AI Can Beat World Champions at Chess—But Still Struggles With a Simple Joke

In 1997, a machine beat the world chess champion. Today's engines are unbeatable. But the same technology still can't reliably tell you why a joke is funny.

In 1997, a machine beat the reigning world chess champion. Today's chess engines are so far past human ability that grandmasters use them mainly to study their own mistakes. And yet the same class of technology — artificial intelligence — still regularly misreads a knock-knock joke, misses the point of sarcasm, or explains a pun so literally it kills the joke. That's not a contradiction. It's a clue about what "intelligence" actually is: not one skill, but many separate ones, and machines have only mastered a few of them.

Why Chess Was Never the Hard Part

Chess looks intellectually demanding because humans find it hard. But from a computational standpoint, chess is one of the most forgiving problems you could hand a machine. The board has 64 squares. There are six piece types. Every rule is fixed and fully visible to both players — no hidden information, no ambiguity, no cultural context required. And there's exactly one measurable goal: checkmate.

That combination — fixed rules, full information, a single scorable outcome — is precisely the kind of problem search algorithms excel at. Deep Blue, the IBM system that defeated Garry Kasparov in 1997, worked by brute force: evaluating up to roughly 200 million board positions per second, guided by evaluation functions that human chess experts had hand-coded into the system.

The Shift From Brute Force to Self-Taught Mastery

Two decades later, DeepMind's AlphaZero took a fundamentally different approach — and made the point even more clearly. Instead of relying on human-crafted chess knowledge, AlphaZero was given only the rules of the game. It then played millions of games against itself, using a technique called Monte Carlo Tree Search paired with a neural network that learned to evaluate positions from experience rather than hardcoded rules. Within about four hours of self-play, it was already outperforming Stockfish, which had been the strongest engine built using traditional, human-tuned methods.

This is the mental model worth holding onto: mastering chess is like solving an enormous equation. There's a right answer, and enough computation will find it. Understanding humor is a different kind of problem entirely — closer to understanding another person's life.

What Makes Humor a Completely Different Problem

A joke doesn't have fixed rules. It doesn't have one correct outcome — what's hilarious to one audience can be baffling or even offensive to another. Landing a joke requires tracking shared cultural references, timing, tone, and the listener's expectations, then deliberately subverting those expectations in a way that feels surprising rather than confusing. None of that reduces to a single number an algorithm can optimize for, the way checkmate does.

Humor researchers have found that humor itself splits into distinct functional types — jokes that build relationships, jokes that boost the teller's own image, jokes at someone else's expense, and self-deprecating jokes — each relying on different social instincts to pull off. It's part of why direct studies of AI vs. human creativity keep landing on the same conclusion: reproducing all of that convincingly is a fundamentally messier task than searching a game tree.

The Paradox Hiding Inside "AI Solved Common Sense"

Here's where the story gets genuinely counterintuitive. In 2012, researchers Hector Levesque, Ernie Davis, and Leora Morgenstern introduced the Winograd Schema Challenge — a test built around sentences where a pronoun's meaning depends entirely on real-world knowledge, not grammar. Their argument was that only a system capable of authentic, "full-bodied" understanding could pass it.

By the early 2020s, large language models had statistically defeated the challenge. That sounds like proof AI achieved common sense. It isn't. The same researchers who built the test later acknowledged that solving it didn't mean the underlying "full-bodied" understanding had actually arrived — the benchmark turned out to be a narrower proxy than its designers intended, one that pattern-matching at massive scale could satisfy without the grounded reasoning behind it. It's the same context comprehension problem showing up again: beating the test and possessing the capability the test was built to detect turned out to be two different things.

So Can AI Actually Tell a Joke?

Somewhat, and this is the part most coverage oversimplifies in either direction. Under constrained, short-form prompts, studies comparing AI-generated jokes to human-written ones have found comparable humor ratings — AI isn't hopeless at humor. But other research on ChatGPT found the model leaning on a narrow, almost hard-coded set of jokes rather than generating fresh material, and models generally struggle once humor requires longer-form timing, an original angle, or culturally specific context. The pattern that emerges: AI can remix and recombine humor it has seen, but it isn't reliably generating humor from an understanding of why something is funny — a limitation that belongs on the same list as most other things AI cannot do right now.

Chess vs. Humor, Side by Side

Put the two problems next to each other and the gap becomes obvious. Chess has a fixed, fully-known rule set; humor's rules are unwritten and shift by culture. Chess has one objective outcome — win, lose, or draw; humor's success is subjective, funny to one audience and confusing to another. Chess engines improve through self-play against a single win/loss signal; language models have no equivalent feedback loop for "did that land" — the same missing correction signal behind why models sometimes confidently state false facts. And where chess has been fully solved at the superhuman level since AlphaZero in 2017, humor generation remains competitive only on short, constrained jokes, and inconsistent on anything original or context-heavy.

What This Actually Reveals About Machine Intelligence

The common belief is that intelligence is one thing — if a system beats the smartest humans at one hard task, it should dominate every other hard task too. The evidence points somewhere more accurate: intelligence is a bundle of separate, specialized abilities. Search and optimization is one of them, and machines mastered it decades ago. Grounded, common-sense understanding of a shared human world is a different ability entirely, and it's the one still largely unsolved — worth remembering the next time a headline claims artificial general intelligence is closer than expected.

Whether more data and bigger models eventually close that gap is genuinely unresolved. Some researchers expect scale to keep chipping away at it. Others argue that common sense requires embodied experience in the physical world — the kind text alone can't teach — meaning no amount of additional reading will fully bridge the gap. Both camps agree on one thing: passing a benchmark and possessing the capability the benchmark was designed to measure are not always the same achievement.

Frequently Asked Questions

Why can AI beat the best chess players but not understand jokes?

Chess has fixed rules and one measurable goal — winning — so an AI can improve by searching millions of positions and scoring outcomes. Humor depends on shared cultural context, timing, and lived experience, which can't be reduced to a single scorable objective the way a checkmate can.

Is chess actually a good test of intelligence?

Not on its own. Chess is a closed, fully-known system, which makes it ideal for search algorithms but a poor proxy for general reasoning. Games like chess were historically used as AI benchmarks partly because they're easy to score, not because mastering them proves broad understanding.

Has any AI ever passed a real common-sense test?

Large language models have technically defeated the Winograd Schema Challenge, a benchmark designed to test common-sense sentence disambiguation. However, the test's own creators note that passing it doesn't mean the models possess the grounded, human-style reasoning the challenge was originally built to detect.

Can AI write jokes at all?

Yes, within limits. Research comparing GPT-generated jokes to human-written ones under constrained, short-form prompts found comparable humor ratings. But studies also show models fall back on a narrow set of familiar joke templates and struggle with longer-form, original, or culturally specific comedy.

What is common-sense reasoning in AI, exactly?

It's the ability to draw on unstated, everyday background knowledge to interpret a situation correctly — knowing a dropped glass will likely break, or that "she" in a sentence refers to the tired person, not the suitcase. It's implicit for humans and difficult to encode for machines.

Will AI ever develop genuine common sense?

Researchers are actively split. Some argue scale and better training data will close the gap; others argue common sense requires embodied experience in the physical world that text-only training can't provide. As of 2026, no model has been shown to reliably match human-level common-sense reasoning across novel situations.