There is an assumption embedded in the widespread adoption of AI meeting transcription: that capturing everything that is said will capture everything that matters. The assumption is wrong in two distinct ways, and the second is more important than the first.
The first problem is that AI reads the surface account. What was said, when, by whom, in what sequence. That is useful for certain purposes — action items, decisions, who committed to what. But the diagnostic value of a meeting rarely lives in what was said. It lives in the discrepancy between what was said and what is actually happening. The gap between the stated problem and the felt one. The contradiction between the person giving the most confident account and the person who knows why that account is incomplete. A transcript captures the speech. It cannot capture the quality of the silence that followed it, or the significance of who did not speak.
The capacity to read that discrepancy is not analytical in the ordinary sense. It is built through repeated exposure to situations where understanding broke down — where what was being said and what was actually going on had quietly separated — and through learning to trust the sensing of that gap before you can articulate why. That capacity is not computable. It is not available to a system trained on text, because the thing it is reading is not in the text.
But the second problem cuts deeper. The most important things are often not in the transcript because they were never said. Not because people forgot, or because the meeting ran short. Because people knew they were being recorded.
The effect is not dramatic. Nobody refuses to speak. Nobody announces they are censoring themselves. What happens is subtler and more consistent: people edit toward the version they are comfortable having on record. The candid aside. The admission that the project is in worse shape than the update suggests. The question that implies the strategy might be wrong. The thing said in the corridor after the meeting ends, quietly, to one person. These are not peripheral to organisational intelligence. They are often where the actual intelligence is — the things people know but have not been able to say in a format that will travel.
Mass transcription changes the ecology of organisational speech. It does not just capture what people say; it shapes what they are willing to say. The meeting becomes more managed. The official account becomes more official. The gap between the transcript and reality — which diagnosis tries to read — quietly widens.
The irony is precise. Organisations adopting AI transcription in pursuit of better knowledge capture may be degrading the quality of the knowledge available to be captured. More data, from a more carefully managed source. More context that is further from the thing it claims to describe.
Erving Goffman described the distinction between frontstage and backstage behaviour — the difference between how people present themselves when they know they are being observed and how they act when they are not. The meeting was already a frontstage environment. Recording it makes it more so. What gets lost is not the performance; that was already there. What gets lost is the residual candour that leaked through at the edges.
This is also why the diagnostic conversation is not a meeting. It is deliberately different in structure, in the position of the person holding it, and in what it promises about what happens to what is said. The conditions that allow someone to say the thing they have not been able to say in a formal context are not accidental. They are specific. And they are not reproducible by recording the meeting and running the transcript through a model.
What you do not know is shaped by what you are able to hear. A system that changes what people say cannot then be used to understand what people actually think.
Further reading:
↳ Repetition and revelation — on how the capacity to read discrepancy is built, and why it cannot be shortcut
Garden notes
- Repetition and revelation — the capacity to sense discrepancy before articulating it is built through experience, not training data.
- The dimensions of not knowing — perspectival ignorance: what is not said is often not said because the conditions for saying it do not exist.
- Separated knowledge — what people know but cannot say in formats that travel is a structural feature of organisations; transcription may deepen the separation rather than bridge it.
- The mirror, the map and the breath — the broader argument about AI reshaping how humans understand themselves and, here, how they speak.
- Internal distortion — meetings are already shaped by relational history and hierarchy; recording adds another layer of managed presentation.
- Contextual excess — more accumulated data does not mean clearer understanding; this is an extreme case of that pattern.