An AI Notetaker Joined Your Meeting. Can You Say No?
Published:
Yes, you can almost always say no, and the moment to do it is the first minute, before anyone has said anything worth transcribing. Asking is not being difficult. In most rooms it takes one sentence and nobody remembers it by the afternoon. The awkwardness people fear comes from asking twenty minutes in, when the request sounds like an accusation about what has already been said.
It helps to know what you are actually objecting to. That bot is not a tape recorder. It produces a transcript, a summary, a searchable document in somebody else’s Drive, and in some cases training data. Those are four different things, and the vendors treat them differently. Here is the shape of it, then the scripts.
What that bot in the participant list is doing
Two very different things wear the same “is recording” label.
The built-in kind. Zoom, Teams and Google Meet each have a note-taker that runs inside the platform you are already on, so no extra guest appears in the list. In Google Meet, when “Take notes for me” is running, Google’s help documentation says Google Meet “informs all meeting participants that notes are being taken,” and “the pencil icon appears on all participants’ screens.” The notification is the product, not a courtesy.
The bot kind. Otter, Fireflies, Read, and the rest join as a participant, usually with a name like “Otter.ai Notetaker” or somebody’s name followed by “Notes”. This is the one that surprises people, because it is a third-party company sitting in a meeting that has nothing to do with them. The host connected their calendar once, months ago, and now the bot shows up to everything automatically. Very often the host has genuinely forgotten it is there. That matters for your tone: the person you are about to ask is usually not scheming, they are on autopilot.
Either kind captures more than notes: a transcript with speaker labels, so every sentence is attributed to a name, plus a summary, action items, and in some tools screenshots of whatever was shared. Where the host has enabled screenshots of presented content, Google’s help page says participants are shown a notice that “Gemini may improve your notes with screenshots.” The offhand remark you would never write in an email becomes a line of attributed text in a document you will never see.
Who reads the transcript afterwards
This is the part people get wrong, and it is the part worth asking about.
The transcript is not yours. It belongs to whoever turned the thing on. In Google Meet, the notes document is “saved in the meeting organizer’s Google Drive,” and the host or co-host picks the audience when they start taking notes: all invited guests including those outside the organization, invited guests inside the organization, or hosts and co-hosts only. You, the guest who was transcribed, are not in control of which of those three applies. You may not receive the document at all.
Then it keeps travelling. A transcript sitting in someone’s Drive is searchable by their admins, subject to their retention rules, discoverable in litigation, and one careless share link away from a much wider audience. If the host is a vendor pitching you, the summary may end up in their CRM, quoted by their sales manager. None of that is sinister. It is just not what “we’ll just get notes” sounds like.
None of this is unique to notetakers. Admin visibility is the normal state of a workplace tool, and the same reach that lets a Team admin see chats covers the transcript of the call you sat in on as a guest.
The honest reframe: a recorded meeting is closer to writing an email than to having a conversation. If you would not put it in writing to that company, do not say it to their bot. That instinct is the same one behind what not to share with ChatGPT, just pointed at your voice instead of your clipboard.
Does it train a model? Where the vendors differ
Here the answers genuinely diverge, so this is worth checking rather than assuming. The three postures you will run into:
⬣ Otter: trains on your content, and says so. Otter’s privacy policy lists among its purposes: “Improve and monitor the Services, including training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information).” Read that parenthesis twice. The de-identifying is described for the audio; the transcriptions are the part the policy itself flags as possibly carrying personal information, and they are trained on too. The policy also puts the permission burden on the user who turned it on, not on Otter: “please make sure you have the necessary permissions from your co-workers, friends or other third parties before sharing Personal Information.” You, the guest, are the third party. Nobody asked you.
● Microsoft Copilot: does not train foundation models on it. Microsoft’s documentation is unambiguous and repeated: “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs, including those used by Microsoft Copilot.” The transcript still exists inside the tenant, still visible to admins through Purview and Content search, still subject to retention policy. Not trained on is not the same as not readable.
▲ Platform-native tools generally: check the tier. The commitments that matter usually live in the business and enterprise terms rather than the free ones, and the wording changes. Check your own plan’s terms on the day, because this is exactly the kind of fact that quietly ages wrong.
The distinction to carry away: who can read it and what trains on it are two separate questions. A vendor can answer one cleanly while the other stays wide open.
Where consent rules generally come from
We are describing the landscape here, not telling you what you are entitled to. For your situation, ask your legal team or your DPO.
Recording rules mostly predate AI by decades and come from wiretap law. California’s Penal Code section 632, for example, reaches anyone who “intentionally and without the consent of all parties to a confidential communication, uses an electronic amplifying or recording device to eavesdrop upon or record” that communication. Some jurisdictions require every party to agree; others require only one. In the EU, meeting recordings that identify people are personal data, so the GDPR’s usual questions apply: lawful basis, transparency, proportionality. Those are the same questions raised by uploading a document to ChatGPT, pointed at audio instead of a file.
Whether those statutes reach a bot sitting quietly in a video call is being tested right now. A class action filed against Otter.ai on 15 August 2025 alleges exactly that: the named plaintiff says he never signed up for Otter, and that a sales call of his was recorded because another participant had OtterPilot running. Those are allegations in a complaint, untested and not ruled on, so nobody should be telling you the answer yet.
The more telling signal is that the platforms are moving without waiting for the courts. In spring 2026, Google announced a Workspace setting letting administrators “require explicit consent from meeting participants before automatic note-taking,” where participants on supported devices “must actively agree to be recorded, be transcribed, or have notes enabled.” The setting is off by default, so it only applies where an admin has turned it on. Read the direction of travel as the industry conceding the point: silently capturing people who never agreed is a problem, even where it might turn out to be lawful.
What to say in the first minute: three scripts
Short, early, no hedging. Pick the one that fits.
When you just want it off. “Before we start, could we run this one without the notetaker? Happy to write up anything we agree at the end.” You have given them the thing they wanted (a record) while removing the mechanism. Hosts almost always say yes, because what they wanted was a record of the decisions, not a recording of you.
When it can stay, but not for everything. “Fine to keep notes running for the overview. When we get to the pricing and the client names, can we pause it?” This is the realistic one. Meetings are rarely uniformly sensitive; it is usually ten minutes in the middle.
When you need to know before deciding. “Which tool is that, and where does the transcript end up?” Not confrontational, just procedural, and it does something useful: half the time the host does not know, and discovers in real time that they do not know. That question changes more behaviour than any objection.
If the answer is no, that is information too. A host who will not turn off a bot for a guest who asked has told you what kind of counterparty they are. Adjust what you say, not whether you stay.
One thing stays yours regardless: what you screen-share. The bot may be capturing screenshots, and a shared window full of real customer names is a leak no consent conversation fixes. Before sharing a document in a meeting you do not control, run it through the paste checker: it flags names, addresses, contact details and card numbers locally, while the text is still on your side of the screen. And if the bot belongs to your own employer rather than a stranger’s, the question shifts to whether your employer can see your ChatGPT.
If you are the one running the bot
Most of us are on both sides of this. The courteous setup takes five minutes, once.
Turn off “join every meeting on my calendar.” Automatic attendance creates the awkward moment, because the bot arrives in rooms you never decided to record. Invite it deliberately instead.
Say it out loud in the first minute rather than relying on the participant list. “I’ve got the notetaker running, shout if you’d rather I didn’t.” Nobody then has to interrupt a meeting to object, which is exactly the barrier that makes silent consent worthless.
Share the output with the people who were in it. If you have a summary of a conversation and they do not, you are holding an asymmetry they never agreed to. Sending the notes round costs nothing and removes the whole objection.
And check the training answer for your own tool before it is your name on the incident. If your notetaker trains on transcripts, every client conversation you have recorded went into that pipeline, and your client’s data protection clause probably has something to say about sub-processors you never mentioned. The same reasoning for text tools runs through whether ChatGPT is safe for work.
None of this requires being the person who makes meetings weird. It requires one sentence in the first minute, from whichever side of the bot you happen to be on.