ManageTime

Article

An AI is taking the notes: do you still need to take your own?

Keep taking notes, but fewer of them. An AI notetaker protects the record of what was said; it does not protect your memory of it. In two experiments with 200 people, those who photographed lecture slides instead of writing notes recalled no more than people who took no notes at all, even with the slides to review. This page is about what you write while the meeting runs, not what you do with the summary afterwards. It covers what the notetaker buys you, why your recall drops when you stop writing, and the few lines still worth taking by hand.

Do you still need your own notes when an AI is taking them?

The bot joins before you do. It sits in the participant list with a name you half recognize, and by the time the call ends there is a summary in the channel, a transcript nobody will open, and a list of action items with names already attached. The obvious conclusion is that your own notes are now redundant. The obvious conclusion is half right.

Start with the half that holds, because that part is real. Cognitive offloading is the general name for using something outside your head to carry what your head would otherwise hold: a shopping list, a calendar alert, a transcript. Burnett and Richmond pooled the experimental literature on it across two meta-analyses in Memory and Cognition, published in 2026, and found the benefit you would expect on memory-based task performance. Their second analysis is the one worth carrying into a meeting. Offloading also reduced the differences between people, and reduced them more in adults than in children. An external record levels the room. The colleague who never writes anything down and the colleague with the immaculate notebook now walk out with the same document.

Read what that benefit is measured against, though, because it is easy to over-collect. Those studies measure how well someone performs a memory task when the external store is there to be used. That is the record, and the record is exactly what an AI notetaker secures. It is not a finding about what stays in your head once the store is closed. Your week is full of moments where it is closed: a question in the corridor, a client call, the point next Tuesday where somebody asks why you agreed to this.

None of it was measured in a meeting. The offloading literature runs on lab tasks, the note-taking studies below on recorded lectures, and nobody in any of them was sitting in a status round with a bot in the participant list. Take the direction of these findings and treat the size as unsettled.

The record is safe. Your memory of it is the part that is not.

Why do you remember less from a meeting you did not write in?

Wednesday, a decision gets made, and you nod along. You were listening properly, too. Three weeks later somebody asks why the team went with the second option, and you have the summary open in front of you, which tells you what was decided and nothing about why. That gap is not laziness and it is not the tool. It is what the writing had been doing for you.

Wong and Lim ran the closest thing to a test of this in the Journal of Experimental Psychology Applied. Across two experiments with 200 participants, they compared three ways through a video-recorded lecture: writing longhand notes, photographing the slides with a phone, and taking no notes at all. Longhand note-takers beat both other groups on a later recall test. The comparison was rigged against them, too, since the photo-takers and the no-notes group got to review an exact transcript of the slides beforehand while the longhand group had only their own partial handwriting. The longhand group still won.

Photo-takers scored about the same as the people who had captured nothing.

The second experiment asked why, and the answer was attention rather than the notes themselves. Longhand note-takers mind-wandered less during the lecture, and that difference accounted for the gap in what they retained. Writing something down keeps you in the room. Pointing a camera does not, and neither, presumably, does knowing a bot has it covered. The focus pillar covers the attention side of this properly; the narrow point here is that a note is a way of staying present, and only incidentally a document.

Then the finding that makes this worth a page. Across both experiments the participants could not tell. They rated all three methods as about equally effective, including the two that left them recalling least. You do not feel this happening while it happens. That is precisely why an AI notetaker is easy to over-trust, because the confidence you carry out of the meeting is not evidence about what you took in, and it never was.

A 2025 paper puts a dent in the comfortable explanation for all this. Lurie, Fabrizio and Westerman ran three experiments in which people viewed artworks and either photographed them, screenshotted them, or simply looked. Memory was worse for anything saved, by any method. Screenshotting hurt most, which is the awkward result, because a screenshot is the least effortful save available. If dividing your attention were the whole story, the easiest capture should have cost the least. It cost the most.

The authors conclude that divided attention alone cannot explain the effect, and they do not settle what does.

What should you write down when a transcript already exists?

Four lines, and none of them is what was said.

The transcript already holds what was said, in more detail and with better spelling than you would have managed under time pressure. Your notes are for the things a transcript structurally cannot carry.

The first is what you now owe. A decision that binds your work, an action item with your name on it, a date you just agreed to out loud: write those in your own words, in whatever system your work already lives in. Getting Things Done calls that step clarifying, and the argument transfers without modification, because a commitment sitting in a summary somebody else generated is not a commitment you have captured. What happens to those items over the following fortnight belongs to the meetings pillar. The point here is only that they leave the meeting with you, or they do not leave.

The second is why. Summaries record outcomes and drop reasoning, since an outcome is quotable and reasoning is spread across forty minutes of half-sentences. Two lines on why the room went the way it did will be worth more in a month than the entire transcript.

The third is what you disagreed with and did not say. That one is the most valuable note you can take and the only one no tool will ever produce, because it never entered the audio in the first place. The fourth is the question you want to ask and cannot ask yet. Write it down and you stop rehearsing it, which is the thing that was quietly costing you the next five minutes of the discussion.

Reading the summary afterwards is worth doing, with one adjustment: read it as a prompt, not as a report. Skim it the same day, while you can still tell what it missed, and the gaps between what it says and what you remember are the useful part. Leave it a fortnight and you have nothing left to check it against, at which point the summary quietly becomes the meeting.

The ordinary limits are worth knowing too. These tools mishear names and in-house jargon, they attribute confidently, and they tend to flatten a room that was uneasy into a consensus that reads clean. No study on this page measured any of that, so treat it as what it is: what people who use these tools report, not a finding.

Writing costs something, and pretending otherwise would repeat the mistake this page is about. Every line you take by hand is a moment you were not following the discussion, which is the same arithmetic that decides what other work survives an hour you cannot leave. Four lines stays on the right side of it. A running transcript in your own handwriting does not, and it is also the one thing the bot has made redundant.

How does your working style change the answer?

Which part of this goes wrong depends on how you already work. The four names below are this site's, from the working-style self-check on the home page.

Readers closest to the Architect style keep taking full notes, and the notetaker changes nothing about that. What it should change is which notes. Duplicating a machine that transcribes better than you do is the one clear waste on this page, and the four lines above are the part of the notebook worth defending. Sprinters have the opposite problem and the more expensive one. Nothing written, full trust in the summary, and a confident account of a meeting on Friday that turns out to be the summary read back. They are the population that the misjudgment finding describes best.

Visionary-leaning readers pay the highest price of the four, because the thing they care about is the reasoning, and reasoning is the first thing a summary drops. Improvisers will tell themselves they will read it later. Later is usually past the point where they can still tell what the summary left out, and an unchecked summary is just a document.

None of this applies evenly, and one case runs the other way entirely. Where a meeting decides nothing that touches your work and you need none of the reasoning, the summary is enough and writing anything is waste. In formally minuted or regulated settings, the official record is the record, and your private notes are a personal aid rather than a second version of events. Know which of the two you are producing before you write.

So take the next meeting with a bot in the participant list. Write four lines and nothing else: what you now owe, why the room went that way, what you disagreed with silently, and the question you could not ask. Then open the summary the same afternoon and mark what it left out. The size of that gap is the size of what you have been losing.

Frequently asked questions

Are AI meeting summaries accurate enough to rely on?

Accurate enough for the record, and no substitute for your own memory of the hour. No study on this page measured how accurate these tools are. What the research measures is the other side of the trade. When people hand capture to a device, they recall less of what they sat through, and they cannot tell that it happened.

Why do you remember less from a meeting you did not take notes in?

Because writing keeps your attention in the room, and delegating the capture lets it drift. In two experiments with 200 people, longhand note-takers outperformed both photo-takers and people who took no notes on a later recall test. The second experiment traced the gap to mind-wandering, which the longhand group did less of. Those were lectures, not meetings, so treat the direction as the finding.

What should you write down when a transcript already exists?

Four things: what you now owe, why the room decided as it did, what you disagreed with silently, and the question you could not ask yet. The transcript already has what was said. Your notes are for what it cannot hold. Reasoning gets dropped from summaries because outcomes are quotable and reasoning is not, and an unspoken objection never entered the audio at all.

Is it worth reading the AI summary afterwards if you were in the meeting?

Yes, the same day, and read it as a prompt rather than as a report. While the hour is still fresh you can see what the summary missed, and the gaps are the useful part. Left a fortnight, you have nothing to check it against and it becomes your only version of the meeting. These tools also flatten disagreement into consensus, which is the gap worth looking for first.

Sources