You tell a customer-service bot what you need, it responds with something almost related, you rephrase, it offers a menu of options that don’t include yours, and three exchanges later you’re typing “REPRESENTATIVE” in all caps. Compare that to an actual misunderstanding with a human: you say something ambiguous, they raise an eyebrow or say “wait, do you mean—,” you clarify, and the conversation moves on in about two seconds. Human conversation is full of small breakdowns too. The difference is that humans fix them almost invisibly, and machines usually can’t fix them at all.
Repair: The Quiet Machinery of Human Talk
Researchers who study real, recorded conversation — a field with roots in the sociological method called conversation analysis, pioneered by Harvey Sacks, Emanuel Schegloff, and Gail Jefferson — noticed decades ago that misunderstanding isn’t rare in everyday talk. It’s constant. What’s remarkable isn’t that people avoid it; it’s how quickly and smoothly they notice it and fix it, usually within a turn or two, through what conversation analysts call repair: a clarifying question, a rephrasing, a “sorry, what?”, a correction offered by the other person before you even finish your sentence.
This machinery runs so fast and so automatically that fluent speakers barely register it happening. It’s only when you watch it fail — in a bad phone connection, a language barrier, or a conversation with a machine — that you notice how much ongoing repair work a normal conversation actually requires.
Mutual Intelligibility Is Built, Not Assumed
Researcher Lucy Suchman, whose work on situated action and human-machine communication drew directly on this conversation-analytic tradition, argued that mutual intelligibility — the sense that two parties understand each other — isn’t a state that communication starts in. It’s something the parties actively build and continuously repair together, moment to moment, using pragmatic cues, shared background knowledge, and constant small checks that everything so far has landed correctly.
Two people having a conversation are both doing this work simultaneously and reading each other’s signals — hesitation, a confused expression, a slightly-off response — as evidence that repair is needed. That’s a genuinely two-sided, real-time collaboration, not a message passed successfully from one head to another in a single shot.
The Interactional Asymmetry
This is where human-machine conversation runs into a structural wall rather than a solvable rough edge. A machine responding to you isn’t a second party doing that same collaborative repair work — it’s producing output based on its model of what you likely meant, without a genuine, moment-to-moment read on whether you’ve actually understood, or whether it has actually understood you. Suchman described this gap between how two humans monitor and repair each other’s understanding, versus how asymmetrically a human and a machine relate, as one of the central design problems in building systems that communicate.
The bot in a customer-service chat isn’t failing to repair the misunderstanding because its language model is too small. It’s failing because it isn’t participating in the interaction the way a second conversational partner does — it has no independent access to whether you’re actually following, only to patterns in the text you’ve typed.
Why More Data Doesn’t Fully Fix This
It’s tempting to assume that today’s much larger language models have closed this gap simply by getting better at predicting what people mean. They’ve certainly gotten better at producing plausible, well-formed responses. But the deeper problem Suchman described isn’t about the quality of the response — it’s about the absence of a live, shared situation that both parties can jointly monitor. Work exploring what large-scale AI reveals about the limits of language learned from text alone keeps finding the same pattern: the parts of communication that depend on real-time, embodied, situational grounding remain the hardest parts to replicate, no matter how much text a system has processed.
What This Means for You
When a conversation with a machine goes sideways and no amount of rephrasing fixes it, that’s not a personal failure to phrase things clearly enough — it’s the absence of the repair machinery that makes ordinary human conversation work, running quietly and constantly beneath words you never even notice being exchanged. The final article in this series looks at the research tradition that first made that invisible machinery visible: a branch of sociology called ethnomethodology, and what it has to do with talking to machines at all.