← Back to all articles Communication

What You Say Is Never What You Mean: Why AI Still Struggles With Human Communication

Here's one of the oldest observations in linguistics, and one of the most useful: what you say is almost never exactly what you mean. Human communication doesn't run on dictionary definitions. It runs on implication, shared context, tone, timing, and a bunch of small social calculations that happen in a split second, mostly without either person noticing.

This is the domain of pragmatics: the branch of linguistics concerned not with what words literally mean, but with what speakers actually do with them. And it turns out to be exactly the place where AI language systems, however fluent they sound, still show their seams.

The cooperative principle, and why we bother being indirect at all

In the 1970s, the philosopher of language Paul Grice proposed something called the Cooperative Principle. The basic idea: conversation works because both speakers assume the other person is trying to be helpful, truthful, relevant, and clear. And when a speaker seems to break one of those rules on purpose, the listener doesn't get confused. They dig for the meaning underneath.

If a colleague asks "Do you know what time it is?" and you answer "Yes," you've technically answered with perfect accuracy, and completely failed to communicate. Everyone understands this instantly. The question was never really about your knowledge of clocks. It was a request for the time, dressed up as a question because a flat command ("Tell me the time") can feel abrupt or rude.

This kind of indirectness isn't a flaw in human communication. It's kind of the whole point. We soften requests, hedge disagreements, and imply refusals instead of stating them, because directness carries a cost. It can threaten someone's dignity, create discomfort, or damage a relationship. Every language does some version of this. Readers of this site will recognize it right away from an earlier piece on Chinese business communication: an outright "no" is rare precisely because indirectness is doing real social work. It's not hiding meaning by accident. It's protecting something on purpose.

Where AI language models actually struggle

Modern AI language systems are trained on enormous amounts of human text, and they've gotten remarkably good at reproducing the surface patterns of polite, indirect, socially fluent language. They can hedge. They can soften. They can phrase a disagreement diplomatically. In that narrow sense, they've learned the choreography of pragmatics.

But choreography isn't the same as the reasoning behind it. Pragmatic competence in humans depends on a working model of what the other person already knows, wants, and would find embarrassing or upsetting in this specific moment, with this specific history, in this specific relationship. That's a constantly updating, deeply contextual judgment call. AI systems approximate it statistically, from patterns in training data, instead of actually tracking a real, ongoing relationship with another mind.

This produces some pretty recognizable failure modes:

• Over-literalism at the wrong moment. Asked an indirect question, an AI system may answer the literal question instead of the implied one. It's the linguistic equivalent of answering "do you know what time it is?" with a simple "yes."
• Missing culturally specific implicature. An indirect refusal that a Chinese business professional would read instantly, something like "we'll need to study this further," carries no reliable signal to a system trained mostly on different conventions, unless it's specifically learned that pattern.
• Over-smoothing disagreement. Because so much training data rewards agreeable, cooperative sounding language, AI responses can default toward excessive hedging or agreement, even when directness would actually be more useful and more honest.
• A certain uncanny evenness. Human writing is full of small imperfections, asides, and shifts in register that come from a real person's fatigue, mood, and personality in the moment. AI text can be pragmatically "correct" and still feel a little too smooth. That's a pattern this site's readers may already recognize from an earlier piece on detecting AI-generated Chinese text, where the real giveaway usually wasn't an error. It was suspicious perfection.

Why this matters more, not less, as AI becomes part of everyday communication

As AI systems increasingly draft emails, summarize meetings, translate messages, and mediate first contact in business communication, the gap between literal meaning and intended meaning doesn't disappear. It just moves somewhere else. A poorly calibrated AI translation of an indirect refusal can turn a polite "no" into something that reads, in the target language, like an enthusiastic yes. A summarization tool that flattens hedged, face-saving language into a blunt bullet point can strip out exactly the social information a human reader needed to read the message correctly.

None of this means AI communication tools aren't useful. They obviously are. It means something more specific: understanding pragmatics isn't a soft skill or a nice-to-have. It's the difference between technically accurate communication and communication that actually works, whether the person on the other end is from a different culture or a language model standing in for one.

What this means in practice

A few practical takeaways, whether you're working across a language barrier, a cultural barrier, or increasingly, an AI-mediated one:

• Don't assume literal accuracy equals successful communication. A technically correct translation or summary can still miss the actual point, if it strips out the implication the original speaker was relying on.
• Watch for flattened nuance in AI-assisted communication. If an AI tool is drafting or translating something sensitive, check whether the hedges, softeners, and indirect phrasing survived the process, or whether they got smoothed into something blunter than intended.
• Cultural context still needs a human who actually has it. Pattern matching can approximate a lot. But the specific, high stakes judgment calls, reading an indirect refusal correctly, catching a face-threatening moment before it happens, still benefit enormously from someone who's actually lived inside that communication system.

The oldest insight in pragmatics turns out to be the newest problem in AI. Language was never really about the words. It was always about what the words were doing. Any system, human or artificial, that forgets this will keep answering the wrong question. Cheerfully, fluently, and completely convinced it got it right.