← All Articles
AI for Real Life

The silent listener in every meeting: the 10-minute audit of your AI notetaker

A notetaker bot joined Sara’s stand-up uninvited — and emailed the summary to everyone on the invite, including people who no longer work there. Ten minutes of settings is what puts you back in charge of where your meetings go.

November 20, 2026 · 11 min read
A meeting room seen from above the table: four people mid-conversation with laptops open, and a small robot with a microphone for a head sitting in an empty chair, quietly taking notes while nobody looks at it
A notetaker bot joined the meeting from someone’s calendar — and nobody in the room had invited it. The audit makes sure you know where it took the conversation afterwards.

It’s 9am on a Tuesday. Sara opens her laptop for the team stand-up, and before anyone has said hello, a new name appears in the participant list: a notetaker bot. Nobody in the room invited it. One colleague connected it to their calendar weeks ago, and now it joins every meeting on its own, records every word, and emails a summary to everyone on the invite afterwards — including the contractor whose contract ended last month and is still on the recurring invite.

Sara didn’t agree to be recorded. Neither did anyone else on the call. And the meeting hasn’t even started.

That bot is now ordinary. In the last two years, AI notetakers have gone from a novelty to a default: a small participant that records the audio, turns it into a transcript, writes the summary, and files it all away on the vendor’s servers. The convenience is real — nobody takes minutes any more. But a meeting is not a document. It is people saying things they would never write down: a salary figure, a health worry, a disagreement about a colleague, a plan that isn’t public yet. Once a bot has heard it, that conversation lives in places Sara never chose: a vendor’s cloud, a training set, an inbox she can’t recall it from.

This piece is the ten-minute check that tells you where your meetings actually go — and the switches that put you back in charge of them.

What the bot does with your meeting

Strip away the brand names and every AI notetaker does the same four things:

  1. It joins the call — either because the host turned it on, or because someone connected it to their calendar and it now joins everything automatically.
  2. It records the audio and turns it into a transcript, usually with each speaker labelled.
  3. It stores the recording and transcript on the vendor’s cloud, where it is searchable — by you, by your workspace admins, and sometimes by anyone the summary is shared with.
  4. It shares the result — a summary emailed to the invite list, a link to the full transcript, sometimes posted straight into a shared channel.

Steps 1 and 4 are where the trouble lives, because both can happen without anyone in the room making a decision in the moment. Auto-join means the bot is there before the meeting starts. Auto-share means the summary leaves before anyone has read it. And step 3 is the one nobody thinks about: the transcript is a permanent, searchable record of a conversation everyone assumed was ephemeral.

The number to remember: 65. In September 2024, an AI notetaker joined a hospital’s regular clinical discussion in Ontario, Canada, through a former physician’s still-active calendar invite. It recorded a conversation about seven patients and automatically emailed a summary — with a link to the full transcript — to all 65 people on the invite list. Twelve of them no longer worked at the hospital. The hospital’s privacy regulator published its findings in October 2025. Nobody in the room had realised the bot was there, and nobody noticed until the summary arrived.

That is the worst version of the story, and it needed several small failures to line up: a stale invite, an auto-join setting, an auto-share setting, a tool never approved for that kind of meeting. But every one of those failures is a setting, not a hack. Which is why the audit below is worth ten minutes: it finds the same settings in your own account before a sensitive meeting finds them for you.

The 10-minute audit: five questions

You need the answers for the notetaker you use — and, if colleagues invite bots to your meetings, for theirs too. The audit is five questions, and every one of them has an answer on the vendor’s own privacy or settings page. Don’t take the marketing page’s word for it; the privacy policy is the document that counts, and it carries a date. If the answer isn’t there, that is itself an answer.

#The questionWhere to find itWhat a good answer looks like
1Where does the audio go?Privacy policy, “where we process data” or security pageStored encrypted, in a region you can name, by the vendor itself — not passed to unnamed sub-processors
2Who can read the transcript?Sharing settings + privacy policyOnly you, until you share it; admins and support only with your explicit consent
3Does it train an AI model on my meetings?Privacy policy, “how we use your data”A flat no — or a switch you can turn off yourself
4How long is it kept?Retention / data-deletion settingsA number of days you chose, with auto-delete you can switch on
5How do I delete it?Settings, or the vendor’s deletion pageOne place, works for the recording and the transcript, confirmed in writing

Work through them in order. Question 3 is the one most people skip, and it’s the one that matters most — because it’s the difference between a tool that works for you and a tool that learns from you.

What the big names actually say

The vendors do not all do the same thing. Read their own words, and the choice gets much simpler.

Otter.ai — trains on your meetings unless you’re on Enterprise. Otter’s privacy policy, effective 16 June 2026, lists “training our proprietary AI technology on de-identified audio recordings and on transcriptions” as one of the purposes it uses your data for. De-identified means Otter strips away the labels that name you — your account, your email — before training. It does not strip away what was said. The words, the figures, the argument: those are the content, and the content is what gets learned from. Otter’s own help pages describe a switch — Account Settings → Meetings → Feedback and Training → “Help improve Otter” — and its help centre says Enterprise workspaces are opted out of training by default. On any other plan, assume training is on until you have checked.

Otter is also the subject of a live federal lawsuit about exactly this. In In re Otter.AI Privacy Litigation, a US federal judge ruled on 13 August 2026 that the core claims — under the federal wiretap law, California privacy law and Illinois biometric-privacy law — could proceed past Otter’s attempt to have the case dismissed. The plaintiffs allege the notetaker recorded meetings without all participants’ consent and used the recordings to train its models. To be clear about what that ruling is and isn’t: it decides that the claims are plausible enough to go forward. It is not a finding that Otter broke the law, and the case is still pending. But it is the reason the consent question in this piece isn’t theoretical.

Fireflies — does not train, on any plan. Fireflies’ privacy policy, last updated 28 September 2026, says: “We do not use personal information for AI model training and we contractually prohibit our vendors from using this information for their own model training.” Its security pages say the same in plainer words — Fireflies does not use your meeting data to train AI models, and its AI vendors are contractually barred from training on it — and that its own staff have no access to meeting content by default. This is the answer question 3 is looking for, and it applies on the free plan as well as the paid ones.

Zoom AI Companion — does not train either. Zoom says it does not use any customer audio, video, chat, screen sharing, attachments or other meeting content to train Zoom’s or its third-party AI models, across all of its AI Companion deployment options. There is a practical difference worth knowing: AI Companion is built into Zoom, switched on by the host for that meeting, and attendees are notified when it’s running. A third-party bot, by contrast, joins as a participant from someone’s calendar and can keep joining meetings its owner forgot about — which is how the Ontario hospital’s bot got into the room.

That comparison is the whole audit in miniature. Same job — notes from a meeting — and three different answers to the only question that decides what happens to your words afterwards.

The fixes: four switches and one habit

The audit tells you where you stand. These are the changes that move you somewhere better. All of them are settings, not promises you make to yourself — a setting can’t be talked out of behaving the way a resolution can.

1. Turn off auto-join. This is the single most valuable switch. A notetaker that only joins when you invite it to a specific meeting cannot wander into a sensitive call on its own, and it cannot join through a stale calendar invite months after someone left. Find it in the notetaker’s Settings — it is usually called “auto-join” — and turn it off. From now on, the bot joins when you choose the meeting, not when your calendar does.

2. Turn off auto-share. A summary that emails itself to the invite list is a copy of your meeting leaving the room before anyone has read it. Turn off automatic sharing of summaries and transcripts, and share them yourself, to the people who need them, after you’ve read them. While you’re there, check who the summary would have gone to: open one recurring meeting invite and look at the guest list. If someone who left is still on it, remove them now — that stale name is how the Ontario summary reached twelve former staff.

3. Turn off training, and set auto-delete. If your notetaker has a training switch, turn it off — for Otter, that’s the “Help improve Otter” switch above. Then set a retention limit: 30 or 90 days, whatever your work allows, with automatic deletion switched on. A transcript you no longer need is a liability, not an archive. And check the deletion actually covers both the recording and the transcript; some tools delete one and keep the other.

4. Ask the room before a sensitive call. Before any meeting where health, money, personnel or anything confidential will be discussed, say it out loud at the start: “Is anyone recording this, or has anyone invited a notetaker?” It takes five seconds, and it catches the bot that joined from someone else’s calendar — the one whose settings you can’t check. If the answer is yes and the topic is sensitive, ask for it to be turned off for this meeting. If your workplace has an approved notetaker, use that one; never bring a personal notetaker into a work meeting that handles other people’s confidential information without an approved system behind it.

The habit that holds it together: treat the transcript as a document, not a memory. A transcript can be searched, forwarded, quoted and kept. Before you share one, read it the way a stranger would: does it contain a number, a name or a sentence that was meant for this room only? If it does, cut it before you send it — or don’t send it at all.

Where it still breaks

An honest audit includes its limits, and this one has four.

  • De-identified is not content-free. Removing your name from a recording doesn’t remove what you said. A training set that learned your team’s pricing discussion has still learned your team’s pricing discussion.
  • Policies change, quietly. Otter’s policy is dated June 2026, Fireflies’ September 2026. Re-run the five questions when you renew, or when a vendor announces a new feature — that is usually when the defaults change.
  • You can’t audit someone else’s bot. Your settings don’t bind a colleague’s notetaker. That’s what the ask-the-room habit is for, and why it matters more than any switch of your own.
  • Consent law differs by place. In some places every participant must agree to be recorded; in others, one person’s consent is enough; at work, your employer’s policy may already decide the question for you. This piece can’t settle that for you — check your employer’s policy and your local rules rather than guessing, and don’t treat a bot’s presence as consent.

None of these is a reason to skip the ten minutes. They are the reason to spend them on settings rather than on trust.

The checklist, ready to copy

The 10-minute AI notetaker audit

  1. Turn off auto-join — the bot joins only meetings you choose.
  2. Turn off auto-share — you send the summary, after reading it, to the people who need it.
  3. Answer the five questions for your notetaker: where the audio goes, who can read the transcript, whether it trains a model, how long it’s kept, how you delete it.
  4. Turn training off where a switch exists, and set auto-delete to 30 or 90 days.
  5. Clean one recurring invite: remove anyone who no longer belongs on it.
  6. Before any sensitive call, ask the room: “Is anyone recording this?”

Copy it and do it once. It holds for every meeting after that.

The honest limit

A notetaker is a useful tool, and this piece isn’t telling you to stop using one. It’s telling you that a meeting is the one place where your words are also other people’s words — your colleague’s salary, your client’s plan, a patient’s care — and a bot that records the room records all of them at once. Ten minutes of settings is what makes sure the convenience stays yours, and the record stays where you put it.

The bot joined the meeting. The question is whether you invited it — and whether you know where it took the conversation afterwards.

Where this comes from

  • Otter.ai, Privacy & Security page and Privacy Policy (effective 16 June 2026) — training on de-identified audio recordings and transcriptions; Enterprise workspaces opted out of training by default; third-party AI providers do not train on customer data.
  • Fireflies.ai, Privacy Policy (last updated 28 September 2026) and Security page — no use of meeting data for AI model training; vendors contractually barred from training; staff have no access to meeting content by default.
  • Zoom, “How Zoom’s terms of service and practices apply to AI features” and AI Companion data-governance pages — no use of customer audio, video, chat, screen sharing, attachments or other meeting content to train Zoom’s or third-party AI models.
  • Information and Privacy Commissioner of Ontario, findings letter, 27 October 2025 — Otter.ai bot joined an Ontario hospital’s 23 September 2024 hepatology rounds via a former physician’s calendar invite; seven patients’ personal health information discussed; summary and transcript emailed to 65 invitees, 12 no longer employed; hospital reported the breach 17 December 2024.
  • In re Otter.AI Privacy Litigation, No. 5:25-cv-06911-EKL (N.D. Cal.), order 13 August 2026 (Judge Eumi K. Lee) — federal wiretap, California privacy, biometric and unfair-competition claims survived a motion to dismiss in part; computer-fraud claims dismissed with leave to amend; a ruling on the pleadings, not a finding of liability.
  • Daytapes, “Do transcription apps train AI on your audio? 10 policies compared” (policies read 14 November 2026) and Speechmark, “Do AI notetakers train on your meetings?” — independent cross-checks of the Otter training switch and the Fireflies no-training position.

Sources

  • Otter.ai, Privacy & Security — Privacy page
  • Fireflies.ai, Privacy Policy — Privacy Policy
  • Fireflies.ai, Policy on keeping information safe — Guide
  • Zoom, Terms of service and AI features — Zoom blog
  • Information and Privacy Commissioner of Ontario findings, via HR Reporter — HR Reporter
  • Otter.AI wiretap lawsuit, explained — Recording Law
  • Otter.AI case summary — DataPacts
  • Daytapes, Do transcription apps train AI on your audio? — Daytapes
  • Speechmark, Do AI notetakers train on your meetings? — Speechmark

Free newsletter

Get guides like this in your inbox

Practical AI for parents, teachers, and everyday people. Copy-paste ready. Always free.

Get the weekly newsletter Subscribe free