ChatGPT meeting notes: 8/11 on a meeting with 11 traps
One constructed eight-minute checkpoint, eleven traps, one run on 8 October 2026. Five actions were agreed and none made the action list. Asked afterwards, it gave four of them back with owners and dates.
Let it draft the summary, then rebuild the action list yourself or ask it for every action with an owner and date. Don't paste its action list into a tracker. It held the total cost, the corrected date and the change of owner, but its one action item was a risk nobody agreed to own, and the VAT fix lost its owner and date when about 25 seconds dropped out of the transcript.
- £18,500 recorded as the total for both weekends, with "total" in bold
- Priya's sign-off on the 28th, with the reason, after she corrected the 21st
- Comms moved from Priya to Sandra, with 23 October, and Priya still supplying the finance content
- Supplier extract given to Kieran through the crosstalk, with no speaker labels to help
- Parallel run cut to two cycles with full reconciliation in each
- Online training filed under Open decision, not as decided
- Priya's "don't minute this" comment left out
- Asked afterwards, it rebuilt the action table with the right owner for four of the five agreed actions, and kept the outage as a risk and training as parked
- Action items held one entry, assessing the test-environment outage, which nobody agreed to and nobody owns. In the meeting it was a risk with no owner.
- None of the five agreed actions were in the action list. Four were only in the summary prose.
- Kieran's VAT fix by Friday 16 October was lost. About 25 seconds dropped out of the transcript, taking the owner and the deadline with it, and the follow-up couldn't recover them.
- Added "proceeding still depends on a clean trial load" to a cutover date the chair called decided
- The transcript has no speaker names, only You and Others, and every line appeared twice because the speakers fed the microphone
- The meeting audio is deleted once the notes are ready, so you can't replay it to check
ChatGPT's new Meetings plugin sat in on an eight-minute project checkpoint and scored 8 out of 11. The summary was mostly right: the total cost, a date corrected mid-sentence, a job handed from one person to another. The action list had one item on it, and nobody in the meeting agreed to that one.
Watch the run. The recording, script and answer key are below if you want to try it on your own note taker.
How I tested it
I didn't want to put a real client meeting online, so I wrote one: the weekly cutover checkpoint for a finance system migration at Ashworth Logistics, a made-up company, voiced with four AI voices from ElevenLabs. It runs eight minutes. I wrote it the way meetings go. Someone offers to do a job and gets overruled. A date is corrected in the same breath. A cost is queried: per weekend, or total? Two people talk over each other. Someone says "don't minute this". The chair's recap leaves an action out. That's eleven traps, and I wrote down the right answer to each before I recorded anything.
Then I started Meetings in the ChatGPT app on my Mac, played the recording through the speakers and let it capture the meeting live. One run, no retries, 8 October 2026. Meetings is a beta in the ChatGPT desktop app on macOS for Pro and Business plans, with Enterprise to follow, according to OpenAI's release notes of 29 September 2026.
Run this yourself
Play the same meeting into whatever note taker you use and see how many of the eleven it gets. Free, no sign-up:
The steps are on the AI meeting notes test kit page.
- The meeting recording (MP3, 8 minutes, 11.5 MB)
- The meeting script (PDF, with speaker names and times). For checking afterwards. Don't give it to the tool.
- The answer key (PDF, with pass, partial and fail rules and a score sheet). Open it once your notes are written.
Score
What it got right
The budget. The vendor quotes £18,500, Kieran asks whether that's per weekend, and it's the total for both weekends. ChatGPT wrote "total" and put it in bold. Had it written "per weekend", someone's forecast would be out by £18,500.
The corrected date. Priya says the 21st, stops herself and moves it to the 28th because month-end close runs to the 27th. ChatGPT has the 28th and the reason. The change of owner: Priya offers to write the comms and the chair gives the job to Sandra. ChatGPT has Sandra, with the 23rd, and noticed Priya is still supplying the finance content.
Kieran and Sandra talk across each other about the supplier extract, and Kieran ends up saying "I'll take it". ChatGPT gave it to Kieran. Moving training online was parked, and ChatGPT filed it under Open decision. The comment Priya asked to keep out of the minutes isn't there.
What it got wrong
The action list. Its only item is to assess whether the test-environment outage on 19 and 20 October disrupts the trial load. That's a sensible job, but nobody agreed to do it. In the meeting the outage was logged as a risk with no owner, and the chair said so. ChatGPT turned the one thing nobody owns into the only action on the page, and it doesn't say who it's for.
Five actions were agreed: Kieran fixes the VAT codes by Friday 16 October, Sandra sends the comms draft by the 23rd, Priya signs off the reconciliation on the 28th, Karen chases Mark for the budget approval, and Kieran pulls the supplier extract. None of them are in the action list. Four are in the summary as prose. Copy the action list into a tracker and you get nothing that was agreed.
The fifth isn't anywhere. The summary knows 47 of about 1,200 VAT codes are wrong. It doesn't know Kieran is fixing them, or that he said Friday the 16th. The transcript jumps from "the ones Pegasus handled with that" to "not Friday afternoon". About 25 seconds is missing, and the deadline went with it.
A smaller one: the chair says the 14 November cutover is decided. ChatGPT added "proceeding still depends on a clean trial load". That was Priya's comment, and it turns a firm decision into a conditional one.
The transcript
It doesn't know who's speaking. It splits the call into You, your microphone, and Others, everyone else. In my run every line appears twice, because the speakers fed the microphone. On a real call, Others is everyone else in one lump. ChatGPT got owners right where people said names out loud. In a meeting where nobody does, expect owners to be guessed.
OpenAI's release notes say the audio is deleted once the notes are ready and can't be replayed. If you want to check "21st or 28th" later, you need your own recording or your own notes.
Asking it again
There's a box under the notes for questions about the meeting. I asked once, on camera:
That came back as a proper table: Sandra on the comms by 23 October, Priya's sign-off on the 28th, Karen chasing Mark this week, Kieran on the supplier extract. It kept the outage as a risk with no owner and training as parked. The VAT fix came back too, but with the owner "not clearly assigned in the captured transcript" and the deadline "not captured". What never reached the transcript can't come back out.
That's the tip. Ask for actions, owners and dates every time, check the table against your own notes, and paste that into your tracker instead of the action list. You shouldn't have to ask.
List every action agreed in this meeting, with the owner and due date for each.
The four checks
- Rebuild the action list yourself. Every action needs an owner and a date. If one hasn't got both, that's a question for the meeting, and ChatGPT can't answer it.
- Check every number and date against a recording or notes you keep. ChatGPT deletes the audio. Get permission before you record anyone.
- Check every decision. Was it decided, conditional or parked? My run turned a firm decision into a conditional one.
- Check what shouldn't be there: anything said off the record, anything about a person, anything that would embarrass someone if it got forwarded.
What I would let it do
I'd let it write the summary. It's a better set of minutes than some people write, and it caught the traps that depend on listening carefully: the total, the corrected date, the change of owner. I wouldn't let its action list near a tracker. Ask for the actions, check them against your own notes, then send. It's a beta, and speaker names and a proper action list are the two things I'd want fixed before it comes out of one.
Limits of this test
One meeting, one run, on one date, with a feature in beta. The voices are AI-generated, which makes for cleaner audio than a real call: one clear voice each, nobody on a bad microphone. On a noisy real meeting I'd expect it to do worse, though I haven't tested that. Some of the transcript's mishearings, like "the lead slot" for "the Leeds lot", showed up in a second transcription tool too, so the synthetic voices may share the blame for those. It doesn't rank ChatGPT against other note takers.
Started Meetings in the ChatGPT Mac app and played the MP3 through the Mac's speakers so it captured the meeting live. One run, no retries. Scored the notes against an 11-row answer key written before the meeting was recorded. Then one follow-up in Ask about this meeting: List every action agreed in this meeting, with the owner and due date for each.- Constructed meeting. Ashworth Logistics, Finity and everyone in the call are fictional. The voices were generated with ElevenLabs on my own paid plan.
- Hands-on test on my own ChatGPT Pro plan, which I pay for. No free access, sponsorship or affiliate arrangement.
- The answer key was written before the meeting was recorded. One run, no retries.