Ask an AI for a joke: you'll get something passable, often a recognisable pun. Then send it an ironic message, without flagging it. Chances are it will take it at face value, or over-interpret and see irony where there is none. This gap is not anecdotal: it's one of the clearest indicators of what these systems actually do.
Humour is a textbook case because it concentrates just about all the difficulties of language. Let's see why.
Why producing a joke is easy
A joke has a structure. A setup, an expectation, a twist. Some forms are highly codified, and a model that has read millions of humorous texts reproduces these structures without difficulty.
That's pattern recognition, exactly what these systems do best, as we explained in our article on LLMs. The result is often technically correct and rarely funny, because a joke that resembles a thousand other jokes no longer has the element of surprise that made it work.
Why understanding irony is difficult
Here, three obstacles pile up.
Irony reverses meaning without changing the words. Saying "what a lovely day" in the rain means exactly the opposite of what is written. No element of the text carries this inversion: it comes from the situation. Yet the model only perceives the text.
Shared context is absent. Humour relies heavily on what two people know in common without saying it: a shared history, a culture, a recent event, the tone of a relationship. The model only has what is written in the conversation, and it has no anchor in the world, as we developed in our article on understanding.
Register is ambiguous. In writing, without intonation or facial cues, telling a joke from a sincere statement is sometimes difficult even between humans. That's indeed the source of most misunderstandings in messaging.
There's an additional factor, and it's structural. These models are trained to respond helpfully and avoid misunderstandings. But interpreting a message as ironic carries a risk: if you get it wrong, you seem not to take the speaker seriously. The safest behaviour, from a training standpoint, is therefore to treat statements at face value. This isn't an inability, it's an implicit choice installed during reinforcement learning from human feedback.
What this limitation reveals
Humour is interesting because it simultaneously demands three things these systems possess unevenly.
You need mastery of language, which they do remarkably well. You need knowledge of the world, which they do partially, through what has been written about it. And you need to share a situation with someone, which they don't do at all.
A witty remark works because it hits something at the right moment, with the right person. It is dated, situated, relational. That's precisely the dimension these systems lack: they have no moment, no situation, no relationship.
What to take away
This limitation is useful for calibrating what we expect from these tools. They excel at what is expressed in the text, and are weak on what depends on what isn't there.
Concretely, two pieces of advice. If you want an AI to grasp an ironic intent, say so explicitly: the gap disappears as soon as it's named. And if you ask it for humour, expect correct structure rather than wit, and use it as a starting point to rework.
There's something rather reassuring about this limitation. Making someone laugh requires knowing who they are, what moves them, and what they won't say. That may be the finest social skill we have, and it holds up rather well.