Can AI meeting summaries make things up?

Yes, and this is the failure mode worth guarding against, because it doesn't look like a failure. Summaries are written by language models, which produce fluent text whether or not the underlying transcript supports it — so a garbled passage can come back as a clean, confident sentence, and a tentative "we could maybe push the date" can be flattened into "the date was pushed." Outright invented facts are rarer than over-confident paraphrase, but both read identically on the page. The mitigation is structural rather than a matter of picking a better model: use a tool that lets you drop from the summary back to the underlying words, and check anything you plan to quote, forward, or act on.

Why does my AI notetaker get names and speaker labels wrong?

Because deciding who spoke is a separate problem from deciding what was said. Splitting audio into speakers — speaker diarization — is hard when people talk over each other, when several voices sound similar, or when someone joins from a conference room where three people share one microphone. Tools that join the call as a bot can sometimes get speaker names handed to them by the meeting platform, which is a genuine advantage; tools that capture your computer's audio hear one mixed stream and have to infer it. Proper nouns are the other common miss: unusual names, internal project codenames, and product acronyms are exactly the words a general speech model has the least reason to expect.

Are live summaries less accurate than notes generated after the call?

At any given moment, a live summary knows less — it has only heard the meeting so far, so a point that gets reversed in the last five minutes will be wrong until it isn't. A summary written after the call sees the whole transcript at once and doesn't have that problem. The honest trade is that a rolling summary revises as the meeting continues and is checkable while the room is still there to correct it, whereas a post-meeting note is fixed at the moment the context needed to verify it has evaporated. A final full-call summary is built from the complete transcript either way.

How accurate are AI meeting notes?

Accurate enough to rely on for recall, not accurate enough to quote from without checking. The useful move is to stop asking about “accuracy” as one number, because AI meeting notes have two accuracies that fail in completely different ways — and only one of them is the one that will embarrass you.

Two accuracies, not one

Every AI notetaker runs the same broad pipeline, described in full in how AI meeting notetakers work. Two stages of it can be wrong:

Transcription accuracy — did it hear the right words? This is the number vendors advertise. It’s usually good, and more importantly it fails visibly: bad transcription looks like nonsense, so you notice.
Summarization accuracy — did the notes represent the meeting? This is the number nobody advertises, and it fails invisibly: a wrong summary is a well-written paragraph that happens not to match what people said.

Treat published accuracy percentages with the same skepticism you’d apply to any benchmark. They’re measured on recorded speech under good conditions, not on your 9am call with two people on speakerphone, one dog, and someone’s slide deck audio bleeding in.

What actually moves transcription accuracy

Ranked roughly by how much damage each one does:

Crosstalk. Two people speaking at once is the single most reliable way to lose a line. It also breaks speaker attribution at the same time.
Microphone and connection quality. A laptop mic across a room, a weak connection, or heavy platform-side compression all remove information before any model sees it.
Proper nouns and jargon. Names, internal codenames, and acronyms are the words a general speech model is least equipped to guess — and, annoyingly, the words most likely to end up in an action item.
Accents and code-switching. Accuracy varies by accent, and it drops further when speakers move between languages mid-sentence.
Where the tool taps the audio. A recorder sitting in the room hears the room, with all its echo. A tool doing system-audio capture reads the stream your computer is already playing, so remote voices arrive as cleanly as they reached your speakers.

Worth knowing about live tools specifically: what appears on screen a fraction of a second after someone speaks is a first guess. Streaming transcription emits interim results that get revised as more audio arrives — so a word that looks wrong mid-sentence has often fixed itself by the end of the sentence. That’s the system working, not failing.

Where summaries actually go wrong

Transcription errors are annoying. Summary errors are the ones that reach other people. In practice they cluster into four kinds:

Omission. The most common by far. A decision made in eight seconds between two longer tangents is exactly the kind of thing that doesn’t survive compression — and unlike a mangled word, a missing decision leaves no trace to notice.
Flattened hedging. “We could probably move the date if procurement agrees” becomes “The date will move.” Summaries prefer clean declaratives, and meetings are mostly conditionals.
Misattribution. The right commitment recorded against the wrong person. This inherits every error from speaker diarization underneath it, which is why action-item detection is only ever as good as the speaker labels feeding it.
Confident paraphrase of something unclear. Where the transcript is garbled, a language model will still write a clean sentence. Fully invented facts are rarer than people fear; over-confident smoothing is routine.

None of this is a reason to skip AI notes — the honest comparison isn’t against a perfect record, it’s against whatever you’d have typed yourself while also talking, which omits far more.

The check that takes thirty seconds

You don’t need to re-watch anything. Verify the parts that carry consequences:

Anything with your name on it. Read only your own action items and confirm you agreed to them.
Anything you’re about to forward. A summary sent to people who weren’t in the room is a primary source for them, so it deserves one read-through first.
Numbers, dates, and names. These are simultaneously the highest-stakes tokens and the ones transcription is worst at.
Anything surprising. If a note says something you don’t remember happening, that’s the flag. Drop into the transcript at that point rather than trusting the paraphrase — which is only possible in tools that keep the words behind the summary. See live transcript vs live summary for why both layers matter.

Where Canary sits on this

Canary is a real-time, bot-free meeting summarizer. It captures your computer’s system audio (no bot in the call, no plugin, no virtual audio device) and shows a live, multi-resolution rolling summary — from what’s being said right now to the whole call — so you can catch up the instant your name is called. It runs on macOS, Windows, and Linux.

Two things follow for accuracy, one in each direction.

In Canary’s favour: because the summary is on screen during the meeting, it’s checkable at the only moment checking is cheap. If the “now” pane says a decision went one way and it didn’t, you’re in the room and can say so. Post-meeting notes get audited hours later, if at all, by which point nobody remembers well enough to correct them. That’s the real argument for live notes — not that the model is better, but that the error has a shorter half-life. The multi-resolution view helps too: a line that looks wrong at the two-minute resolution can be checked against the full-call view without leaving the summary.

Against it, honestly: capturing system audio means Canary hears one mixed stream, the same one you hear. A tool that joins as a meeting bot can sometimes get speaker identities from the meeting platform directly, which is a genuine advantage for attribution — and if your audio is poor at your end, it’s poor for Canary too. And as covered above, a rolling summary is provisional by construction: it reflects the meeting so far, and something settled in the final minute will only appear once it’s been said. Tools like Otter and Notta come at this from the transcription-first direction, which is a reasonable preference if a verbatim record matters more to you than knowing where the conversation is right now.

Bottom line

AI meeting notes are reliable for recall and unreliable for quotation. Transcription is mostly a solved problem on decent audio and degrades in ways you can see; summarization is the layer that quietly loses things, and no vendor’s accuracy percentage describes it. Pick a tool that keeps the transcript reachable underneath the summary, check the handful of lines that carry consequences, and — if you can — read the notes while the meeting is still running, when a correction costs one sentence instead of a follow-up thread.