Do AI meeting notes make things up?

Sometimes, yes. But not usually the way people fear. AI meeting notes rarely invent a meeting that didn’t happen. More often they describe a meeting that was more decided than it really was. A “we could probably” turns into “the team agreed,” a suggestion picks up an owner, and “sometime next week” turns into a date. Each line reads fine on its own, which is the problem.

The useful question isn’t “does this tool hallucinate?” (they all can). It’s “where did this line come from?” There are three possible sources, and the transcript tells you which one you’re looking at.

Three places a made-up line can come from

Every AI notetaker runs roughly the same pipeline: capture the audio, transcribe it, work out who spoke, then have a language model write the notes. The full version is in how AI meeting notetakers work. Made-up content can enter at three points.

Where it started	What it looks like in the notes	What the transcript shows	What helps
The speech model invented words	A sentence or phrase nobody said, often from a quiet stretch, hold music or keyboard noise; sometimes a repeated line	The invented words are in the transcript too, at a moment when nobody was talking	Voice activity detection so silence never reaches the model; better audio
The speech model misheard	A confident note built on the wrong word: “fifty” for “fifteen,” a codename turned into a common word, “can” for “can’t”	The wrong word is there, in a sentence that otherwise makes sense	Clean audio, custom vocabulary where offered, and checking numbers and names
The summary went beyond the transcript	A decision nobody made, an owner nobody volunteered, a date nobody said, a consensus that was really a disagreement	The right words are there, but they don’t say what the summary says	Checking the summary against the transcript, and tools that leave fields blank instead of guessing

The first two are transcription problems, and a word error rate partly measures them. (Invented words in silence count as insertions, which is one reason a published WER can look better than your messy call.) The speech-model hallucination is best known from open models such as Whisper, which can produce a fluent sentence from a quiet stretch. That’s covered in Canary vs. Whisper. The third is a summary problem, and no transcription benchmark measures it at all. It’s the one that most often reaches other people, because a summary is what gets forwarded.

The quick test: find the moment in the transcript. If the doubtful words are there, it’s a transcription problem (or someone really said it). If they aren’t, the summary wrote them.

Why summaries add things

Meeting summaries are abstractive: a language model writes new sentences rather than copying spoken ones. That means a summary can be wrong about what it says, not only about what it left out. The mechanics are in extractive vs. abstractive summarization. Three pressures make the “more decided than it was” failure common in meetings in particular.

• Meetings are tentative and notes are declarative. Spoken decisions are full of “probably,” “if,” and “let’s see.” A summary that kept all of them would read as mush, so the model tidies them up, and sometimes the condition goes with them.
• Templates ask for details the meeting didn’t contain. An action-item list with an Owner column and a Due column invites the model to fill both. If nobody said who or when, the most plausible name and date can appear anyway. A blank field is honest. A filled-in guess looks like a fact. This is the weak spot in action-item detection, and it’s worse when speaker diarization has put the commitment against the wrong voice.
• The model fills gaps in what it heard. A summary only knows its input, but it writes as if it knew the whole meeting. Anything the tool never heard can get papered over with a plausible guess: the ten minutes before you pressed start, a decision typed in the chat, a number that was only on a shared screen, or a passage lost to crosstalk. “As discussed, the team will proceed with the plan” is often a model referring to something it never heard.

The tells: what made-up content usually looks like

You don’t need to check every line. These are the ones worth a second look:

• Specifics nobody said. A calendar date where people said “Friday” or “next week.” A rounded or tidied number. A full name for someone the transcript only knows as “Speaker 2.”
• Owners on things nobody picked up. “Sam will send the deck” when the transcript has “someone should send the deck.”
• Consensus words. “Agreed,” “decided,” “aligned,” “confirmed.” Check that one person actually said yes.
• References with nothing behind them. “The plan,” “the issue raised earlier,” “as discussed,” where the transcript you have doesn’t contain the earlier part.
• Anything surprising. If you were there and don’t remember it, trust that instinct and look it up.

How accurate are AI meeting notes? covers the broader accuracy picture, including omissions, which are more common than invention and harder to spot.

When you check matters more than how

The same wrong line costs very different amounts depending on when someone notices it.

• During the meeting, a doubtful line costs one sentence: “Just to confirm, are we committing to Friday, or only if QA clears it?” Everyone who knows the answer is still there.
• Right after, it costs a scroll through the transcript, if your tool keeps one, and some memory that’s already fading.
• A week later, after the notes were forwarded to people who weren’t there, it costs a correction thread, and only if someone notices. For readers who weren’t in the meeting, the notes are the only account they have.

That’s the strongest case for reading notes live. It isn’t that a live summary is less likely to be wrong. At any moment it has heard less of the meeting, and something settled in the last minute isn’t in it yet. The case is that its mistakes get noticed while fixing them is cheap. A live transcript and a live summary also complement each other here: the transcript holds the exact words, and the summary shows you where to look. Canary vs. Tactiq compares the two approaches. If you need a verified, word-for-word record instead, that’s a job for a transcript with a human check. Canary vs. Rev covers that trade.

Shared notes carry other people’s words

There’s a transparency side to this. When AI notes go to the team, each line is attributed, explicitly or implicitly, to someone who was there. A made-up commitment isn’t only an error. It puts words in a colleague’s mouth. Two habits help. Tell people the notes are AI-written (how to tell participants you’re using an AI notetaker), and give anyone named in them an easy way to say “that’s not what I said” before the notes become the record. That matters most for formal minutes, which people treat as settled.

Where Canary fits

Canary is a real-time, bot-free meeting summarizer. It captures your computer’s system audio (no bot in the call, no plugin) 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.

Everything above applies to it. Its summaries are written by a language model from a streaming transcript, so they can be wrong in the same three ways. What Canary does about that, specifically:

• Owners and dates are left blank when nobody stated them. Canary’s action-item step asks the model to omit the assignee and deadline if they weren’t stated, and to return no action items at all if there weren’t any. Its decisions step asks for commitments the team agreed to, not discussions or open questions, and returns an empty list if there were none. Those are instructions to a model, not guarantees, but they target the most common invented details.
• Silence isn’t summarized. If a time window has no transcript text, Canary doesn’t send it to the model, so a quiet stretch doesn’t produce a summary of nothing.
• You read it while the meeting is running. The narrow windows (about the last 30 seconds, the last 2 minutes) stay close to what was just said, so a wrong line shows up while you can still ask.

Where it’s weaker, honestly:

• Speakers are numbered, not named, because Canary hears one mixed audio stream. Treat who-said-what in the summary as approximate.
• Summary lines don’t link back to the transcript. After the call, the review view has the full timestamped transcript, which you can search. That’s where to check a number, a name or an “if.”
• On long meetings, the full-call view, decisions and action items are built from the most recent part of the transcript, roughly the last 10–15 minutes, so earlier discussion may be missing from them. Don’t read their silence about an early topic as “it didn’t come up.”
• It only hears the audio. The chat, the shared screen and anything before you started capture aren’t in its input. Some tools that join as a meeting bot can see the shared screen, which closes part of that gap.

If you’d rather have exact words than a gist, use a transcript-first tool. If what you need is to know where the conversation is right now, and to catch a wrong line before anyone acts on it, a live summary that you check while the room is still there is the better fit.