When the Transcript Gets It Wrong

Everything looks equally correct

The dangerous thing about a machine written summary is not that it makes mistakes. Everything makes mistakes. It is that the mistakes arrive in exactly the same font as the facts.

A person taking notes by hand leaves a trail of their own uncertainty: a question mark, a name half written, a gap. You can see where they lost the thread. A generated summary has no such tells. The line the tool is sure about and the line it invented are formatted identically and read with the same calm authority. There is no visible seam, which is why people stop checking.

So this lesson is about the three failures worth knowing by name, and the ninety seconds that catch them.

Names, and why they matter more than they should

Speech to text handles common words well and unusual ones badly, and names are the most unusual words in any meeting.

Your colleagues' names get mangled, particularly names the system has heard less often, which in practice means names from outside the dominant language of the training data. That is not a neutral technical quirk. If Adaeze appears as "Adze" in every set of notes for a year while Sarah is always Sarah, the tool has produced a small, repeated, entirely visible insult, and you are the one who signed the email.

The same goes for your own vocabulary: product names, internal systems, acronyms invented in some forgotten workshop, suppliers, the client's brand. Anything that sounds vaguely like a real word gets replaced with that real word, and the summary carries the substitution forward without hesitation.

The wrong person said it

Speaker labelling is a guess, and it is a harder guess than it looks. Two colleagues with similar voices, a shared room with one microphone, a caller whose connection dips, somebody who joined from a different device halfway through: any of these can swap a name at the front of a sentence.

Nine times in ten this is harmless. The tenth time, the sentence was "honestly, the client is impossible to work with", and it now sits in a document with a name in front of it, and that name belongs to someone who never said it.

This is the error that damages relationships rather than schedules. It is also easy to miss, because the sentence itself is accurately transcribed and reads perfectly naturally. Nothing about it looks wrong. So the rule is narrow and worth keeping: if a summary quotes or attributes anything critical, anything about a person, or anything you would not want repeated, check that line against the transcript before it goes anywhere.

Agreement that never happened

This is the serious one.

People do not only speak in order to communicate meaning. They make noises to show they are still listening. "Yeah." "Mm." "Sure." "Right, okay." Linguists call these backchannels, and in a normal conversation they mean "I am following you", not "I agree with you". Sometimes they mean "I am thinking". Occasionally they mean "I am pretending to think while I work out how to say no".

A transcript records the word and loses everything that made it mean nothing. The summariser then reads "Sure" immediately after a proposal and draws the obvious conclusion. So you end up with a line like this in your notes:

"Agreed that Dan's team will absorb the extra work this quarter. Confirmed by Dan."

What Dan actually did was say "yeah" while frowning, and then ask two more questions. The frown is not in the transcript. Nor is the pause, nor the tone, nor the fact that everyone in the room understood him to be sceptical. In person, nobody thought Dan had agreed. On paper, Dan has agreed, and the paper is what survives.

This is worse than a misheard name because there is no error you can point at. Every individual word is correct. The meaning was created by the summary, and once it goes out to eight people unchallenged, it becomes what happened.

The defence is to make the tool show its working.

For every item you list under DECISIONS, add the exact sentence from the
transcript that shows it being agreed, and who said it.
If the only supporting evidence is a short response such as "yeah", "sure",
"ok" or "mm", do not list it as a decision. Move it to a section called
NEEDS CONFIRMING and say who needs to confirm it.
Do the same for actions: if nobody explicitly took the task on, put it in
NEEDS CONFIRMING rather than assigning it.

Then read the NEEDS CONFIRMING list and send two messages. It takes a minute and it is the difference between a record and a fiction.

Numbers, dates and quiet invention

Figures said out loud in meetings are approximate, often misheard, and sometimes wrong to begin with. A summary strips away the hedging around them, so "roughly forty, maybe fifty, I would need to check" becomes "approximately 50" and then, two documents later, becomes 50.

Treat every number in a generated summary as unverified. If a figure matters, get it from wherever the real figure lives, and if it does not matter, take it out.

The ninety second read

You do not need to check everything. You need to check five things, and it is quick because you were there.

1. Names. Is everyone spelled right? Fix them, every time, including in the transcript if your tool allows it.
2. Attribution. Is anything quoted or credited to a person? Does it sound like that person?
3. Decisions. For each one, can you remember it actually being decided? If you hesitate, it goes in the list of things to confirm.
4. Owners and dates. Did the person named actually take the task? Did anyone say that date out loud?
5. Numbers. Are they right, and do they need to be in there at all?

Do this on every summary before it leaves your hands. Do it twice as carefully when the notes are going to a client, going into a file that will be read in a year, or forming the record of a decision somebody might later dispute.

You are the author

Here is the mental shift that makes all of the above automatic.

The tool did not write your meeting notes. It produced a draft, and you are the author. When you press send, the document becomes yours: your name is at the top, your judgement is what people are relying on, and "the AI wrote it" is not a sentence that has ever improved anybody's situation once something has gone wrong.

That is not a burden, it is a relief. You do not have to trust the machine, and you do not have to check every word. You have to read a draft written by a fast, tireless, slightly literal minded assistant who has never met your colleagues. That is a job you already know how to do.