Back in January, a lot of people were infuriated by the very idea of a posthumous, AI-generated George Carlin special—at least until it turned out to be written by a human. Jeff Ganim was not one of those people. “I want to see an AI Jerry Seinfeld cracking jokes in a hundred years about the towels in a hotel on the moon,” the AI prompt engineer says.
Ganim’s recent creation, PFFT, named after one of the sounds people’s lips allegedly make when they find something funny, uses an AI model to dispense The Onion-style headlines based on user prompts. It’s the latest in a growing number of tools that rely on machine-learning to generate jokes, including standup chatbot AI Comedian; Punchlines.ai, which provides the kicker to user’s setups, and AI Comedy Skit, which generates scripts for decidedly non-SNL caliber sketches out of user-provided premises. But are any of these digital vaudevillians actually pushing the creativity envelope, or are they turning a vital artform into a satire of humanity?
PFFT has a no-frills interface with a bright orange news ticker at the bottom, rattling off previous, humor-adjacent headlines. (“Unicorn dies of glitter lung,” “Study finds every type of parenting produces miserable adults,” etc.) A prompt-field dead center solicits users to enter an opinion. Type one in—say, “Spring is an underrated season”—hit the “Pfft” button, and after a brief, buffering throat-clearance, the machine confidently offers a handful of zingers. (Cherry Blossoms’ Brief, Fleeting Beauty Completely Wasted On Fallible, Stupid Humans, etc.) Users can then keep repeating this process until a joke riffing on that initial opinion produces sufficient guffaws.
Ganim approached the task of building PFFT as both a creative and a comedy lover. The AI prompt engineer, who has a background in digital marketing, provided large language models with ample doses of humor to build the comedy tool. (He would not specify which sources of humor he used.) Ganim was amazed initially by the creative potential of OpenAI’s early models, such as davinci-002—the AI that reportedly freaked out comedy writer Simon Rich last year with its joke-generating abilities. Ganim felt somewhat stifled by the newer models, which now have more guardrails against overtly offensive output, but pressed on anyway.
Like a road comic tailoring material to different cities, AI could potentially mimic any number of comedic modes. Ganim chose satirical news headlines because AI is especially adept at pattern-recognition and imitation, which he thought made it well-suited for this kind of parody. He describes PFFT as a way to democratize the power of satire for anyone who doesn’t feel served by The Onion or The Babylon Bee. However, he also hopes it will inspire other creatives to work with AI and build new kinds of art together. Even comedians.
“Comedy writers talk about how difficult comedy writing actually is,” Ganim says. “Many describe it as, counterintuitively, quite a painful process. I think tools like this can help take away some of the drudgery of comedy writing so comedians can focus on the parts of comedy that they really love.”
Curious about how well this tool performed the drudgery of comedy writing, I put it to the test.
“Black labs are the best dogs,” I wrote into the blinking field hungry for opinions. After a brief wait, the machine shares its work. The vast majority of these headlines are hot garbage. “Black Lab Kills Owner in Cold Blood,” for instance, is essentially the exact opposite of the prompt I had just entered. Some of the entries, however, stood out in positive ways.
Black Lab Has No Time For Your Bull*hit, Wants To Play Ball, hinges on an amusingly familiar situation. Unfortunately, this joke could apply to most dog breeds. It’s cute, but not specific enough. It also telegraphed another issue that became more pronounced in later searches; several headlines rely on an unlikely curse word to do the heavy lifting of making a joke work.
Racist Black Lab Refuses to Chase Own Tail is another that caught my attention, but mainly because I hadn’t yet considered how and whether race would factor into PFFT’s jokes. (Ganim says his team needs to use OpenAI’s moderation filter to comply with the terms of service of the model, preventing it from landing in truly tasteless territory.)
A caption on PFFT’s main page suggests experimenting with detail. I flailed around for the type of opinion I might have considered tweeting. Here’s what I came up with: “When a server at a restaurant asks if you need another minute with the menu, and you say yes, they will always give you way too many minutes.”
One of the responses skews close to the tried and true Onion format of framing mundane moments as major news: Waitress Gives You 50 More Minutes to Decide. Another takes a different tact with the same core joke: Server’s ‘More Time with the Menu’ Offer Turns Out to Be for the Rest of the Day. Neither so much as elicits a chuckle, but they do make me smile.
PFFT seems to have a near-competent command of some Onion styles, and at least flicks in the direction of being funny. But is it actually funny? I decided to ask a professional comedian.
Django Gold is a former writer for The Onion and The Late Show with Stephen Colbert, who has joked about AI in his recent standup special, Bag of Tricks, which is free to watch on YouTube. Although he has not taken in the full range of comedic AI offerings currently available, Gold disputes the idea that machines can learn what makes jokes funny—let alone make funny jokes.
“You just know when a joke works. And that knowledge derives from an impossible number of inputs,” he says. “It comes from your own personal experiences, it comes from the audience’s experiences, and it also comes from how you interpret the overlap between them. It comes from our shared culture, and it also comes from the way words sound and how their meaning subtly changes when combined in different ways, and—when spoken out loud—the vocal inflection used to deliver them. And, very often, it comes from the identity of the joke teller and our related expectations.”
When a joke comes from a comedian with a well-defined persona, for instance, the audience filters each joke through all the other jokes they’ve ever heard from that comedian, along with everything else they know about her or him. (See: Mulaney, John.)
Conversely, AI-generated jokes may instead suffer from the fact that most audiences will experience them in the context of everything they know and feel about AI.
According to Gold, though, the problem is more that the material is not up to par.
“PFFT maybe can produce joke elements, little idea-pieces that the machine has correctly identified as occasionally being contained in a joke,” he says. “But the machine can’t make them cohere in a meaningful way; it’s just throwing them together willy-nilly and presenting this collection of elements as a joke.”
Since he hadn’t seen any of PFFT’s jokes yet at this point, I showed Gold some examples from the headlines it created from my prompts, starting with the one about the racist black lab.
“This is technically a joke,” he says. “If a person said this to me, I would smile politely.”
Next, I offered one PFFT made based on my opinion about how oranges aren’t worth the amount of effort it takes to peel them: “Oranges Just a Vibrant, Healthy Way to Get Your Hands Sticky.”
We both agreed that the use of the word “vibrant” seemed almost violently unnatural in this context. I thought the joke was basically okay, though, while Gold felt otherwise.
“This isn’t good, and worst of all it’s not good in a very boring way,” he says. “Oranges are sticky, that’s the ‘nut’ of the joke here. Not very interesting. How about: Man Takes Break from Peeling Orange to Put in Contact Lenses? Not exactly groundbreaking, but better. More human.”
Curious about what PFFT would have to say about Gold’s disses, I entered a prompt about a comedy AI making mediocre jokes. It responded with the headline, Comedy AI: 404 Not Funny, which is probably the funniest thing it came up with in the hours I spent noodling with it.
One area where the comedian and the comedy AI tool sync up somewhat is in the idea of comedy as a numbers game. PFFT boasts that while professional comedians may get one in 20 funny bits, it averages one in 10. (Ganim says this statistic is a rough estimate.) Meanwhile, Gold often has to sift through a lot of meh iterations of a joke before getting to a winner. One critical difference between how man and machine approach producing jokes in bulk, though, in Gold’s view, is that machines can’t distinguish between good and bad jokes.
“When writing a bunch of monologue jokes, for example, the idea would be to write 30 and get 10 good ones out of that. And if you wrote 60, those 10 good ones would be stronger than the 10-from-30 batch,” he says. “But one key element of this process is that you learn from the bad jokes and figure out how to improve them. A machine can’t teach itself to get closer to a good joke; it just keeps generating material without making adjustments, forever.”
It’s unclear if PFFT could ever teach itself to be funnier, but its creator would like his machine to do more.
Ganim wants to see PFFT eventually evolve into a new kind of social media site, one where everyone can post their own satirical news based on in-jokes with their friends. He also sees such a place becoming a hub where aspiring comedians can post the material they come up with, start building an audience, and maybe even get discovered. He’d also like to work with heavy-hitter marquee comedians and—with their permission, of course—train custom AI models just for them, turning something like that fake George Carlin AI special into a reality.
“People don’t think it’s possible to do this in a way that does justice to their creative minds,” he says, “but I think these people are giving up too soon.”
For Gold, on the other hand, humanity can’t throw in the AI comedy towel fast enough.
“I don’t think AI will ever produce comedy that anyone enjoys or has demand for. If it does, I’ll eat my hat,” he says, adding: “And it’s a large, brittle hat.”
Back in January, a lot of people were infuriated by the very idea of a posthumous, AI-generated George Carlin special—at least until it turned out to be written by a human. Jeff Ganim was not one of those people. “I want to see an AI Jerry Seinfeld cracking jokes in a hundred years about the towels in a hotel on the moon,” the AI prompt engineer says.
Ganim’s recent creation, PFFT, named after one of the sounds people’s lips allegedly make when they find something funny, uses an AI model to dispense The Onion-style headlines based on user prompts. It’s the latest in a growing number of tools that rely on machine-learning to generate jokes, including standup chatbot AI Comedian; Punchlines.ai, which provides the kicker to user’s setups, and AI Comedy Skit, which generates scripts for decidedly non-SNL caliber sketches out of user-provided premises. But are any of these digital vaudevillians actually pushing the creativity envelope, or are they turning a vital artform into a satire of humanity?
PFFT has a no-frills interface with a bright orange news ticker at the bottom, rattling off previous, humor-adjacent headlines. (“Unicorn dies of glitter lung,” “Study finds every type of parenting produces miserable adults,” etc.) A prompt-field dead center solicits users to enter an opinion. Type one in—say, “Spring is an underrated season”—hit the “Pfft” button, and after a brief, buffering throat-clearance, the machine confidently offers a handful of zingers. (Cherry Blossoms’ Brief, Fleeting Beauty Completely Wasted On Fallible, Stupid Humans, etc.) Users can then keep repeating this process until a joke riffing on that initial opinion produces sufficient guffaws.
Ganim approached the task of building PFFT as both a creative and a comedy lover. The AI prompt engineer, who has a background in digital marketing, provided large language models with ample doses of humor to build the comedy tool. (He would not specify which sources of humor he used.) Ganim was amazed initially by the creative potential of OpenAI’s early models, such as davinci-002—the AI that reportedly freaked out comedy writer Simon Rich last year with its joke-generating abilities. Ganim felt somewhat stifled by the newer models, which now have more guardrails against overtly offensive output, but pressed on anyway.
Like a road comic tailoring material to different cities, AI could potentially mimic any number of comedic modes. Ganim chose satirical news headlines because AI is especially adept at pattern-recognition and imitation, which he thought made it well-suited for this kind of parody. He describes PFFT as a way to democratize the power of satire for anyone who doesn’t feel served by The Onion or The Babylon Bee. However, he also hopes it will inspire other creatives to work with AI and build new kinds of art together. Even comedians.
“Comedy writers talk about how difficult comedy writing actually is,” Ganim says. “Many describe it as, counterintuitively, quite a painful process. I think tools like this can help take away some of the drudgery of comedy writing so comedians can focus on the parts of comedy that they really love.”
Curious about how well this tool performed the drudgery of comedy writing, I put it to the test.
“Black labs are the best dogs,” I wrote into the blinking field hungry for opinions. After a brief wait, the machine shares its work. The vast majority of these headlines are hot garbage. “Black Lab Kills Owner in Cold Blood,” for instance, is essentially the exact opposite of the prompt I had just entered. Some of the entries, however, stood out in positive ways.
Black Lab Has No Time For Your Bull*hit, Wants To Play Ball, hinges on an amusingly familiar situation. Unfortunately, this joke could apply to most dog breeds. It’s cute, but not specific enough. It also telegraphed another issue that became more pronounced in later searches; several headlines rely on an unlikely curse word to do the heavy lifting of making a joke work.
Racist Black Lab Refuses to Chase Own Tail is another that caught my attention, but mainly because I hadn’t yet considered how and whether race would factor into PFFT’s jokes. (Ganim says his team needs to use OpenAI’s moderation filter to comply with the terms of service of the model, preventing it from landing in truly tasteless territory.)
A caption on PFFT’s main page suggests experimenting with detail. I flailed around for the type of opinion I might have considered tweeting. Here’s what I came up with: “When a server at a restaurant asks if you need another minute with the menu, and you say yes, they will always give you way too many minutes.”
One of the responses skews close to the tried and true Onion format of framing mundane moments as major news: Waitress Gives You 50 More Minutes to Decide. Another takes a different tact with the same core joke: Server’s ‘More Time with the Menu’ Offer Turns Out to Be for the Rest of the Day. Neither so much as elicits a chuckle, but they do make me smile.
PFFT seems to have a near-competent command of some Onion styles, and at least flicks in the direction of being funny. But is it actually funny? I decided to ask a professional comedian.
Django Gold is a former writer for The Onion and The Late Show with Stephen Colbert, who has joked about AI in his recent standup special, Bag of Tricks, which is free to watch on YouTube. Although he has not taken in the full range of comedic AI offerings currently available, Gold disputes the idea that machines can learn what makes jokes funny—let alone make funny jokes.
“You just know when a joke works. And that knowledge derives from an impossible number of inputs,” he says. “It comes from your own personal experiences, it comes from the audience’s experiences, and it also comes from how you interpret the overlap between them. It comes from our shared culture, and it also comes from the way words sound and how their meaning subtly changes when combined in different ways, and—when spoken out loud—the vocal inflection used to deliver them. And, very often, it comes from the identity of the joke teller and our related expectations.”
When a joke comes from a comedian with a well-defined persona, for instance, the audience filters each joke through all the other jokes they’ve ever heard from that comedian, along with everything else they know about her or him. (See: Mulaney, John.)
Conversely, AI-generated jokes may instead suffer from the fact that most audiences will experience them in the context of everything they know and feel about AI.
According to Gold, though, the problem is more that the material is not up to par.
“PFFT maybe can produce joke elements, little idea-pieces that the machine has correctly identified as occasionally being contained in a joke,” he says. “But the machine can’t make them cohere in a meaningful way; it’s just throwing them together willy-nilly and presenting this collection of elements as a joke.”
Since he hadn’t seen any of PFFT’s jokes yet at this point, I showed Gold some examples from the headlines it created from my prompts, starting with the one about the racist black lab.
“This is technically a joke,” he says. “If a person said this to me, I would smile politely.”
Next, I offered one PFFT made based on my opinion about how oranges aren’t worth the amount of effort it takes to peel them: “Oranges Just a Vibrant, Healthy Way to Get Your Hands Sticky.”
We both agreed that the use of the word “vibrant” seemed almost violently unnatural in this context. I thought the joke was basically okay, though, while Gold felt otherwise.
“This isn’t good, and worst of all it’s not good in a very boring way,” he says. “Oranges are sticky, that’s the ‘nut’ of the joke here. Not very interesting. How about: Man Takes Break from Peeling Orange to Put in Contact Lenses? Not exactly groundbreaking, but better. More human.”
Curious about what PFFT would have to say about Gold’s disses, I entered a prompt about a comedy AI making mediocre jokes. It responded with the headline, Comedy AI: 404 Not Funny, which is probably the funniest thing it came up with in the hours I spent noodling with it.
One area where the comedian and the comedy AI tool sync up somewhat is in the idea of comedy as a numbers game. PFFT boasts that while professional comedians may get one in 20 funny bits, it averages one in 10. (Ganim says this statistic is a rough estimate.) Meanwhile, Gold often has to sift through a lot of meh iterations of a joke before getting to a winner. One critical difference between how man and machine approach producing jokes in bulk, though, in Gold’s view, is that machines can’t distinguish between good and bad jokes.
“When writing a bunch of monologue jokes, for example, the idea would be to write 30 and get 10 good ones out of that. And if you wrote 60, those 10 good ones would be stronger than the 10-from-30 batch,” he says. “But one key element of this process is that you learn from the bad jokes and figure out how to improve them. A machine can’t teach itself to get closer to a good joke; it just keeps generating material without making adjustments, forever.”
It’s unclear if PFFT could ever teach itself to be funnier, but its creator would like his machine to do more.
Ganim wants to see PFFT eventually evolve into a new kind of social media site, one where everyone can post their own satirical news based on in-jokes with their friends. He also sees such a place becoming a hub where aspiring comedians can post the material they come up with, start building an audience, and maybe even get discovered. He’d also like to work with heavy-hitter marquee comedians and—with their permission, of course—train custom AI models just for them, turning something like that fake George Carlin AI special into a reality.
“People don’t think it’s possible to do this in a way that does justice to their creative minds,” he says, “but I think these people are giving up too soon.”
For Gold, on the other hand, humanity can’t throw in the AI comedy towel fast enough.
“I don’t think AI will ever produce comedy that anyone enjoys or has demand for. If it does, I’ll eat my hat,” he says, adding: “And it’s a large, brittle hat.”