You ask a voice assistant to “call the restaurant we went to last night,” and it asks you to repeat yourself, then offers to search the web for “restaurant.” You know exactly which restaurant. It has access to your location history, your calendar, your call log. And it still doesn’t get it.
The easy explanation is that the software just isn’t smart enough yet — that next year’s model, or the one after that, will finally close the gap. But this exact frustration was being formally studied before most personal computers had a graphical screen, and the person who studied it most carefully concluded that the problem isn’t a missing feature. It’s a mismatch between how machines are built to process language and how people actually use it.
The Planning Model of Communication
Early artificial intelligence and cognitive science borrowed a tidy model of human action: a person forms an intention, builds a plan to achieve it, and then executes that plan step by step. Communication, on this view, works the same way — you have a message in mind, you encode it into words according to a plan, and the listener decodes it back into your original intention. Understanding is a matter of successfully passing a plan from one head to another.
This model is appealing partly because it’s exactly how a computer program works: a defined goal, a sequence of steps, a predictable outcome. If human communication worked this way too, building a machine that understood language would mostly be an engineering problem — get the steps right, and understanding follows.
The Trouble With Plans
In 1987, researcher Lucy Suchman published Plans and Situated Actions, a study built on close observation of people trying to use an early “intelligent” office machine. Her conclusion challenged the planning model directly: real human action, she argued, is not the execution of a pre-formed plan. It’s improvised, moment to moment, in response to the specific, unrepeatable details of a particular situation. Suchman’s own phrase for this was that plans are “a resource for action,” not a blueprint that determines it — something people refer to before or after acting, much more than something they consult line by line while acting.
Communication works the same way. Two people talking don’t successfully exchange a fixed plan; they build shared understanding together, in real time, using an enormous amount of context that never gets spoken aloud.
What Machines Actually Miss
That unspoken context is where most of the trouble lives. When you say “the restaurant we went to last night,” the phrase carries no meaning on its own — its meaning depends entirely on who’s asking, what happened last night, and what both of you already know. Linguists call this kind of context-dependence part of pragmatics: meaning that isn’t in the words themselves but in how they’re used, by whom, and in what situation.
A machine can be extremely good at semantics — the general meanings of words and how they combine — while still having almost no access to the situated, personal context a listener would need to resolve a phrase like that. It isn’t failing to look up the right definition. It’s missing the shared history a real conversational partner would draw on automatically.
It’s Not Just an AI Problem
It’s worth noticing that Suchman’s foundational case study wasn’t a chatbot or a language model — it was a photocopier with an onscreen expert-help system, built in the 1980s. The same basic gap shows up decades later in systems built on entirely different technology, including the large language models behind today’s voice assistants and chatbots. Researchers studying why AI can struggle to actually understand language keep running into a version of the same wall: what AI reveals about the limits of purely textual learning is that a great deal of meaning depends on grounding in a specific, lived situation — not just on processing more words.
That continuity across forty years of very different technology is the strongest evidence that this isn’t a temporary engineering shortfall. It’s a structural feature of the problem.
What This Means for You
The next time a device seems to miss an obvious point, the honest description of what’s happening isn’t “it doesn’t know that word” — it’s “it doesn’t share the situation you’re both supposedly in.” That’s a much harder problem than vocabulary, and it’s the animating question behind an entire field of human-computer interaction research that predates the smartphone by a generation. If you want the fuller argument, Suchman’s original book, Plans and Situated Actions, is still the clearest statement of it — and the rest of this series works through its major ideas one at a time.