For most of the history of workplace meetings, speech was ephemeral. What someone said in a room existed afterward only in the memories of the people present and in whatever notes they chose to take.
That condition ended quickly. Automatic transcription is now a default feature of the major meeting platforms, and a meeting in 2026 is more likely than not to produce a searchable, attributed, shareable record of who said what.
The methodological literature has a long-standing interest in what happens to behaviour under observation.[2] The workplace question is narrower and newer: when participants know that an AI system is transcribing and summarising, does the way they speak change?
The available evidence
The largest dataset addressing this is proprietary. Read AI, which sells meeting analytics, published an analysis in January 2026 drawing on 159,870 virtual and hybrid meetings across more than 30 industries.[1] Its headline finding concerns the distribution of speaking time by gender: with an AI notetaker present, women contributed approximately 9% more than men, against a reported baseline in which women speak roughly 25% less than men.
If that comparison holds, the presence of a transcription system does not simply record the meeting. It changes who talks.
What to hold in mind about this finding
The publisher sells the product being studied. Read AI's analysis uses its own customers' meetings and supports a favourable conclusion about its own category. That does not make it wrong, and the disclosed sample size is unusually large for a vendor study, but it has not been independently replicated.
The comparison groups are unclear from the published summary. Meetings with a notetaker and meetings without one may differ in type, formality, size, or organisation. Without knowing whether the comparison is within-organisation or across, selection cannot be ruled out.
"Contributed more" requires definition. Speaking time, turn count, and word count are different measures that can move in different directions.
Self-reported mechanism is absent. The proposed explanation, that participants become more conscious of how much they are talking when a record is being made, is plausible and untested.
Why the direction is plausible anyway
Two independent lines of research make a behavioural shift under recording unsurprising.
First, disfluency rates are known to vary with the public or private character of a speech setting, and with the speaker's role and relationship to the listener.[3][4] Speech is adjusted to audience as a matter of course. A recorded meeting has a larger and less defined audience than an unrecorded one, since the transcript may be read by people who were never in the room.
Second, survey evidence from 2026 describes workers as acutely conscious of how their communication will be perceived and reviewed, with 44% reporting that they sometimes freeze in person because they cannot review or edit their words first.[5] A population already sensitive to the reviewability of its communication is a population likely to respond when speech becomes reviewable.
What has not been measured
The gender-distribution finding, if real, is the most socially interesting result in this area. It is also the only one that has been reported at scale. Several adjacent questions appear to be untouched:
- Does recording change disfluency? Filled-pause rates are the standard measure of processing load and self-monitoring in speech. Nobody has published a comparison of disfluency rates in recorded versus unrecorded meetings.
- Does recording change candour? Practitioner accounts suggest participants become more careful and less willing to speculate aloud when a transcript exists. This is anecdotal.
- Does it change meeting outcomes? More equal speaking time is treated as a benefit. Whether it produces better decisions is a separate empirical question.
- What happens to the unrecorded conversation? If candour moves out of recorded meetings, it moves somewhere. A displacement effect would be invisible to any analysis conducted on transcripts.
The fourth is the structural problem in this area. Every large dataset of workplace speech is now assembled by transcription products, so the corpus that could answer the question is generated by the very condition under study. There is no straightforward control group, because the meetings without a notetaker are the meetings nobody has a record of.
Implications for measurement
Anyone measuring workplace speech from meeting transcripts is measuring recorded speech specifically. The resulting figures describe how people talk when they know they are being transcribed, which may not generalise to speech in unrecorded settings.
The same caution applies to our own published corpus work. Practice sessions conducted inside an application are recorded by definition, and the figures we report for answer length and filler rate[6] describe speech produced under known recording. That is a limitation of the measurement instrument rather than a flaw in the data, but it belongs stated rather than buried.