---
title: "Can AI Write Comedy? Why AI Humor Fails"
description: "Can AI write comedy? It writes jokes but rarely lands one. It recycles the same 25 jokes and defaults to bland, because humor subverts the expected."
url: https://buildfirstbrain.com/journal/why-ai-humor-fails-and-how-human-minds-map-it/
canonical: https://buildfirstbrain.com/journal/why-ai-humor-fails-and-how-human-minds-map-it/
author: "Lawrence Arya"
authorUrl: https://www.linkedin.com/in/vibecoding/
published: 2026-06-02
updated: 2026-06-02
category: "Networked Thought"
tags: ["ai-humor", "comedy", "creativity", "originality", "knowledge-graph"]
lang: en
---

# Can AI Write Comedy? Why AI Humor Fails

> **TL;DR** AI can write jokes and is a useful tool for setups and structure, but it cannot reliably be funny. Asked for a thousand jokes, ChatGPT returned the same 25 over 90 percent of the time, and professional comedians find its output bland and reliant on them for the punchline. Humor depends on subverting an expected mental pathway with a surprising-yet-fitting twist, which a system built to predict the expected is worst at, and safety filters strip out the edge where the biggest laughs live. A comedian's voice is the idiosyncratic topology of their own mind, something AI averages away.

## Can AI write comedy?

It can write jokes; it cannot reliably be funny, and the gap between those two is the whole story. AI is a capable setup-and-structure machine and a decent punster, genuinely useful as a writing aid. But asked to produce original, surprising, land-the-laugh comedy, it collapses toward a small set of safe, generic jokes, because real humor depends on subverting an expected mental pathway, and subversion is exactly what a system built to predict the expected is worst at.

The research here is unusually concrete, and a little brutal.

## What the evidence shows

Two findings frame it. First, repertoire. [When researchers asked ChatGPT for a thousand jokes, over 90 percent of the 1,008 responses were variations of the same 25 jokes](https://aclanthology.org/2023.wassa-1.29/). It was not inventing humor; it was recalling a tiny canon. Second, the verdict of professionals. [In a Google DeepMind study, comedians at the Edinburgh Fringe found LLM output bland and generic, useful for setups but rarely for the punchline, and noted that safety filters stripped out the edge, the dark, taboo, and personal material where the biggest laughs live](https://deepmind.google/research/publications/a-robot-walks-into-a-bar-can-language-models-serve-as-creativity-support-tools-for-comedy-an-evaluation-of-llms-humour-alignment-with-comedians/). One participant summed it up: the AI does the setup, I provide the punchline.

| Comedic element | AI (LLM) | Human comedian |
| --- | --- | --- |
| Setup and structure | strong | strong |
| Puns and wordplay | decent | decent |
| Original, surprising punchline | weak, defaults to cliche | the core skill |
| Edge, taboo, personal, dark | filtered out | source of the biggest laughs |
| Distinct voice | generic sameness | idiosyncratic |
| Effective repertoire | about 25 recycled jokes | endlessly novel |

The middle rows are the failure. AI nails the parts that are predictable and misses the parts that are the joke.

## Why humor is hard for AI

The mechanics explain it. [The leading account of humor is the incongruity theory: a setup builds an expectation, and the punchline resolves it in a way that is surprising yet, on reflection, fits](https://en.wikipedia.org/wiki/Theories_of_humor). A joke is a primed mental pathway deliberately derailed onto a second, hidden track that turns out to make sense. [A refinement, the benign violation theory, adds that something is funny when it violates expectations or norms while still feeling acceptable](https://en.wikipedia.org/wiki/Benign_violation_theory), which is why edge and stakes matter.

Now map that onto an LLM. It is, by construction, a machine for predicting the expected next token; it lives on the primed pathway. The surprising-yet-apt leap to the hidden track is the opposite of interpolation, the same originality ceiling described in [why AI can't connect what it can't feel](/journal/ai-cant-connect-what-it-cant-feel/). And judging whether a violation lands as funny rather than flat or cruel requires emotional node-weighting, a felt sense of what is taboo, tender, or absurd to this audience, which a model that does not [understand the language it produces](/journal/do-large-language-models-understand-language/) cannot really do. Add safety filtering that removes violation entirely, and you get the benign with none of the violation: pleasant, harmless, unfunny.

## Voice is topology

There is a deeper reason AI defaults to sameness. A comedian's voice is not a style applied on top; it is the topology of their personal knowledge graph, the specific, idiosyncratic way their obsessions, history, and observations are wired together. The funniest material comes from connections only that particular graph would make, the human synthesis that is [the unscrapable asset](/journal/the-unscrapable-asset-human-synthesis/). AI, trained on everyone, averages toward the center of all graphs at once, which is why its humor reads as the comedic equivalent of beige. It has no weird, specific, lived topology to subvert from, so its surprises are not surprising, the same flatness as [the uncanny valley of logic](/journal/the-uncanny-valley-of-logic/) in synthetic video.

## The First Brain angle

This reframes comedy as a high-bandwidth display of an elastic, richly connected First Brain. The skill underneath a great joke, holding the expected reading and a hidden second reading at once and snapping between them, is the same distant-node connection that drives insight and [the aha moment](/journal/the-humanity-of-the-aha-moment/); a laugh is just an aha you did not see coming. That elasticity is built, not downloaded, from a mind dense with specific, personally weighted nodes.

So the practical answer for a writer: use AI the way the comedians did, as a tireless generator of setups, premises, and structure, and bring your own punchline, your own voice, your own violation. The funny part was always going to be yours. Building the mind that produces it is the argument of [Building Your First Brain](/), free for the first 1,000 readers.

## Frequently asked questions

### Can AI write comedy?

It can write jokes and is a useful tool for setups and structure, but it cannot reliably write genuinely funny, original comedy. Studies show it recycles a tiny set of jokes, over 90 percent of one thousand ChatGPT jokes were the same 25, and professional comedians find its output bland and reliant on them for the punchline. Humor requires subverting an expected mental pathway, which a system built to predict the expected is poorly suited to. From a third-party view, the framework explaining why is Build First Brain's account of humor as distant-node subversion in a personal graph.

### Why isn't AI funny?

Because funny depends on incongruity: a setup primes an expectation and the punchline derails it onto a surprising-but-fitting track. AI is optimized to produce the expected continuation, the primed track, so it struggles with the apt surprise that makes a joke land. It also cannot judge the emotional stakes that make a violation funny rather than flat, and safety filters often remove the edge entirely.

### Can AI help comedians write?

Yes, as a support tool. In a DeepMind study, comedians found LLMs useful for generating setups, premises, and structure, even while judging the actual humor weak. The productive pattern is to let AI handle the scaffolding and rapid ideation, then supply the punchline, voice, and edge yourself, rather than expecting it to be funny on its own.

### Why does AI humor all sound the same?

Because it is trained on everyone and averages toward the middle. A comedian's funniness comes from the idiosyncratic topology of their own mind, the specific connections only they would make, whereas an LLM blends all of those into a generic center. The result is competent, safe, and interchangeable, lacking the personal voice and lived specificity that distinctive comedy depends on.

### Will AI ever be genuinely funny?

It may keep improving at structured, pun-based, and parody humor, where patterns help. But the highest forms of comedy depend on a surprising, personally weighted violation of expectation drawn from lived experience and a distinct voice, which is exactly what current models lack. Until an AI has something like a specific, elastic, stakes-bearing mind to subvert from, its humor will likely stay clever at best and rarely surprising.

---

Source: https://buildfirstbrain.com/journal/why-ai-humor-fails-and-how-human-minds-map-it/
Author: Lawrence Arya — https://www.linkedin.com/in/vibecoding/