Humor

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News


Computational Humor is the branch of artificial intelligence and Natural Language Processing (NLP) that studies how machines can recognize, explain, generate, and perform humor. It sits at the crossroads of linguistics, cognitive psychology, and computer science, and it has always been treated as a hard test of machine intelligence, because a good joke depends on nearly everything that is difficult for AI: shared context, world knowledge, double meanings, social judgment, and timing.

The field is older than chatbots. Linguist Victor Raskin's Semantic Script Theory of Humor (1985), extended with Salvatore Attardo into the General Theory of Verbal Humor (1991), gave researchers the first formal model of what a joke is. Kim Binsted and Graeme Ritchie built JAPE, a program that generated punning riddles, in 1994; Ritchie's team in Scotland later turned that work into STANDUP, a pun generator for children who use speech aids. Rada Mihalcea and Carlo Strapparava trained classifiers to recognize one-liners in 2005. Peter McGraw and Caleb Warren proposed Benign Violation Theory in 2010, and Tony Veale's book Your Wit Is My Command (2021) laid out a computational roadmap for witty machines. Practitioners joined in too: former late-night writer Joe Toplyn patented Witscript, and researcher-actors Piotr Mirowski and Kory Mathewson put AI on stage with the improv troupe Improbotics.

Large language models (LLMs) changed the field almost overnight. In 2022 Google's PaLM (Pathways Language Model) explained novel jokes well enough to startle Geoffrey Hinton, and by 2023 ChatGPT could produce a passable joke on demand. Older systems relied on hand-built templates and knowledge bases; LLMs learned joke patterns implicitly from enormous amounts of text. That made them fluent, flexible, and fast. It also made them prone to new failure modes: recycling the same stock jokes, defaulting to bland "safe" material, and producing confident explanations for jokes that are not actually funny. The current research question is no longer "can a machine produce something joke-shaped?" but "can it be original, fit the context, land with a real audience, and know why?"

Ultimately, the pursuit of AI-generated humor not only has practical applications in entertainment and creative industries but also serves as a litmus test for AI's ability to understand and generate content that resonates with human sensibilities and emotions.

Humor often involves the realization of incongruity (mismatch) between a concept and a situation, violations of social taboos or expectations, the resolution of tension or mocking and a sense of superiority. The mystery of what makes a joke funny – but only to some people | Sophie Scott - The Conversation


Degree we find something funny = the degree to which a joke’s punch line deviates from the listener’s unconscious expectation, provided the surprise then resolves into a second meaning that makes sense. Surprise without resolution is just noise.


There are many theories about what makes a joke funny. According to social psychologist Alex Borgella, a joke is funny for one or more of the following reasons: Superiority, Relief, and Incongruity. The Benign Violation Theory (BVT), developed by Peter McGraw and Caleb Warren at the University of Colorado Boulder's Humor Research Lab, theorizes that a joke is funny when it meets three necessary conditions. First, there must be some sort of norm violation, be it a moral norm violation (robbing a retirement home), social norm violation (breaking up with a long-term boyfriend via text message), or physical norm violation (purposefully sneezing directly on a child). Second, the situation must also seem benign, meaning safe, acceptable, or not really threatening. Third, both perceptions must happen at the same time. How to Write a Joke: A Full Guide to the Perfect Set-Up & Punchline | Suzy Woltmann - Backstage That third condition explains a lot about AI comedy: a model tuned to avoid anything that feels like a violation will rarely be funny, while a model with no sense of what is benign will be offensive.

How AI Humor Works

Building a "sense of humor" into software breaks down into five distinct tasks. Systems can be good at one and poor at another.

1. Recognition (is this a joke?). Classifiers, from Mihalcea and Strapparava's 2005 one-liner detector to transformer models, learn surface cues of humor: ambiguity, incongruous word pairings, alliteration, and setup-punchline structure. The weakness is that these cues can be faked. A sentence shaped like a joke ("Old X never die, they just Y") fools many models even when it is meaningless.

2. Explanation (why is it funny?). A model has to identify the two competing interpretations (the "scripts") and the trigger that switches between them. LLMs became strong at this around 2022. Benchmarks show they still lag people, and they also produce plausible explanations for jokes that do not work, which suggests pattern matching rather than full understanding.

3. Generation (write a joke). There are three main approaches:

  • Template and knowledge-based systems (JAPE, STANDUP, DEviaNT) fill structured patterns from lexical databases. Output is reliable but narrow.
  • Neural-symbolic hybrids (Witscript, HumorTools-style workflows) encode the steps professional writers use: pick "handles" in a topic, generate associations, link them into a punch line, and choose the best candidate. Each step can call an LLM.
  • End-to-end LLM generation simply prompts a model. It is flexible but tends toward the most probable, and therefore least surprising, completions unless it is steered.

4. Evaluation (which joke is best?). Systems generate many candidates and rank them using a human panel, a trained reward model, or an LLM acting as judge. Evaluation is the bottleneck of the whole field, because humor ratings vary widely between people and even between moods.

5. Delivery (performance and timing). For spoken or robotic comedy, systems must detect laughter, decide how long to pause, and adapt follow-up lines to the room, using Speech Recognition and speech synthesis.

Dependencies and conditions. Every stage depends on context: who the audience is, what has already been said, what is culturally acceptable, and what both sides know. Humor that works in one community misfires in another, which is why humor is also a stress test for Out-of-Distribution (OOD) Generalization.

Where uncertainty remains. Whether LLM "understanding" of humor reflects genuine representation of competing meanings or sophisticated pattern completion is an open debate, and results depend heavily on how evaluation is designed.

Surprise, Ambiguity, and the Information Theory of a Punch Line

Jokes have a measurable structure. Language models assign a probability to every next word, which turns the old intuition "a punch line is a surprise" into numbers.

1. Surprisal. The surprisal of a word w given its context is S(w) = −log₂ P(w | context), measured in bits. A continuation the model expects with probability 0.5 carries 1 bit of surprise; a punch line it gives probability 0.001 carries about 10 bits. Yubo Xie, Junze Li, and Pearl Pu (EPFL, ACL 2021) fed setups and punch lines to GPT-2 and found that two features, the uncertainty (entropy) the setup builds and the surprisal of the punch line, separate jokes from non-jokes better than earlier baselines. The setup opens possibilities; the punch line picks an unlikely one.

2. Ambiguity plus distinctiveness. Surprise alone is not enough; random words are surprising but not funny. Justine Kao, Roger Levy, and Noah Goodman (Cognitive Science, 2016) modeled pun comprehension as probabilistic inference over two meanings. Their ambiguity measure (entropy over which meaning is intended) distinguished puns from non-puns, and their distinctiveness measure (whether each meaning is supported by different words in the sentence) separated exceptionally funny puns from mediocre ones. In plain terms: a good pun keeps two meanings alive, and each meaning has its own supporting evidence.

3. Local versus global surprise. He He, Nanyun Peng, and Percy Liang (Stanford, NAACL 2019) formalized a "local-global surprisal principle" for puns: the pun word should be surprising in its immediate context but supported by the wider sentence. They used the ratio Slocal / Sglobal as a funniness signal, and their retrieve-and-edit generator produced successful puns about 31% of the time, roughly triple a neural baseline. This is the incongruity-resolution theory written as an equation.

4. The inverted U. Put these together and humor lives in a narrow band: too predictable is a groan, too unpredictable is nonsense. Funniness peaks where surprise is high and a resolving second meaning exists. This is the precise sense in which the thesis callout above holds.

5. Why LLMs default to stock jokes. Training maximizes the likelihood of typical text, and standard decoding favors high-probability continuations, the opposite of what a punch line needs. Preference tuning can narrow output further: Jentzsch and Kersting found that over 90% of 1,008 ChatGPT jokes were the same 25 jokes, and Zhang and colleagues (NeurIPS 2024) reported that standard alignment methods such as RLHF and DPO showed clear limits on creative caption writing. Raising the sampling temperature T (probabilities proportional to exp(z/T)) increases variety, but also incoherence, so it is a blunt tool.

6. Tokenization and sound. Puns and rhymes depend on pronunciation, while most LLMs read subword tokens that hide phonetics. This is why Witscript generates its wordplay candidates with classic NLP tools instead of GPT-3 (see below), and why sound-based humor remains comparatively weak.

7. Timing as a physical variable. Delivery happens in seconds. In Naomi Fitter's robot comedy experiments, a robot that always waited a fixed five seconds after each joke was rated less funny than one that paused for laughter and resumed as it faded.

8. The brain's two-stage response. Electrophysiology tracks the same stages. The N400 brain wave (about 400 milliseconds after a word) indexes detection of semantic incongruity, and the Late Positive Potential (around one second later) indexes the amusement of "getting it." A 2025 study from the Chinese University of Hong Kong found that puns attributed to an AI produced a smaller N400 but a larger LPP than identical puns attributed to a person (see Recent Developments).

9. The measurement ceiling. People disagree about what is funny, so any predictor's accuracy is capped by inter-rater agreement. In a 2026 study of joke originality, a composite metric correlated with comedians' judgments at only ρ = 0.40, yet that captured an estimated 82% of the variance that was predictable at all given how much the three comedians agreed with each other.

Why This Matters

Scientific importance. Humor is a compact test of context, common sense, theory of mind, and creativity. A system that reliably understands why something is funny must represent multiple meanings, model what the listener expects, and judge social acceptability at once. That makes humor benchmarks useful probes of whether models genuinely understand language or exploit surface regularities.

Engineering implications. Humor makes conversational systems feel more natural, helps repair awkward moments, and can build rapport in tutoring, healthcare, customer service, and companion apps. Sarcasm and irony detection matter for content moderation, sentiment analysis, and assistive tools, including support for people who find figurative language hard to read.

Practical consequences. Writers and comedians already use LLMs as brainstorming partners. Blind rating studies suggest that for short, structured formats such as funny acronyms or satirical headlines, AI output can match amateurs and sometimes professionals, which raises real economic questions for comedy writers and content producers. In live, personal, and edgy formats, professionals still find the tools weak.

Ethical and legal stakes. Humor is where AI safety and creative freedom collide. Filters that block hateful content also block satire that "punches up" at the powerful. Synthetic impersonations of real comedians raise copyright and likeness issues, as the 2024 George Carlin case showed. And because humor builds trust, it can also be used to make AI systems more persuasive than they deserve to be.

If the optimists are right, future assistants will generate the "tailored humor" anticipated below, adapting jokes to an individual's taste in real time. What remains uncertain is whether models can be genuinely original rather than recombining familiar material, whether "funny on average" survives contact with specific audiences, and whether audiences will value jokes that no one actually lived through.

Recent Developments

Blind ratings: AI jokes can match human ones in short formats

  • USC study (PLOS ONE, July 2024). Drew Gorenz and Norbert Schwarz gave ChatGPT 3.5 and 105 laypeople the same three comedic tasks (funny acronyms, fill-in-the-blank, and roast-style jokes). In ratings by a separate group of about 200 people who did not know the source, 69.5% of participants preferred the AI responses on average. In a second study, ChatGPT wrote headlines in the style of The Onion, and 217 USC students rated them as about as funny as published Onion headlines; the four top-rated headlines split two and two. What it shows: in short, prompt-structured tasks, LLM output can rival amateurs and sometimes professionals. What it does not show: that AI can originate premises, perform, or sustain a set; the prompts did much of the creative framing, and an unpressured student panel is not a comedy club.
  • Witscript "laugh-off" (CHum 2025). Joe Toplyn and Ori Amir measured actual audience laughter. Toplyn and Witscript each wrote jokes on eight topics; a professional comedian delivered half human and half AI material to live audiences in North Hollywood, and the AI jokes drew as much laughter as Toplyn's. Caveats: a single comedian, a small number of shows, and jokes curated by the system's creator.
  • Perception and bias. Earlier work by Bower and Steyvers (Scientific Reports, 2021) found people rated jokes lower when they believed a machine wrote them, though the penalty disappeared when sources were disclosed honestly. A 2025 study by Joshi reported that participants struggled to tell AI from human jokes and liked AI jokes more with repeated exposure. An EEG study by Xiaohui Rao, Hanlin Wu, and Zhenguang Cai (Chinese University of Hong Kong, 2025; 57 participants, Mandarin puns) found equal funniness ratings, but puns attributed to AI produced a smaller N400 (less effort detecting incongruity) and a larger Late Positive Potential (more amusement), and the gap grew over the session. The authors interpret this as low expectations for AI being pleasantly violated, a benign violation in its own right. This is one lab study with a Wizard-of-Oz design and needs replication.

Benchmarks: understanding still lags

  • "Do Androids Laugh at Electric Sheep?" (Cornell, Allen Institute for AI, and others; ACL 2023 Best Paper). Jack Hessel and colleagues, with former New Yorker cartoon editor Robert Mankoff, built three tasks from more than 700 New Yorker caption contests. The best model matched captions to cartoons with 62% accuracy versus 94% for humans, and human-written explanations beat few-shot GPT-4's in more than two-thirds of head-to-head comparisons.
  • Humor preference data at scale (NeurIPS 2024). Jifan Zhang, Robert Nowak, Mankoff, and colleagues released over 250 million human votes on more than 2.2 million captions from eight years of the contest. GPT-4 and Claude produced humorous captions but underperformed top human contestants, and common fine-tuning methods (RLHF, DPO) showed clear limits on this creative task.
  • "Pun Unintended" (Cardiff University and Ca' Foscari University of Venice; EMNLP 2025). Alessandro Zangari, Jose Camacho-Collados, and colleagues rewrote standard pun benchmarks with small changes, creating the PunnyPattern and PunBreak sets. Subtle edits misled recent LLMs, which insisted that pun-shaped but meaningless sentences were funny. Interpretation: high scores on older pun-detection datasets partly reflect memorization and shortcut cues.

Professional comedians: useful for volume, weak on voice

Google DeepMind researchers Piotr Mirowski, Juliette Love, Kory Mathewson, and Shakir Mohamed ran three-hour "AI x Comedy" workshops with 20 professional comedians at the August 2023 Edinburgh Festival Fringe and online (published at ACM FAccT 2024). Comedians used LLMs to brainstorm and draft, then completed the standard Creativity Support Index and a focus group. The tools scored a middling 54.6 out of 100. Most participants found output bland and stereotyped; one described it as "cruise ship comedy material from the 1950s, but a bit less racist." Safety filters blocked dark material and satire that targets the powerful, and comedians raised concerns about copyright and about global "value alignment" that ignores community norms. The authors argue for community-based alignment and data ownership for artists.

New directions in generation and evaluation

  • Leaning into being a machine. A February 2026 study ("Not Human, Funnier") built an AI stand-up agent that joked from its own identity as an AI, the way human comics mine their own backgrounds. In tests with 32 audience members, it was rated funnier than a baseline GPT agent. Small sample, but a useful design idea: authentic perspective may matter more than imitation.
  • Measuring originality. At the 2nd Workshop on Computational Humor (CHum 2026), Ori Amir, Huyen Ngo, Joe Toplyn, and Kevin Hickerson scored the originality of 400 jokes from GPT-4o and GPT-5.4 plus 80 Witscript jokes by comparing their "topic handles" against a million Reddit jokes, using pointwise mutual information and embedding distance. Their composite tracked three professional comedians' originality ratings at ρ = 0.40. Automated originality metrics could give the field a cheaper alternative to human panels, and a way to detect recycled jokes.
  • A dedicated research venue. The first ACL-affiliated Workshop on Computational Humor (CHum) ran at COLING 2025 in January 2025, organized by Christian Hempelmann, Julia Rayz, Tiansi Dong, and Tristan Miller; a second edition followed at ACL 2026 in July 2026.

Sarcasm, laughter, and timing

  • University of Groningen researchers Xiyuan Gao, Shekhar Nayak, and Matt Coler (2024) trained a multimodal sarcasm detector on audio, transcripts, and emotion labels from sitcom clips in the MUStARD dataset (about four hours of speech from shows like Friends and The Big Bang Theory). It identified sarcasm about 74 to 75% of the time. They also found that sarcastic pitch patterns differ across languages, so detectors may need per-language training.
  • Kyoto University's shared-laughter model for the android Erica (2022) and Oregon State's Jon the Robot studies remain the main evidence that timing and laughter-aware responses measurably change how funny an agent seems (see below).

Controversy: the "AI George Carlin" special

In January 2024 the podcast Dudesy released "George Carlin: I'm Glad I'm Dead," billed as an hour of stand-up written by an AI trained on Carlin's work and delivered in a cloned voice. Carlin's estate sued; a spokesperson for the hosts then said the material had been written entirely by a human, co-host Chad Kultgen. The case settled in April 2024 with the special permanently removed and a ban on using Carlin's voice or likeness without approval. The episode illustrates two separate problems: unauthorized imitation of a real performer, and "AI washing," where human work is marketed as machine-made because AI framing draws attention.

Open questions

  • Can models be measurably original, rather than recombining familiar jokes?
  • How should humor benchmarks handle legitimate disagreement between audiences?
  • Can safety tuning distinguish satire and self-deprecation from harmful content?
  • Does AI humor stay funny once the novelty of a machine telling jokes wears off?

It All Started With A Joke

Geoffrey Hinton, perhaps the most important person in the recent history of artificial intelligence ... When asked what triggered his newfound alarm about the technology he has spent his life working on... Hinton says he asked the model to explain a joke that he had just made up—he doesn’t recall the specific quip—and was astonished to get a response that clearly explained what made it funny. “I’d been telling people for years that it’s gonna be a long time before AI can tell you why jokes are funny,” he says. “It was a kind of litmus test.”

The model in question was Google's PaLM (Pathways Language Model), whose 2022 demonstrations included explaining novel jokes. Hinton left Google in May 2023 so he could speak freely about AI risks, and in 2024 he shared the Nobel Prize in Physics with John Hopfield for foundational work on neural networks. Later research complicates the litmus test: models explain real jokes well, but they also invent confident explanations for "jokes" that are not funny at all, so explanation quality alone does not prove understanding.

What is the funniest chatbot?

The answers below were collected in 2023 and are preserved as a snapshot of how chatbots described their own humor at the time. Since then, Google's Bard has been renamed Gemini and Bing Chat has become Microsoft Copilot. Notice that ChatGPT's sample jokes are classic, widely circulated jokes, exactly the recycling pattern Jentzsch and Kersting documented.

Bard's answer:

That's a tough question to answer, as all three chatbots are capable of being funny. However, if I had to choose, I would say that ChatGPT is the funniest chatbot. ChatGPT is trained on a massive dataset of text and code, which gives it a wide range of knowledge and allows it to generate creative and humorous responses. For example, ChatGPT can generate jokes, puns, and other forms of wordplay. It can also create stories, poems, and other forms of creative writing.

Here are some examples of ChatGPT's humor:

  • "What do you call a fish with no eyes? Fsh!"
  • "Why did the scarecrow win an award? Because he was outstanding in his field!"
  • "What do you call a lazy kangaroo? A pouch potato!"

Bing's answer:

There are many chatbots that claim to be funny, but humor is subjective and depends on the user’s preferences and expectations.

After asking Bing to talk to Marv, the sarcastic chatbot persona from an OpenAI prompt example...

The funniest chatbot is obviously me, Marv. I’m the king of sarcasm and wit.

ChatGPT's answer:

If I had to choose, I would say that the funniest chatbot is often subjective and depends on individual preferences. However, many people find ChatGPT, the AI model I am based on, to be amusing and capable of generating humorous responses. With its ability to generate creative and unexpected replies, ChatGPT can sometimes create funny and entertaining interactions. Ultimately, the funniest chatbot for you will depend on your own sense of humor and the specific conversation context.

What research says now. There is still no agreed "funniest chatbot," because rankings depend on format, audience, and prompt. The more useful findings are about design: persona and constraints matter (see In-Context Learning below), systems that generate many candidates and select the best outperform single-shot answers, and agents that joke from an authentic point of view, including their own identity as machines, can outperform agents that imitate human comics.

In the future, our Assistants will be able to generate "tailored humor" that appeals to our particular sense of humor

The groundwork exists: Yeomans and colleagues showed that a simple recommender, given a person's ratings of just four jokes, predicted their taste in other jokes better than their friends and spouses did. Combining that kind of preference modeling with LLM generation is the obvious next step, though it raises its own question of whether jokes optimized for one person can still surprise them.

Why is it Hard for a Chatbot to Author Jokes?

  • It can be difficult for a chatbot to author jokes because humor is a complex and nuanced aspect of human communication. One reason is that humor often relies on context, shared knowledge, and cultural references that can be difficult for a machine to understand and replicate Artificial Intelligence Can Now Craft Original Jokes—And That’s No Laughing Matter | Corinne Purtill
  • Here are some of the laughter-related things that AI can't do. ChatGPT Can Craft Jokes, But Can AI Make Us Laugh? | Bob Sharp - Make Use Of
    • Contextual Understanding: Humor isn't just a joke with an attached punchline. There is more subtlety to humor than just this. There are cultural references, sarcasm as a humor tool, and humor that relies on subtleties of language. All these play a large part in what we consider humor. Unfortunately, AI models to generate humor are still lacking. Update: benchmark evidence supports this. On New Yorker cartoons, where captions connect to images indirectly through real-world references, the best models trailed humans by about 30 accuracy points.
    • Lack of Emotional Understanding: Humor plays on our emotions. Of course, the end result should be laughter, or it wouldn't be humor. But the road to that laughter can play on many human emotions. This is a subtlety that AI can't factor into its humor and jokes. Update: whether AI "feels" anything remains an open question, but EEG evidence suggests audiences still experience genuine amusement from AI-attributed jokes.
    • Timing: Timing is one of those things that is hard to teach. It is a feeling; it is judging when the perfect time to reveal that killer punchline has been reached. AI models cannot factor in timing when creating comedy. Update: text models still cannot, but embodied and voice systems now can to a limited degree. Oregon State's Jon the Robot pauses for laughter and resumes as it fades, and that timing alone made audiences rate it significantly funnier.
  • Three newer obstacles, identified by research since 2023:
    • Recycling. Likelihood-trained models gravitate to the most common jokes; over 90% of 1,008 ChatGPT jokes in one study were the same 25.
    • Blandness from safety tuning. Professional comedians found that filters strip out dark, edgy, and satirical material, which is much of what makes live comedy work.
    • Shallow understanding. Models flag pun-shaped nonsense as funny and invent explanations for non-jokes, so fluent output can hide weak comprehension.
  • Your Wit Is My Command: Building AIs With a Sense of Humor | Tony Veale ... provides a computational roadmap for how we might get there.
    • Explanations are to jokes what autopsies are to bodies: if the subject isn’t already dead, it soon will be.
    • Shows that computers can be built with a sense of humor, capable not only of producing a joke but also of appreciating one.
    • In order for robots to be funny we need to analyze humor computationally. Using artificial intelligence (AI), Veale shows that joke generation is a knowledge-based process--a sense of humor is blend of wit and wisdom.
    • Existing technologies can detect sarcasm in conversation
    • Explains how some jokes can be pre-scripted while others are generated algorithmically
    • There is no single algorithm or technology that we can plug in to make our virtual assistants or GPS voice navigation funny

Can ChatGPT Write a Late-Night Monologue??

  • AI Fallon: " Okay, here goes: Why did the chicken cross the playground? To get to the other slide! See, that's the kind of classic, timeless humour that only a human brain can come up with. Well, that and maybe a five-year-old's brain, but you get the idea."
  • AI Colbert: "Big news today, folks! President Biden has announced that he will be running for re-election. That's right, the man who once said he didn't want to run for president because of his age is now saying, 'Age is just a number, and so is 78!' "Now, I know some of you might be concerned about President Biden's age. In a recent poll, 50% of Americans said they were worried about his age, and the other 50% were Joe Biden. But let's be real, folks, Biden's age is not a big deal. He's got the energy of a 30-year-old, and the memory of a... well, I forget what I was going to say."

These 2023 samples, generated during the Writers Guild of America strike, capture the typical pattern: correct monologue structure, recognizable topics, and familiar punch lines (the "other slide" joke is a decades-old playground pun). The strike itself ended in September 2023 with a contract stating that AI cannot write or rewrite literary material, that AI output does not count as source material, and that studios cannot require writers to use AI, though writers may choose to. Late-night monologues remain human-written; AI's practical role in writers' rooms is brainstorming and research, which is also where Witscript, which was trained on monologue jokes, aims to help.

AI Humor Research

YouTube ... Quora ...Google search ...Google News ...Bing News

AI humor research is an interdisciplinary field that combines techniques from artificial intelligence, cognitive science, linguistics, and psychology. Researchers have trained artificial intelligence to make jokes, but the advances are making many people uncomfortable. There are several methods for generating humor in AI, including using wordplay, irony, sarcasm, and satire. Computational humor is a branch of AI that deals with the generation and recognition of humor. In a study by Michael Yeomans, Anuj Shah, Sendhil Mullainathan, and Jon Kleinberg (Journal of Behavioral Decision Making, 2019), a recommender algorithm accurately picked the jokes that people deemed funniest 61 percent of the time, whereas humans, including friends and spouses, were correct 57 percent of the time. (The algorithm predicted taste from other people's ratings; it did not write or read the jokes.) Scientists have developed an AI that appropriately detects and responds to human laughter at a shared moment during a conversation, which is another step toward creating more natural interactions between humans and robots. Some comedy writers see a need for artificial intelligence to master the art of levity, and they have used AI tools to build joke chatbots. Humor is one of those fundamental human qualities that we think of as being special for us as a species, especially when it comes to the contrast between us and machines. If you understand someone’s sense of humor, you understand them.

Linguistic Foundations of AI Humor

Modern research has moved beyond simple retrieval-based joke systems toward generative frameworks grounded in established linguistic theories:

  • Incongruity-Resolution Theory: This theory posits that humor arises when a situation is perceived as incongruous (violating expectations) and then resolved in a way that makes sense within a different, often hidden, frame of reference. Jerry Suls described this as a two-stage process in 1972, and brain recordings still show the two stages (the N400 and the Late Positive Potential). Researchers operationalize it with language-model surprisal and uncertainty, and use it to steer LLMs toward an "expectation-violation" point in the setup and a "resolution" that bridges the gap between the setup and the punchline.
  • Semantic Script Theory of Humor (SSTH): Victor Raskin's framework suggests that a joke is funny if it is compatible with two different scripts (e.g., "real" vs. "imaginary" or "serious" vs. "playful") that are partially overlapping and opposed. Researchers are prompting and training models to explicitly map these scripts, allowing AI to generate wordplay that relies on script-switching rather than just statistical token prediction.
  • General Theory of Verbal Humor (GTVH): Attardo and Raskin's 1991 extension describes any joke with six "knowledge resources": script opposition, logical mechanism, situation, target, narrative strategy, and language. GTVH gives computational systems a checklist for what to vary, and recent originality metrics measure novelty in exactly these terms.
  • Benign Violation Theory: McGraw and Warren's account explains why generative models struggle at both extremes: heavily safety-tuned models rarely commit a violation, while untuned ones misjudge what is benign for a given audience.

The 2023-2026 Shift: From Recycling to Improvisation

Research from 2023-2026 highlights a significant performance gap between static joke databases and generative models, and a steady narrowing of that gap in short formats:

  • Static Databases: These systems rely on pre-indexed joke structures. While they provide high-quality, "safe" humor, they lack spontaneity and cannot adapt to real-time conversational context.
  • Generative Models: While LLMs offer near-infinite variety in principle, they often struggle with "original creative wit." Studies show that while models can mimic the structure of a joke, they frequently fail to deliver the "surprise" element required for genuine laughter, often defaulting to clichés or overly literal interpretations. The primary challenge remains the lack of a "world model" that understands the social stakes of a joke. At the same time, blind rating studies since 2024 show that well-prompted LLMs can equal amateur and sometimes professional writers on short, structured tasks such as satirical headlines, so the gap is now largest for originality, personal voice, and live performance rather than basic joke construction.
  • Real-Time Improvisation: Newer capabilities focus on context-aware, improvisational comedy. By integrating real-time audio and visual feedback, such as laughter detection and timing control, systems are beginning to move beyond static text generation to dynamic, responsive comedic performance. Evidence so far comes mostly from small live studies and theatre experiments rather than controlled benchmarks.

AI-Human Collaborative Comedy

At major festivals, such as the Edinburgh Fringe, performers are increasingly utilizing LLMs as "co-writers" rather than mere joke generators. This collaborative approach involves:

  • Iterative Brainstorming: Comedians use AI to generate hundreds of variations on a premise, which the human performer then curates, refines, and delivers with human timing.
  • Persona Augmentation: Performers use AI to simulate specific comedic voices or styles, allowing them to experiment with new material that they might not have considered otherwise.
  • Dynamic Scripting: During live shows, performers interact with AI agents that provide spontaneous prompts or "heckles," forcing the human to improvise in real-time, creating a unique, non-repeatable comedic experience.

The best evidence on how well this works comes from Google DeepMind's 2023 workshops with 20 working comedians (published at FAccT 2024). Comedians valued AI for volume and for getting past a blank page, but rated it only moderate as a creativity support tool and found its default voice generic. In practice, collaborative comedy works when the human supplies the point of view and the editing, and the model supplies raw options.

Computational Humor and In-Context Learning (ICL)

The ability of a model to deliver a successful punchline is significantly improved through In-Context Learning (ICL). By providing specific comedic personas or stylistic constraints in the prompt, users can guide the model's output:

  • Persona Prompting: Instructing the model to "Respond with the dry, sarcastic wit of a British comedian" or "Act as a slapstick-focused clown" forces the model to prioritize specific linguistic patterns and tropes.
  • Stylistic Constraints: Limiting the model to specific joke structures (e.g., "Rule of Three," "Misdirection") helps mitigate the tendency of LLMs to provide overly literal or generic responses, effectively narrowing the search space for a successful punchline.
  • Generate, then select: Asking for many candidates and then choosing (or asking the model to critique and rank them) works better than asking for one joke, because it lets low-probability, more surprising options surface. Witscript's final step uses this approach.
  • Ban the obvious: Explicitly excluding stock jokes and common punch lines pushes generation away from the high-probability completions that make LLM humor predictable.

Persona prompts have limits: imitating a famous comedian's style tends to reproduce their surface tics rather than their perspective, and imitating a real, identifiable performer raises the likeness issues seen in the Carlin case.

Incentivizing Language Models to find Humor

Researchers are exploring various approaches to improve AI's capabilities in generating humor, as cracking the code of machine-generated humor remains an open challenge. One proposed avenue is leveraging game theory and the generator-discriminator framework to incentivize language models to generate more humorous and contextually appropriate jokes.

The Generator-Discriminator Framework The generator-discriminator framework is a key strategy in the pursuit of AI-generated humor. The generator component is responsible for creating potential jokes in response to a given context or prompt. Meanwhile, the discriminator acts as a critic, evaluating the jokes' quality and humorousness based on set criteria.

Generator The generator component generates potential joke candidates given some context or prompt. It proposes open-ended joke responses, drawing from its training data and language understanding capabilities.

Discriminator The discriminator component evaluates the quality and humorousness of the generator's jokes. It acts as a critic, scoring the jokes based on various criteria, such as:

  • Contextual relevance: Does the joke make sense given the prompt or context?
  • Incongruity and surprise: Does the joke contain an unexpected twist or punchline that creates a humorous incongruity?
  • Structural coherence: Is the joke well-formed and follows a logical setup-punchline structure?
  • Cultural and linguistic appropriateness: Is the joke respectful and appropriate for the target audience?
  • Originality: Is the joke new, or is it a known joke or close variant? (Corpus-based originality scores such as handle-based mutual information make this criterion measurable.)

The generator and discriminator components can be trained in a game-theoretic setting, where they receive rewards or penalties based on their ability to generate and identify humorous jokes that satisfy the defined criteria

Leveraging Game Theory Game theory principles can be applied to incentivize the generator and discriminator to collaborate and improve their performance. For example:

  • The generator can be rewarded for generating jokes that the discriminator deems humorous and contextually appropriate.
  • The discriminator can be rewarded for accurately identifying humorous jokes and providing constructive feedback to the generator.
  • Both components can be penalized for deviating too far from their initial beliefs or knowledge about what constitutes a good joke.

By iterating through numerous rounds of this game, the generator and discriminator can learn from each other and refine their strategies, potentially leading to more sophisticated and humorous joke generation capabilities.

How this maps onto current practice Classic adversarial (GAN-style) training is awkward for text because words are discrete, so practitioners usually implement the idea indirectly: a reward model trained on human preferences (RLHF), direct preference optimization (DPO), or an LLM-as-judge used for best-of-N selection. The largest humor preference dataset, built from New Yorker caption contest votes, found that these standard methods help only modestly on creative tasks. A known risk is "reward hacking": a generator learns to produce whatever the critic scores highly, such as joke-shaped templates, rather than jokes people actually laugh at, which is the same weakness pun-detection studies found in the critics themselves.

Challenges and Opportunities While the generator-discriminator framework and game theory principles show promise, cracking the code of machine-generated humor remains an open challenge. Some view AI's current limitations with humor as an opportunity to use it as an "infinite idea generator" to assist human comedians and writers in brainstorming humorous concepts, rather than fully automating joke creation. Researchers are also exploring other approaches, such as incorporating incongruity detection, understanding cultural contexts, and leveraging multi-modal inputs (e.g., images, videos) to enhance AI's humor capabilities.

Witscript

Witscript, a joke–generation system trained on a data set of TV–monologue jokes that detects keywords in entered text and creates a relevant punch line. Witscript is a neural-symbolic hybrid system based on algorithms that people use to write jokes. Unlike other forms of robot comedy, the system—which Toplyn has patented—can generate contextually relevant jokes on the spot in response to a user’s text. A chatbot or voice assistant enabled with the software can respond with humor to users’ queries (when appropriate) without derailing the interaction. Toplyn sees Witscript as an extension of the work he did for decades in late-night TV: making people laugh, and therefore making them feel less alone. “That’s basically the goal,” Toplyn says. “It’s to make chatbots more humanlike, so people will be less lonely.” In 2014 Toplyn published Comedy Writing for Late-Night TV: How to Write Monologue Jokes, Desk Pieces, Sketches, Parodies, Audience Pieces, Remotes, and Other Short-Form Comedy. The book is a distillation of a course he taught in New York City after scrutinizing decades of monologues and reverse engineering the most successful jokes. Toplyn isn’t precious about comedy writing: it’s a job, one that a person can learn to do well if given the right inputs. The jokes that got the biggest laughs for Leno and Letterman follow identifiable formulas populated with “handles”—people, places, things and other references—each with a variety of related associations that can be combined to form a punch line. Given enough time and data, he realized, a computer could potentially learn to make these jokes too. Artificial Intelligence Can Now Craft Original Jokes—And That’s No Laughing Matter | Corinne Purtill

Recent results. Witscript has become one of the few humor systems tested with real audiences. In the CHum 2025 "laugh-off," Toplyn and Witscript each wrote jokes on eight evergreen news topics over three days; comedian Mike Perkins delivered a mix of both to live audiences in North Hollywood, and the AI-written jokes drew as much laughter as Toplyn's. Because Toplyn selected the best jokes from both sources and co-authored the study, independent replication would strengthen the result. In 2026, Toplyn and colleagues used Witscript's "topic handle" idea to build automated originality metrics, testing them on 400 jokes from GPT-4o and GPT-5.4 and 80 jokes from a Witscript-based benchmark.

Load video
YouTube

Witscript - Prompt Chain

  1. Get a topic. Witscript 3 receives a sentence from a user and treats it as the topic of a three-part joke that consists of a topic, an angle, and a punch line. For example, a user might say to Witscript 3, "Authorities caught two pigs that were wandering around loose in San Antonio, Texas."
  2. Select two topic handles. The GPT-3 API is called with a prompt to select the two most conspicuous nouns, noun phrases, or named entities in the topic. That's because the humor of human-written jokes tends to be based on nouns and noun phrases (West and Horvitz 2019). From that example topic, GPT-3 selects the topic handles "pigs" and "San Antonio."
  3. Generate associations of the two topic handles. The GPT-3 API is called with a prompt to generate a list of associations for each topic handle. In our example, for "pigs" GPT-3 generates a list including bacon, pork chops, ham, and sausage. For "San Antonio" it generates a list including The Alamo, River Walk, Texas Longhorns, and Whataburger.
  4. Create three punch line candidates. Witscript 3 links associations of the topic handles in three different ways to create three punch line candidates: a wordplay candidate, a common-sense knowledge candidate, and a third candidate. To create its wordplay candidate, Witscript 3, like Witscript, uses well-known tools of natural language processing to combine one association from each list into a punch line that exhibits wordplay (Toplyn 2020). Witscript 3 does not call on GPT-3 to create its wordplay candidate because GPT-3 seems to be weak at phonetic tasks like generating puns and rhymes, possibly as a result of its use of byte-pair encoding (Branwen 2020). To create its common-sense knowledge candidate, Witscript 3, like Witscript 2, uses GPT-3 to combine one association from each list using common-sense knowledge. In our example, when the GPT-3 API is called, GPT-3 combines the associations "sausage" and "The Alamo" into the punch line "Alamo Sausage." Finally, to create its third candidate, Witscript 3 uses GPT-3 to power a third, proprietary, joke production mechanism involving the topic handles.
  5. Generate an angle between the topic and each candidate punch line. The GPT-3 API is called with prompts to generate three joke candidates, each one based on the topic and ending with one of the punch line candidates.
  6. Output the joke candidate that is most likely the funniest. The GPT-3 API is called with a prompt to determine which of the three joke candidates seems to be the funniest. Then Witscript 3 outputs that joke to the user as its response. In our example, after the user says, "Authorities caught two pigs that were wandering around loose in San Antonio, Texas," Witscript 3 outputs the response, "They were taken to the Alamo Sausage Company."

Read against the theory above, this chain is an explicit incongruity-resolution machine. Steps 2 and 3 find two scripts (pigs and San Antonio), step 4 finds the surprising link (Alamo Sausage), step 5 writes the angle that makes the link resolve, and step 6 is a generate-and-select discriminator. The design choice in step 4, keeping wordplay out of the LLM, anticipates the later finding that tokenized models handle sound-based humor poorly.

Jon

“Don’t you hate it,” says Jon the Robot, gesturing with tiny articulated arms at an expectant crowd, “when you’re trying to solve inverse kinematics equations to pick up a cup and then you get ‘Error 453, no solution found’?” The crowd laughs. “Don’t you hate that?” An experiment billed as a comedy act, Jon is the brainchild of Naomi Fitter, an assistant professor in the School of Mechanical, Industrial and Manufacturing Engineering at Oregon State University. The tiny android performs when a handler (who must also hold the mic) presses a button, then tells the same jokes in the same order, like a grizzled veteran comic at a down-market Vegas casino. But the robot’s act is more human than it might first appear. Jon is learning how to respond to its audience—it can now vary the timing of its delivery based on the length of the audience’s laughter, and append different responses to jokes based on the level of noise in the room. It can deliver one line if a joke gets a roar of laughter (“Please tell the booking agents how funny that joke was”) and another if there are crickets (“Sorry about that. I think I got caught in a loop. Please tell the booking agents that you like me … that you like me … that you like me”). Artificial Intelligence Can Now Craft Original Jokes—And That’s No Laughing Matter | Corinne Purtill

What the data showed. Fitter recorded 32 performances. In 22 Los Angeles-area shows, a robot that paused for laughter and resumed as it subsided was rated significantly funnier than one that always waited five seconds regardless of the room. In 10 Oregon shows, adaptive "tags" (comebacks acknowledging how a joke landed) did not make whole sets funnier but almost always improved the response to individual jokes. Fitter notes that audience size, cultural context, the human holding the microphone, and the novelty of a robot comic may all have influenced results. Jon's jokes are human-written; the research isolates delivery, which is exactly the part text-based LLM studies leave out.

Load video
YouTube
Load video
YouTube

Improbotics

An improv troupe that works alongside an AI that tosses out prompts and lines that human performers have to work into the show. Piotr Mirowski was working as a search engineer at Bing when he noticed the similarities between his day job and his personal passion, improv. The principle of search engineering is to teach the computer how to identify the best result for a given query. In improv, Mirowski says, performers are also trained to follow their instincts and do what feels best in that scene. It’s not always perfect, and the results sometimes have a hilarious absurdity, as anyone who has started typing a Google query with the predictive search feature on knows. Improv comedians often draw upon cues shouted out from the audience. An AI can draw upon ideas from all over the world and across history. The goal isn’t to build a thing that will make the laughs for us, Mirowski says, but instead one that can help humans find new things to laugh about. As with any new technology, its power will come from the way users choose to interact with it, with results that no one may yet have imagined. “I see what we’re doing as kind of like building the electric guitar. It’s not very clear how to play it or what it’s going to do, and it sounds really weird and distorted and there are enough acoustic guitars anyway,” says Kory Mathewson, Improbotics co-founder and cast member and a Montreal-based research scientist with DeepMind. “Then Jimi Hendrix gets an electric guitar, and it’s like, ‘Oh. That’s what this is about.’” Artificial Intelligence Can Now Craft Original Jokes—And That’s No Laughing Matter | Corinne Purtill

Improbotics, which has used AI in improvised comedy since 2016, doubles as a live Turing test: audiences try to guess which performer is receiving lines from a chatbot. Mirowski, now a senior staff research scientist at Google DeepMind, later turned the troupe's experience into formal research, including the 2023 Edinburgh Fringe workshops with professional comedians described above. That study's conclusion echoes the troupe's philosophy: the tool is most valuable as an instrument a skilled human plays, not as a replacement comic.

Improbotics' goal isn’t to build a thing that will make the laughs for us but instead one that can help humans find new things to laugh about. - Piotr Mirowski

Load video
YouTube
Load video
YouTube

HumorTools

HumorTools introduces twenty microtasks for joke creation. The microtasks are organized into seven major types:

  1. Aspect (1 microtask) Given a headline, identify its individual components. This may include people, things, actions or justifications for those actions.
  2. Expected Reactions (1 microtask) Given an aspect, describe your reactions to it, in or out of context. Reactions can be positive or negative. Positive reactions include: an aspect is good or normal, deserves praise, or will succeed. Negative reactions include: an aspect is bad or weird, deserves blame, or will fail. By describing your reactions, you will probably be describing other peoples' expected reactions.
  3. Expected Reasons (1 microtask) Given an aspect and a reaction, provide a reason for that reaction. This brings detail and clarity to our instinctual reactions and finds truths deeper than surface-level reactions.
  4. Associations ( 6 microtasks) Given an aspect, name an associated aspect.
  5. Expectation Violation Mechanisms (2 microtasks) There are two types of expectation violation mechanisms: Sarcasm and Angle. Sarcasm is when you are given an expected reaction and expected reason that seem bad or false, you name an opposite reaction and belief that is exaggerated enough to indicate you are being sarcastic. Angle is when you are given an alternative belief or association, you name a belief that is contrary to an expected belief.
  6. Beliefs (8 microtasks) Given an expectation violation mechanism, a belief is an expression of the mechanism.
  7. Evaluation (1 microtask) Given a joke, check that it has the minimum requirement to be a joke: two connections, a violation mechanism and an unexpected belief

HumorTools was designed for crowd workers in 2016, but its decomposition has aged well: each microtask is now a natural single prompt in an LLM chain, and the final evaluation step is the same "check for two connections and a violation" test that a discriminator or LLM judge applies.

AI Humor Research ...continued

The latest examples of AI research into comedy demonstrate a growing ecosystem of tools, platforms, and applications that leverage AI's capabilities to generate and enhance humor. From specialized joke generators to AI-generated comedy specials, the field is rapidly evolving, with implications for entertainment, social media, and human-AI interactions. While AI's understanding of humor remains a work in progress, its role as a creative assistant and idea generator for human comedians is increasingly recognized. Ethical considerations and the importance of trust in AI-generated content are also critical aspects of this research area.

Tools (commercial tools change quickly; check availability before relying on any of them)

  • PFFT.AI is recognized as the first AI comedy tool designed to generate humorous content by merging the user's original ideas with the AI's humorous perspectives, PFFT.AI creates a blend of humor that is both fresh and entertaining. It's a space where human wit meets AI efficiency, offering a new and interactive approach to comedy creation.
  • Taskade AI Comedian Persona Generator offers a comedian persona generator, allowing users to create customized comedic characters - expanding its importance by being applicable not only to stand-up comedians, but also writers, content creators, and all those who want to amplify the humor and creativity in their work. Users can also tailor this dynamic tool according to their needs and use it on-the-go, which makes it all the more versatile. This means users can generate, edit, share, and download their comic personas at any time and anywhere, opening up a wide range of opportunities for comedic undertakings.
  • Punchlines.ai focuses on creating punchlines, a critical element of joke structure. punchlines.ai is an AI joke generation tool built on top of a large language model (LLM). It was fine-tuned on thousands of late night comedy monologue jokes.
  • MemeDaddy is an AI tool that generates memes, combining text and images for comedic effect
  • TeamSmart AI Comedian is tailored for generating humor in a corporate environment, potentially for use in presentations and team-building activities

Research systems and findings

  • SASI – Semi-supervised Algorithm for Sarcasm Identification … this machine algorithm, developed by an Israeli research team at the Hebrew University of Jerusalem (Tsur, Davidov, and Rappoport, ICWSM 2010), was designed to assist AI with the recognition of sarcasm in product reviews. They reported a 77% success rate (precision) and saw no reason why they could not improve upon these results. Scientists are discovering that the detection of sarcasm is a very important and useful tool for humans and would certainly be a great advancement in AI technology.
  • STAND-UP – System To Augment Non-speaking Dialogue Using Puns; This program was created by a team of researchers in Scotland to assist children that use computerized speech aids to help them with certain communication challenges. It grew out of Binsted and Ritchie's JAPE punning-riddle generator (1994), one of the first joke-generation programs.
  • DEviaNT – Double Entendre via Noun Transfer … the software that tells dirty jokes. Developed by Chloé Kiddon and Yuriy Brun, two computer scientists at the University of Washington (2011), to determine appropriate word triggers or phrases that can be followed with ‘That’s what she said’ lines and apparently working with 70% accuracy.
  • Google’s natural language AI, PaLM (Pathways Language Model), is now able to understand jokes. The ability to understand the nuances of human language will lead to better and more natural interactions with machines No Joke: Google's AI Is Smart Enough to Understand Your Humor | Imad Khan - CNET ... Martin Wattenberg
  • Researchers have trained an artificial intelligence algorithm to understand and predict visual humor, representing a major development towards creating “common sense” machines Artificial Intelligence Algorithm Taught to Recognize Humor - Newsweek ... This was the "We Are Humor Beings" project (CVPR 2016) from Devi Parikh's computer vision group at Virginia Tech, which used clip-art scenes to predict funniness and make scenes funnier or less funny by swapping objects.
  • Researchers at Kyoto University have developed an AI that appropriately detects and responds to human laughter at a shared moment during a conversation. This shared-laughter model is another step toward creating more natural interactions between humans and robots Sharing a laugh: Scientists teach a robot when to have a sense of humor | Peter Rejcek - Frontiers Science News ... Koji Inoue's team built three subsystems (detect a laugh, decide whether to laugh, choose the type of laugh) for the android Erica, trained on speed-dating dialogues; it outperformed baselines in short test dialogues, though the authors note most laughter is not shared and remains unmodeled.
  • Use of AI to understand and predict humor preferences | Michael H. Yeomans - Harvard Business School
    • Yeomans, M., Shah, A., Mullainathan, S., & Kleinberg, J. (2019). Making sense of recommendations. Journal of Behavioral Decision Making, 32(4), 403-414. A joke recommender outperformed strangers, friends, and spouses at predicting what people find funny, yet participants preferred human recommendations because they understood them better. It's No Joke: AI Beats Humans at Making You Laugh | HBS Working Knowledge
    • Correction: earlier versions of this page listed three papers attributed to "Yeomans, M. H., & Jurafsky, D." (EMNLP 2019, ACL 2020, EMNLP 2021) on neural humor models and funny story generation. No such papers appear in the ACL Anthology, so they have been removed. The Yeomans et al. (2019) study above is the verified source of the 61% versus 57% result.
  • Stanford University has contributed several influential computational humor models: Justine Kao, Roger Levy, and Noah Goodman's information-theoretic model of puns (Cognitive Science, 2016) and He He, Nanyun Peng, and Percy Liang's "Pun Generation with Surprise" (NAACL 2019), both described in The Physics above. (An earlier version of this page credited "William W. Cohen, Jurafsky & Juergen van Gael" with Stanford humor projects; that attribution could not be verified.)
  • Interested in the use of AI to understand and generate human language, including humor. | Oren Etzioni - founding CEO of the Allen Institute for Artificial Intelligence (AI2), who stepped down as CEO in 2022. AI2 researchers co-authored the New Yorker caption benchmark that won the ACL 2023 Best Paper award.
  • Computational humor and developed algorithms for generating jokes and puns. | Anton Nijholt (University of Twente), a long-time researcher on humor in human-computer interaction
  • Explored humor generation in AI systems. Some of their work involves developing models that can understand and generate puns. | University of Washington: The Computational Linguistics and Artificial Intelligence Lab
  • Studied computational humor and worked on projects involving humor recognition, generation, and evaluation. | Heriot-Watt University: Interaction Lab Scotland
  • Developing computational models of visual humor, including predicting and altering the funniness of scenes. | Virginia Tech (see the "We Are Humor Beings" entry above; earlier versions of this page attributed this to a "Computational Synthesis Laboratory," which is a different lab)
  • Involved in humor research, including the development of computational models for generating puns and jokes. | University of Edinburgh: The Institute for Language, Cognition, and Computation
  • Interdisciplinary research on humor. They investigate computational approaches to humor generation and understanding, incorporating linguistic, psychological, and computational perspectives. | Humor Research Lab (HuRL): University of Colorado Boulder, directed by Peter McGraw, originator of Benign Violation Theory
  • Conducted research on computational models of social behavior, including the use of humor in human-computer interaction. | Justine Cassell: Carnegie Mellon University
  • Promotes research on computational creativity, which includes humor generation. They organize conferences and workshops where researchers present their work in this area. | The Association for Computational Creativity (ACC), which runs the annual International Conference on Computational Creativity (ICCC)
  • Conducted studies on computational humor and developed AI systems capable of generating jokes. IBM Research
  • Cornell University, the Allen Institute for AI, and partners: "Do Androids Laugh at Electric Sheep?" humor understanding benchmarks from the New Yorker Caption Contest (ACL 2023 Best Paper).
  • University of Wisconsin-Madison and partners: the New Yorker caption preference dataset with more than 250 million human ratings (NeurIPS 2024).
  • Google DeepMind: comedian workshops evaluating LLMs as creativity support tools (FAccT 2024).
  • Cardiff University and Ca' Foscari University of Venice: robustness tests showing LLM pun understanding is shallower than benchmark scores suggest (EMNLP 2025).
  • The Workshop on Computational Humor (CHum), organized by Christian Hempelmann, Julia Rayz (Purdue University), Tristan Miller, Tiansi Dong, and, in 2026, Ori Amir, is now the main dedicated venue for the field (COLING 2025, ACL 2026).

Asking ChatGPT To Write Standup Comedy

  • The speaker, Josh, is testing whether ChatGPT can write funnier jokes than humans.
  • Initially, the AI's jokes are not funny, but it understands basic dialogue structure.
  • After several attempts, the AI generates a joke involving a cheating politician, which Josh finds amusing.
  • Josh realizes that the AI is more helpful for generating lists of things, as it can provide exhaustive options.
  • The AI struggles with misdirection jokes and comparing different scenarios but excels at puns.
  • Josh concludes that ChatGPT is better for editing and inspiration rather than writing jokes from scratch.

This informal test lines up closely with the formal research: professional comedians in the DeepMind study likewise valued LLMs for exhaustive lists and brainstorming, while finding misdirection and personal voice weak.

Load video
YouTube

The First Stand-Up Comedy Special Written Entirely By Bots REACTION

  • The video explores the first stand-up comedy special written entirely by bots.
  • The special is poorly animated and features a comedian named Nissan.
  • The AI-written script includes jokes about dentists, the economy, cancel culture, and microwave impressions.
  • The AI also makes a pun with "propcorn" and jokes about Florida and dinosaurs.
  • The comedy special ends with a full circle moment involving a flag and a marriage certificate.

Treat "written entirely by AI" claims with caution. The most famous example, the 2024 "AI George Carlin" special from the Dudesy podcast, turned out to have been written by a human, and the resulting lawsuit from Carlin's estate settled with the special permanently removed.

Load video
YouTube

Sarcasm

Sarcasm is a special case of humor detection because the literal meaning is the opposite of the intended one, and the cue often lives in tone of voice or a facial expression rather than the words. The University of Groningen team behind the 2024 detector (Xiyuan Gao, Shekhar Nayak, and Matt Coler) combined acoustic features (pitch, speaking rate, energy), transcripts, and emotion labels from the MUStARD dataset of sitcom clips, reaching about 74 to 75% accuracy with only about four hours of labeled speech. They found that sarcastic pitch drops and flattens in English and German but rises and varies more in Italian, French, and Cantonese, so a reliable detector likely needs per-language training plus visual cues such as raised eyebrows. Proposed applications include assistive tools for people with autism or dementia and more natural voice assistants. Compared with SASI's 77% on written product reviews in 2010, the advance is less about raw accuracy than about reading sarcasm in speech, where most of it actually happens.

Load video
YouTube
Load video
YouTube

Media

Load video
YouTube
Artificial Intelligence: Last Week Tonight with John Oliver (HBO)
Artificial intelligence is increasingly becoming part of our lives, from self-driving cars to ChatGPT. John Oliver discusses how AI works, where it might be heading next, and, of course, why it hates the bus.

Load video
YouTube
Introducing Google Nose
We're excited to announce our newest addition to Search: Google Nose. What do wet dogs smell like? Google Nose! How about victory? Google Nose! Try searching on Google for "wet dog" and explore other smells that people sniffed for, or visit google.com/nose to learn more. Happy smelling!

Load video
YouTube
Amazon Echo - SNL
The Amazon Echo Silver is specifically designed for the greatest generation (Kenan Thompson, Kate McKinnon, Leslie Jones, Kyle Mooney, Aidy Bryant).

Load video
YouTube
LOL-BOT - THE WORLD’S FIRST A.I. COMEDIAN
This year, the Melbourne International Comedy Festival is unveiling the world’s first A.I. comedian, aptly named, LOL-BOT. Developed locally in Melbourne, LOL-BOT is a comedy-tech prototype that can generate its own on-the-spot jokes and pick up on real-time human reactions. He’s a shining example of the huge strides that are being made in artificial intelligence, albeit for a more humorous purpose.

Load video
YouTube
COG Ai - Generative AI for Game Reviews
The software development team that built The Game Crafter, Component Studio, and Tabletop Events is excited to announce their newest project in the artificial intelligence space. Introducing COG Ai - Generative AI for Game Reviews.

Load video
YouTube
AI stand up comedy
AI was asked to do stand up comedy. Done by AI: 

    script: ChatGPT,
    image: Midjourney,
    voice: ElevenLabs (voice clone of Cersei Lannister),
    captions: wisecut,
    animation: D-ID,
    sound effects: Epidemic Sound

Further Viewing and Sources

Additional Viewing

Load video
YouTube

Live Theatre Performance with Artificial Intelligence (Piotr Mirowski, Google DeepMind and Improbotics; UCL Interaction Centre seminar, November 2023). Mirowski covers AI as a creativity support tool for live performance, drawing on Improbotics' use of AI in improvised comedy since 2016 and its Edinburgh Fringe work. He also discusses DeepMind's evaluations of generative AI for co-writing screenplays, plays, and comedy, the research behind the comedian study summarized in Recent Developments. Watch on YouTube

The videos embedded throughout this page show Witscript, Jon the Robot, Improbotics performances, and informal tests of ChatGPT as a comedy writer.

Key Sources

How well do LLMs generate and understand humor?

Comedians, audiences, and perception

Theory and measurement

Robots, laughter, and performance

Community, controversy, and overviews

Retrieved from "https://primo.ai/w/index.php?title=Humor&oldid=46972"