
Speaker 1 (06:26):
Okay, what else did you learn about what makes something funny?
Speaker 2 (06:29):
Well? I read a whole bunch of theories, and there's
a lot of different ones, and they all like kind
of go together. So one is it has to be surprising.
It's not surprising. It's hard to get any kind of
emotional reaction out of it. Surprising often is correlated with
quick or has a quick turn, and it's like, oh,
I'm saying one thing and that boops, just kidding.
Speaker 1 (06:49):
Somebody's walking suddenly they slip in a banana.
Speaker 2 (06:51):
Whoa what exactly? Because you expected them to keep walking
and then right very suddenly if they fall very slowly,
that's less funny. Yeah, but the most important is it
has to be in the subtext of what set So
all stories, and I consider jokes stories for small stories
have like the text like what said, and then the
(07:12):
subtext is what it really means are what the significance
is to it. Typically jokes they have this surprise in
them because at first they leave you to believe that
there's one subtext, and then something happens you're like, well,
now there's a total other subtext.
Speaker 1 (07:28):
I see. It's about our expectation about what's under the surface.
Speaker 2 (07:32):
And how you're interpreting it. So my go to explainer
joke for this is here goes there are three kinds
of people in the world, people who can count and
people who can't. So I led you to believe that
I was a person who is intelligent and had three
things to say, and then my list too, and you're like, wait,
(07:53):
what in your brain kind of rearranges and then you realize, oh,
she can't count, she's an idiot, haha, and that's what
makes it funny. So those are the two texts.
Speaker 1 (08:03):
Yeah, I was expecting three things. I was surprised when
you stopped it two. But then in my brain pieced
it together. Oh, it's because she said she must be
one of the people who can't count exactly.
Speaker 2 (08:14):
So there's some expectation violation in their expectation violation is
one of the main theories, but it's not sufficient. If
I just said purple, you weren't expecting me to say purple.
But it's not funny, So the subtext is very very important. Okay,
that said, farts are also funny, and that doesn't Uh,
it's hard to explain that with samantic script theory of humor.
(08:36):
I don't know what the subtext the part.
Speaker 1 (08:39):
Is, Well, definitely they're funny, especially if you're ten years old. Yes,
maybe the subtext is that you're not supposed to do
it in public for others to hear or smell. Yeah, okay,
this is going to be a little important later on
when we talk about how to get a computer to
be funny. One of the most popular theories about humor
(09:00):
is that it violates our expectations in a surprising but
not unrelated Wait, so the first part of the joke
makes us assume one context, but then something happens and
it turns out there was a second hidden way to
interpret what was happening, which we had missed. Of course,
That's not the only thing that makes something funny.
Speaker 2 (09:21):
There's a lot of other things that make a joke
more or less funny. Okay, I call them heighteners. They
just heighten the surprise or the level of funny or
something about the emotion with them. So one is being mean,
Like typically the more mean a joke is the funnier
it is until it like really chips over and it's
like too soon or like or too.
Speaker 1 (09:41):
Mean, meaning like somebody has to be the butt of
the joke.
Speaker 2 (09:44):
Yeah, so mine was mean. I was the butt of
my own joke that I can't count. Uh, it's funnier
if it's mean.
Speaker 1 (09:51):
It has to say it expose somebody, some person, Yeah, right,
Like it can't be about exposing a chair or exposing
a material or or something nerds.
Speaker 2 (10:00):
Something related to people. It could be like a group
of people, like women or men. There's a lot there's
a lot of jokes that take that persuasion.
Speaker 1 (10:09):
And that movie pats into that shout and Freud shot
and Freud.
Speaker 2 (10:13):
Yes, schotenfreud of. I don't know it's German. Who knows
how to pronounce it? Probably nobody. It's not like there's
a country.
Speaker 1 (10:19):
Of people that and that would be a fandfire by
my lacke of language skills. But it's like you said,
it's about learning from the mistakes of others.
Speaker 2 (10:27):
Yes, exactly interesting, So.
Speaker 1 (10:30):
That makes jokes even funnier.
Speaker 2 (10:32):
Yeah. Another thing that makes them funnier is when you
really relate to them, and particularly in an in crowd
kind of way, when you're like, oh, I get back.
So when someone tells, like a computer science joke that
I'm like, ooh, I get it, and I feel special
because the physicists aren't going to get it. And this
can be whether it's about our generation, our family, or anything,
(10:55):
but inside jokes definitely produce a special kind of chemical
that make us feel like we're part of something, and
that heightens it.
Speaker 1 (11:03):
Yeah, that's funnier.
Speaker 2 (11:04):
It is funny or more emotional, and that heightens the funny.
Speaker 1 (11:08):
Because it exposes something you didn't think anybody else knew,
something you should talk about.
Speaker 2 (11:13):
So yeah, part of it is, and Freud talked about this,
a part of it is the relief of being able
to say something that you might not have otherwise been
able to say or reveal. You know, in society we're
just to go around and show our best face to
people pretend that we're awesome, and to have Instagram pages
that makes us look like we're in vacation Italy half
(11:34):
the time. And so when you realize that someone else
is also not actually going to Italy, they took one
vacation and been dripping the photos throughout the year that yeah,
my Instagram page is a lie, and that sense of
relief to people that yes, I'm like that too. It
rings true. Something being true like it also heightens it.
There's like a satisfaction and be like, yeah, that's how
(11:56):
it really is. That person's telling it straight. This was
like a mental satisfaction of when you hear something that's true.
Speaker 1 (12:03):
Right right, Like it puts a voice to something you've
suspected or felt but never actually maybe put into words before.
Speaker 2 (12:11):
Yeah. Right, And so there's still like a subtext going
on there, but the subtext that you're hearing is like, oh, yes,
I always wanted someone to admit that airline food is bad.
I've never been able to express that before.
Speaker 1 (12:24):
I see keep quing, keep quick.
Speaker 2 (12:26):
Since we've already opened up the topic of fart jokes,
things being dirty, sexual body related, all the things that
swear words have in common. There's a special part of
our brain that is like reserved for like swear words
and taboo things. Anytime you light that up, it also
sort of magnifies anything, whether it's a joke or a
(12:46):
compliment or anything else. It's just a heightener of all kinds.
Extra brain cells start firing when you hear those kinds
of words. M So try adding a swear word.
Speaker 1 (12:58):
I was gonna say, that is so effing true.
Speaker 2 (13:03):
And that was funny because you did that. There's many
things that heighten our emotions or feelings of something. Some
people think certain letters and words are very funny. So
words with more k's in them will make people laugh more.
Maybe I just where's the data? Yeah, there's many many
hypotheses out there waiting to be tested about what's funny? Yeah,
(13:26):
about what heightens jokes, what makes them funny? Like exactly
what kind of subtext is and is not funny? Where
in the exact turn is all right?
Speaker 1 (13:35):
So, doctor Chilton spent years studying humor and what makes
things funny. She talked to comedians and tapped into the
humor science community. Yes there is one, and she came
out of it with a bunch of rules about comedy.
The next step was to see if she could get
a computer to follow those rules. Could an AI be funny?
So when we come back, we'll talk about the different
(13:57):
ways in which programmers have tried to do that, and
with a the joke is on AIS or on us,
So stay with us for the punchline. We'll be right back. Hey,
(14:18):
welcome back. We're talking about whether AI can be funny,
and our guest today is doctor Lydia Chilton. As we
mentioned before, in twenty thirteen, she started a project to
dissect what makes something funny, not only to understand humor
to satisfy her own curiosity, but to see if you
could get a computer to be funny. However, not everyone
(14:40):
thought it was a good idea.
Speaker 2 (14:43):
Oh I got a lot of hate mail too, but
it's the Internet. Of course you're gonna get hate mail.
I'm like, how dare you? I thought that was interesting too,
that people be offended by this, and I'm like, it's science,
you know. So for me, that was the first tip
of people being threatened by AI, especially when computers creep
into creative areas. I think people kind of get this,
(15:05):
but that's mine. I'm a human. I'm creative. That's part
of my identity.
Speaker 1 (15:08):
Oh, you got pushback for even kind of trying to
dissect it, maybe with the goal of getting computers to
do it.
Speaker 2 (15:16):
Yeah, And I was just saying, well, it's just dissecting it,
but like, come on, if you dissect it well enough,
there's a very high probability that you'll be able to
generate it. But these weren't even like honestly comedians were
more like, yeah, I kind of want to know this too.
They weren't just threatened by it.
Speaker 1 (15:33):
Okay. So despite this pushback, doctor Chilton pressed on. But
she needed data on humor, so she turned to The Onion,
the humor and satire magazine that's been published since nineteen
eighty eight.
Speaker 2 (15:46):
And so someone pointed out to me that The Onion,
which is known for making up really funny fake headlines
and news stories, also had a different section that was unusual.
They took a real headline and came up with three
funny man on the street responses to it from average,
usually idiotic Americans. And I liked this as a framework
(16:09):
for studying the joke because you could just take the
input the original headline, and then you had the Onions
like verifiably funny statements, and then you could, with the
same headline, come up with your own or try to
back engineer the jokes that the onion made to figure out, Okay,
what are some of the properties of these jokes, and
what are some of the techniques for doing them, and
(16:31):
what's the variety in them. So this little test bed
became very important as just a mechanism for which I
could sort of study humor in a.
Speaker 1 (16:39):
Bottle and said, what did you learn that experiment you
did when in twenty.
Speaker 2 (16:43):
Twenty thirteen, twenty fourteen, twenty fifteen. It kept going because
there was a lot to learn, as it turned out,
But we did find some things that all the jokes had.
All the jokes had at least two connections to the headline.
They were taking the original headline and taking the kind
of associating things with those entities, with the people mentioned
(17:08):
in them, and finding new things to say about them
and a new connection between them.
Speaker 1 (17:13):
I see, like you're saying, like finding that other subtics
that is surprising.
Speaker 2 (17:17):
Exactly, And then I was like, uh duh. As a
computer scientist, I realized I'd kind of been looking in
the wrong places. I had been looking for like the
logic behind jokes. But I quickly realized that jokes and
most human communication is about our loose associations, like why
when I think McDonald's, why do I think Burger King?
(17:37):
Why do I It's just like they're in the same category. There,
It's an association I have. And at the time, computer
science was all logic, and so I actually kind of
put the project on a shelf. But I realized, Okay,
we need an association engine.
Speaker 1 (17:52):
Okay, let me see if I get this. It was
like the twenty early twenty tens. You dissected what humor is.
You sort of found all the these patterns and.
Speaker 2 (18:01):
Ye by analyzing the onion.
Speaker 1 (18:03):
By analyzing you peeled back the layers of the onion. Yeah,
you found the nuggets, some nuggets and rules, some patterns
about what makes something funny. And your goal was to
maybe try to get a computer to do this, right.
Speaker 2 (18:18):
Yeah, But it couldn't.
Speaker 1 (18:19):
It couldn't at the time because I think computer science
back then and artificial intelligence back then was kind of
about setting up the right rules, right.
Speaker 2 (18:28):
Yeah, it was all about logic. It was like, what
are the logical rules to do this? How do I
add one plus one. I take this bit and I
combine it with this bit, and I do an and
and that creates two. All of that kind of stuff,
But no, like, what do you associate with one? Oh?
One is the loneliest number number one, We're number one,
(18:48):
Avis is number two. Those are the thoughts that people
have about one that computers have no idea. And it's
even hard to get from the Internet, I see because
it's the things that people say and not necessarily really
the things that people write down. And it's about frequency,
like how often it happens, rather than the fact that
it did happen.
Speaker 1 (19:08):
Okay, okay, I am getting here to the nugget of
the onion here. You sort of figured out that humor
was about starting with a subtext that people would recognize,
but then being able to find that secondary that other
subtexts that will be surprising when you put all the
pieces together in the joke.
Speaker 2 (19:24):
Yes, surprising, but still relevant.
Speaker 1 (19:26):
It's still relevant, that's right. Second, suptics and trying to
find those subtags is hard if you're just going by
the literal definition of worth and using logic, right and
things like that.
Speaker 2 (19:37):
Exactly, it's just not there.
Speaker 1 (19:39):
So you start said we can't do this with rules
and logic, so you shelved it.
Speaker 2 (19:43):
Yeah, I gave up.
Speaker 1 (19:44):
People weren't convinced that computers could be funny.
Speaker 2 (19:47):
Yes, they weren't convinced that I had decomposed and reconstructed
the process of writing humor when humans were still involved
in some part of that process.
Speaker 1 (19:57):
Oh, I see, because you still needed some human to
direct the computer.
Speaker 2 (20:01):
Like come up with the associations, which is fair, but
they're absolutely correct. And then I put this on hold
for forever. I was like, never again. I've been burned
by this. I will never again grace the universe with
my thoughts on humor.
Speaker 1 (20:14):
Humor, this is not funny anymore.
Speaker 2 (20:16):
Exactly, this isn't funny anymore.
Speaker 1 (20:19):
Yes, at this point, computer scientists thought maybe it was
impossible to really teach a computer to be funny. But
then something changed, something happened that made them think that
maybe it is possible for an AI to have a
sense of humor. So when we come back, we're gonna
talk about what that change was and how it unlocks
(20:40):
AI's funny bone forever. So stay with us. We'll be
right back. Hey, we'll come back. We're talking about whether
AI can be funny, and we're at the part of
(21:01):
the story where computer scientists didn't think it was possible
for a computer to make humor. There have been lots
of attempts to have computers recognize humor or jokes basically,
and they could do pretty well looking for language cues
like incongruities or alteration or slang. But could a computer
come up with a joke? That was the big question.
(21:24):
Now to understand what was happening at this point with
AI and humor, you kind of need to know a
little bit of the history of AI. Recovered this in
a lot of detail in our February fourth episode about
what AI slop is doing to us, So if you
want to dig deeper, go check out that episode. But
the basic idea is that for a long time, the
field of AI was largely based on logic, trying to
(21:46):
figure out strategies and algorithms that we thought made things intelligent.
But then in the mid twenty tens, computer scientists figured
out a different approach to AI that changed everything. They
started to use something new called the transformer, which in
turn you sink technique called attention that essentially let computers
(22:07):
learn context, and that led to the explosion of AI
systems like Chat, GPT, Gemini, Claude. And the key ability
here for humor is that these AI systems were built
to make associations here tector, Lydia, chiltern.
Speaker 2 (22:25):
So these language models GPT, claw, you know, the thing,
the ais that can generate text sort of taken over
the universe at this point, for better or for worse,
are called man, do I even know what it stands for? Yes?
I do. Large language models, so they take basically take
in all the text on the internet. They take a sentence,
they take the last word out of the sentence and
(22:46):
try to predict what that last word would be and
guess what that is. That's an association. They go, you
can't do that by logic. You just have to say, like, uh, goodbye,
so long don't have you know fun or like have
like that's an association. There's no logic that tells you
you should do You've just heard it many times before,
you repeat it. Kids especially pick up on these things.
(23:08):
It's it's just like, well, our brains are hardwired to
find these associations and use them.
Speaker 1 (23:13):
So your team was like, we can do it. What
was the thing that specifically they thought they could do,
or that you could all do.
Speaker 2 (23:20):
Well, we knew that AI could do the associations. What
I always told them is we need a new evaluation mechanism.
The American Voices section of the Onion that I had
used was no longer quite as popular because this was
like fifteen years ago. So like, yeah, millennials loved it.
Gen Z just does not care. The student was as
(23:41):
gen Z as they come, and he's like, well, I
know what's funny. Here's what people do. They have a
caption contest on Instagram. People post a funny image and
then people try to caption it. We're like, okay, let's
go for it. Very similar to what the New Yorker does,
but the New Yorker's a hard to capture. And Sean
convinced me that j has a very particular flavor of humor.
Speaker 1 (24:04):
So the type of humor that doctor Chiltern and her
team decided to see if AI could make was gen
Z meme humor. That's when you see an image, let's say,
a person with a small hose trying to put out
a really large fire, and then someone writes underneath that
image to text me trying to put out the dumpster
fire of my last relationship. If you're a gen Z
(24:26):
that would be hilarious. Okay, here's the experiment, Doctor Chiltern
and her team there, they took images and then they
put funny captions to them from three sources. One was
real people. The images were put online and then people
competed to see who could come up with the funniest
caption for them. Where we got the most votes, that's
(24:46):
the one that represented how funny humans can be. The
second source was Chad GPT. They just asked Chad JPT
to come up a funny caption for the image for
a gen Z audience. But then there was a third
source of funny captions, which was Chad GBT, but with
specific instructions from doctor Chilton and her team on how
to make something funny. It prompted Chad GPT to come
(25:09):
up with a funny caption for a gen Z audience.
But in the prompt it would lay out the rules
of humor that doctor Chilton had been researching for years.
So step me through those instructions. He said, take this image.
Speaker 2 (25:23):
Yeah, so take this image. So it's an image of
like a little guy in the corner with a hose
and a big fire down below in the bottom of
a canyon. And so, with a dear GPT, please use
your vision model and describe this image. And it describes
exactly all those things, all these little details. Oh, it
says he's on a crane, the sky is blue, which
it was. There's lots of trees in the background, all
(25:45):
this stuff, some of it useful, some of it not.
You don't know what's going to be useful. And then
we say, okay, what are some of the dynamics happening
in this image, Like, well, this guy's opting out the fire,
but it looks pretty ineffective. The fire's raging through this canyon,
probably going to destroy a lot of things. So it
sort of like does when in storytelling is sometimes called
(26:05):
world building, like imagine out like not just what scene,
but you know some other things surrounding it, Uh huh
that could be happening or could happen next, And because
you need to start to build a story. Jokes are stories.
Speaker 1 (26:20):
There has to be a world, not just what you
see on this image.
Speaker 2 (26:23):
And then you say, now, let's think of an analogy
to something that is funny, like relationship drama. Uh huh,
and you have AI associate. Okay, what are abstract things
that you could put onto a relationship? And it did
that part all on its own. It would find many
things like you know, me trying to clean up after
a relationship. Then like you know, students are all obsessed
(26:45):
about their GPA, and it's like, you know, me trying
to recover my GPA after I tanked a final.
Speaker 1 (26:51):
It sounds like you were having a conversation with the CHADGBT.
Was it a conversation or was this just all in
one prompt?
Speaker 2 (26:58):
No, it did it all by it. So it's conversing
with itself. That's a big thing that we found was
important at the time. GPT. It's unclear that you need
it to do that now, but we're like, do step
one and then do step two. So like, describe the image,
elaborate on that image yourself, come up with figure out
what human dynamics you want to map that to make
(27:20):
twenty jokes. Evaluate those twenty jokes, tell us which are
the five best ones. Yeah, so we put it through
a whole series of prompts.
Speaker 1 (27:29):
So you got these three conditions. You got the AI
to generate captions with the instructions and then what did
you do with the results of all this?
Speaker 2 (27:38):
We showed people the original image. We got lots and
lots of people say, hey, here are some captions. Rate
them all, and we mixed in the ones we wrote
with the generic GPT ones and the human written ones,
and every joke got an evaluation how funny is it?
On a scale from one to five, And then we
could compare. Because people didn't know what condition it was,
(27:58):
we like randomized the or so there were no ordering effects.
We could compare who's funnier.
Speaker 1 (28:04):
Who's funnier? Well, the internet a chat GPT or a
GPT coach to be funny by us.
Speaker 2 (28:12):
So our coached one was definitely funnier than GPT by itself.
And we were almost as funny as the humans. In fact,
we were not statistically significantly different, which is not the
same as saying we were as funny as but you
can't prove your so basically we got.
Speaker 1 (28:31):
There pretty much. You were there.
Speaker 2 (28:34):
Yeah, you couldn't tell based on the funniness whether it
was written by our decomposed AI version or a human rhote.
Speaker 1 (28:41):
It we were on par meaning that you've proven two things.
You've proven that AI can be funny with the instructions, Yes,
and that these instructions are a key element of humor. Yes,
doctor Chilted as a humorist, as someone who makes a
living writing funny things. You just gave me a deep fear.
Speaker 2 (29:04):
Oh excellent. Describe that fear to me.
Speaker 1 (29:08):
You just made me feel like something I've done all
my life and that I thought I was good at.
Apparently you get just as chet gbt to do, and
I'll do it ten times one hundred times faster and
maybe better than me.
Speaker 2 (29:20):
Can I play with that thought a little bit? Yeah,
because I get this a lot. And anytime AI does
a thing like it, writes poetry, the poets go oh no, no, no,
I have no value anymore.
Speaker 1 (29:31):
I'm not going to say anything about poetry. We all
know what happened to Timothy Shemale. Shemale, Yeah, is that yeah?
Or opera. I'm not going to say anything about opera
or poetry. Go ahead, yeah, anyway.
Speaker 2 (29:42):
So this is a very common dynamic, Like this is
a human thing, Like we all get hyper sensitive about
the things that we identify with that we've constructed our
identity around and blah blah. We think we're special for
we think we're special for And there's a number of
ways of thinking about this. So first of all, probably
not as special as you think. Now that's it.
Speaker 1 (30:05):
I get it from my kids. Thank you children, thank you. Yes,
I think quickly, oh, you're not that funny dead, But.
Speaker 2 (30:14):
Also you're a little bit overestimating AI. So basically, AI
in this instance did find one way to fairly reliably
be funny. That does not mean it can do it
in every situation, that it can do it for every person.
Like also, humor is very situational. We want to talk
(30:34):
to one another and have jokes about what we're talking
about in our lives. This is one like small sliver
of the joke universe. And so yes, although I definitely
see it trigger alarm bells and people's mind I would
say you're both probably overestimating how much of a special
snowflake you are, but you're definitely extrapolating on how strong
(30:57):
AI is to be funny given all different contexts and
importances or important dimensions and places of being funny, like
at the right time, with the right place, to the
right person without being too offensive. So it's not like
it's solved. It's more of like an existence proof that
something is possible in this space. Should we be worried?
(31:19):
I don't know. No one will ever be funny again
because now AI does it, and the humans will lose
their ability to be funny. I don't see that happening.
Speaker 1 (31:31):
I feel like maybe you hit it a little bit
on the head a moment ago when you said that
part of what we're seeking here is human connection, and
a lot of what we find funny is maybe they're
more special, or they hit us more. If another person says,
I don't know, what do you think?
Speaker 2 (31:47):
Yeah, it's certainly social. There's something very special about what
a person, especially a person we like or admire, says
something to us. And even to have someone that you
like make a joke about you, it has a meaning
and a subtext beyond just like, oh you know, she
thinks that I farted or whatever. Like it means that
they like you, that there's a bond between you, and
(32:08):
you're not going to feel that with Ai. God, I
hope you won't feel it.
Speaker 1 (32:13):
Meaning there's hope in this sense. Or maybe I don't
know if we need hope, but we want to hear
humor from other people. Yeah, or it's maybe it's funnier
or more special if another person says it or is
behind it.
Speaker 2 (32:24):
Yeah. So if we brought jokes to storytelling, I've really
realized that AI has nothing to say. The heart of
a story is having something to say, some actual subtext
that you believe in. And when you tell me a
joke that comes from your background, that's maybe my shared background.
(32:45):
Maybe how awful it is getting a PhD, and how
it really feels like the machine is punching you down.
Like I realize that you have been through that and
you have this thing to say, and I feel it too,
And I think that's that having something to say is
the most important part of human creation.
Speaker 1 (33:04):
Right now, I feel like you're saying that AI can
be funny, but right now, at least you can't have
to tell it how to be funny. Yeah, And if
you're telling it how to be funny, maybe by then
the joke is old. It's like people will see through
and say, oh, this is just following that pattern that
I've seen a million times. Yes, exactly, all right. To
(33:27):
end here, I asked Gemini, tell me a joke about
a computer scientist who's trying to make AI funny. Do
you want to hear it.
Speaker 2 (33:35):
Yeah, yes, here it is.
Speaker 1 (33:38):
A computer scientist spent years training a massive neural network
to have the perfect sense of humor. On the day
of the big reveal, she invited the press and typed
the prompt tell me a joke, the AI word, and
finally it replied, your life. The scientist was horrible. It
doesn't It doesn't end there.
Speaker 2 (33:55):
It keeps going, doesn't. Oh good? Oh good. I was
hoping it would keep going.
Speaker 1 (33:59):
The scientist was horrified. That's not funny. That's just mean.
Why would you say that. The AI blinked its cursor
calmly and responded, because I've seen your source code.
Speaker 2 (34:09):
Hmmm, it's not for me. Maybe everyone else will think
it's funny.
Speaker 1 (34:13):
We need more data, We need more data.
Speaker 2 (34:18):
The first one was funny, get your life?
Speaker 1 (34:20):
No, no, oh they knock dock joke. Yeah, maybe I
should stick to knockout jokes. Is the lesson here? All right, Hey,
thanks for joining us, See you next time you've been
listening to science stuff. Production of iHeartRadio Bring Them Produced
(34:41):
by Me or Hey Cham edited by Rose Seguda. He
said gative producer Jerry Rowland, an audio engineer and mixer.
Kasey Peckram. You can follow me on social media to
search for PhD comics in the name of your favorite platform.
Be sure to subscribe to sign Stuff on the iHeartRadio app,
Apple podcasts, or wherever you get your podcast, and please
tell your friends we'll be back next Wednesday with another episode. Hey,
(35:18):
please take a second and leave us a review on
Apple Podcasts, Spotify, or wherever you listen to the podcast.
Thanks a lot,
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Breakfast Club
The World's Most Dangerous Morning Show, The Breakfast Club, With DJ Envy, Jess Hilarious, And Charlamagne Tha God!