AI meeting assistants now join calls, transcribe everything and email a summary before anyone has left the room. Two things are well evidenced: automatic speech recognition still makes systematic errors, and the summarisation layer on top of it can invent content. What is not evidenced, because almost nobody has studied it, is whether having a recorder in every meeting changes how people pay attention, speak up or follow through. Anyone telling you it does, in either direction, is guessing.
What the products say they do
Otter.ai’s pricing page (accessed 14 September 2026) lists a free tier with 300 monthly transcription minutes and a Pro plan at $8.33 per user per month on annual billing, and states that the product joins Zoom, MS Teams and Google Meet to automatically write and share notes. Granola takes a different approach: its site says no bot joins your video call, audio is transcribed from the computer directly, and audio is deleted after the text is processed.
These are vendor descriptions of features, not evidence of accuracy. None of the three publishes an independently verified error rate for its transcripts or summaries.
The transcription layer: measured, and imperfect
Automatic speech recognition (ASR) is the foundation everything else sits on, and its failure modes are documented in peer-reviewed work. A 2020 study in PNAS tested five commercial ASR systems from Amazon, Apple, Google, IBM and Microsoft on interviews with 42 white and 73 Black speakers, and found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers. The gap persisted on identical phrases, pointing to the acoustic models rather than vocabulary. Systems have improved since 2020, but the study established that error is not evenly distributed across a room.
A second failure mode is worse than mishearing. A 2024 arXiv preprint (later presented at the ACM FAccT conference) evaluated OpenAI’s Whisper and found that roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences which did not exist in any form in the underlying audio, and that 38% of those hallucinations included explicit harms such as invented violence or false attributions. Hallucinations clustered around pauses and silence. One per cent sounds small until you multiply it by every meeting in a company, and the errors are not flagged as uncertain; they read as confident transcript.
The summarisation layer: a second place to go wrong
Meeting summaries are generated by a language model reading the transcript, so they inherit the transcript’s errors and add their own. A 2024 arXiv preprint on meeting summarisation notes that large language models still struggle to maintain relevance and avoid hallucination, and releases a dataset of 200 automatically generated meeting summaries annotated by humans on nine error types, including structural, omission and irrelevance errors. That is a preprint, and its purpose was to build a correction pipeline, but the annotated error taxonomy is itself the point: omission (a decision that was made but not recorded) and irrelevance (a tangent promoted to an action item) are quiet errors that nobody notices until a deadline is missed.
The practical consequence is that a summary is a claim about what happened, produced by a system that cannot tell you which lines it is unsure of.
Attention: what we know, and what we do not
The hope behind these tools is that people can stop taking notes and pay attention. There is no published trial of whether that happens. The nearest evidence comes from note-taking research, and it cuts both ways.
A widely cited 2014 study found that, in three experiments, students who took notes on laptops performed worse on conceptual questions than students who wrote longhand, attributed to laptop users transcribing verbatim rather than processing. A 2021 direct replication with a larger sample did not find that difference, and its authors concluded the results do not support the idea that longhand note taking improves immediate learning via better encoding. So the strongest claim you can make is that we do not know whether writing your own notes helps you remember a meeting, which means we also do not know whether letting a machine do it costs you anything.
What is solid is that attention is fragile. In a 2014 experiment with sequential tasks, interruptions averaging 4.4 seconds tripled the rate of sequence errors, and interruptions under three seconds doubled them. That is a lab task, not a meeting, but it is a reasonable caution against the assumption that a live transcript scrolling in a side panel, or a “catch-up” summary pinging mid-call, is free of cost.
Participation and the recorder in the room
Whether people speak differently when they know they are being recorded, transcribed and summarised for absent colleagues is an obvious question with, as far as we could find, no controlled answer for AI notetakers specifically. It is plausible that some people self-censor and others perform for the transcript. It is also plausible that nothing changes once the novelty wears off. We are not going to pretend either way.
Two points do follow from the error evidence. First, if ASR error is higher for some accents and speech patterns, the transcript systematically under-represents those speakers, and a summary built from it will too. Second, hallucinations cluster around silence and disfluency, which is where the hesitant contributions live. The people most at risk of being misquoted by the machine are not a random sample.
Follow-through
A summary with action items is only useful if the action items are right. The summarisation error taxonomy above lists omission and irrelevance as recognised failure types, and there is no independent data on how often commercial tools get the action list right. The honest position is that an AI-generated action list is a draft that someone who was in the meeting should confirm within the hour, while memory is fresh.
The bottom line
Transcription error is real, measured, and uneven across speakers. Summary hallucination is real and mostly silent. The behavioural effects on attention, participation and follow-through are unstudied. Use the tools for search and recall, treat every summary as unverified, and do not let “the AI has it” replace someone writing down the decision.
Common questions
How accurate are AI meeting transcripts?
Accuracy varies by speaker and audio quality, and vendors do not publish independently verified figures. A 2020 peer-reviewed test of five commercial systems found average word error rates of 0.35 for Black speakers and 0.19 for white speakers. A 2024 study of Whisper found about 1% of transcriptions contained fabricated sentences. Expect errors, and expect them to be uneven.
Can AI meeting summaries make things up?
Yes. The summary is generated by a language model from a transcript that may itself contain fabricated text, and research on meeting summarisation identifies hallucination, omission and irrelevance as recognised error types. The output does not flag which lines are uncertain, so treat action items and attributed statements as unverified until someone who attended confirms them.
Does letting AI take notes help you pay attention?
Nobody has tested this for AI meeting assistants. Note-taking research is mixed: a 2014 study found handwriting beat typing for conceptual recall, but a larger 2021 replication found no difference. Separately, very brief interruptions of a few seconds measurably increase errors on sequential tasks, so live transcripts and mid-meeting summaries may carry a cost.
Is it better to use a tool that joins the call or one that records locally?
The evidence does not distinguish them on accuracy; both rely on similar speech recognition. The difference is visibility and data handling: Otter’s bot joins as a participant, while Granola states it captures computer audio with no bot and deletes audio after transcription. Which you prefer is a consent and privacy judgement, not an accuracy one.
Did this answer your question?
Thanks. That helps us decide what to fix.
Leave a comment