Do AI Notetakers Train on Your Meetings? Tool by Tool

Some do, and the ones that do are usually the ones you've heard of. As of August 2026, Otter trains its proprietary AI on de-identified recordings and transcripts, Fathom trains its in-house models on de-identified customer data with an opt-out, and Granola uses anonymised data on Free and Business plans by default. Fireflies and Zoom say they never do. Tools that run entirely on your Mac have nothing to train on.

Full disclosure: Speechmark is our app, and it's on-device. Every claim about another tool below is quoted from that vendor's own privacy or help page, linked, and qualified "as of August 2026" — these policies change, sometimes quietly.

Which AI notetakers train on your meeting data?

Two things stand out. First, the split isn't cloud-versus-local: Fireflies and Zoom are cloud services with strict no-training positions, stricter on paper than some tools marketed on privacy. Second, where training does happen it is almost always on by default, so the people most exposed are the ones who never went looking for the setting.

Fathom deserves a specific note in its favour: it states that none of its AI subprocessors — Anthropic, OpenAI, or Google — "are contractually permitted to use our users' data to train their AI models." That's a meaningful commitment that many vendors don't make. The training it does do is its own, on de-identified data, and you can turn it off.

What does "de-identified" actually mean for a meeting transcript?

De-identification strips the metadata that names you — your account, your email, the file. It does not, and cannot, strip what was said. A transcript in which your CFO says the acquisition price out loud is still that transcript after the account label is removed.

Otter is unusually direct about the mechanism: it says it "uses a proprietary method to de-identify user data before training our models so that an individual user cannot be identified," and that "audio recordings and transcripts are not manually reviewed by a human" (as of August 2026). Take that at face value — it's a real control, honestly described. It's also a control over identification, not over content.

For most meetings that distinction is academic. For a client engagement under NDA, a therapy session, a due-diligence call, or an HR investigation, the content is the confidential thing, and "we anonymised it first" isn't the answer to the question your client is actually asking.

Isn't "we don't train on your data" enough?

It's a good signal, and worth having. But it's a narrower promise than it sounds, in three ways.

It's about use, not location. A vendor that never trains on your meetings still stores them. They remain subject to subpoena, to breach, to an employee with production access, and to whatever the retention policy says. Granola, for instance, deletes meeting audio after transcription but keeps transcripts and notes indefinitely on US servers unless you set a retention policy.

It's revisable. A privacy policy is a document, and documents get updated — Otter's was last revised in June 2026, Fireflies' in March 2026. Nothing prevents the next revision from reading differently, and the assets a company holds tend to get used when incentives shift or ownership changes.

It's usually about first-party training only. "We don't train on your data" and "our model providers don't train on your data" are separate claims, and only some vendors make both.

The architectural version of the promise is different in kind: if the audio never leaves your machine, there is no server-side copy that a future policy could reclassify. That's the whole argument for on-device processing, and it's the reason we built Speechmark that way.

Why the training question is really a consent question

The reason this matters more in 2026 than it did in 2023 is that you are frequently not the person who chose the notetaker. In a July 2026 survey of 500 employed US adults commissioned by Kolmogorov Law and distributed by Stacker, one in three (33.4%) said an AI notetaker had been present in their work meetings — and among those, only 34.7% said they were always asked first. A quarter (25.1%) were never asked and simply saw the bot appear. Nearly one in five of all workers surveyed (18.8%) later discovered a call had been recorded without their knowledge.

"For companies deploying AI notetakers — and for the far larger number whose employees bring them uninvited — the survey suggests the safest assumption is the one California law already makes: Nobody on the call has consented until they are asked."

— Kolmogorov Law survey, July 2026, via Stacker

That gap is exactly what the courts are now examining. In August 2025 a federal class action was filed against Otter alleging its notetaker recorded conversations without all-party consent and that the recordings were used to train its models; NPR covered the filing. A motion to dismiss was argued in May 2026 and, as of August 2026, no ruling has issued — no court has found Otter's practices lawful or unlawful, and there is no certified class or settlement.

The practical point stands regardless of outcome: your opt-out only covers your own account. It does nothing about the notetaker someone else brought to the call.

How to turn training off in each tool

If you're staying on a cloud notetaker, these are the settings worth ten minutes today (as of August 2026):

• Granola — Settings → Preferences → Data & sharing → disable "Use my data to improve models for everyone." Applies to Free and Business plans; Enterprise has separate admin controls.
• Fathom — opt out in account settings. On Team Edition, an admin can disable it for every user in Organization Settings, which is the version worth doing.
• Otter — Account Settings → Meetings → Feedback and Training → set "Help improve Otter" to Don't allow sharing. Worth knowing what this control covers: Otter's help page describes it as permission for Otter "and its third-party service provider(s)…to access your conversations for training and product improvement, including through human review." That sits oddly beside the privacy and security page, which says "audio recordings and transcripts are not manually reviewed by a human" — the likely reading is that the toggle governs a feedback-review path separate from the de-identified model-training pipeline. Either way, "Don't allow sharing" is the setting you want, and it's per-account, not per-meeting.
• Fireflies, Zoom — nothing to change; both state they don't train on customer content.

Do it at the team level where you can. An individual opt-out on a shared workspace protects one account's worth of a conversation that had six people in it.

When a cloud notetaker is the better choice

On-device isn't the right answer for everyone, and pretending otherwise would be a sales pitch rather than an argument.

Pick a cloud tool if you need cross-platform, cross-device access — Speechmark is Mac-only, and Fireflies, Otter, and Fathom work anywhere with a browser. Pick one if your workflow depends on server-side integrations like automatic CRM enrichment or a shared searchable archive for a whole sales team. Pick one if someone else administers your tooling and a local app won't pass IT review, or if you want compliance certifications — Fathom lists HIPAA and SOC 2 Type II; we hold neither, and we argue architecture rather than certification.

And if the deciding factor is simply that you want a no-training policy with zero setup, Fireflies' and Zoom's stated positions are strong, and both are free to start with.

Where Speechmark sits, including the caveat

Speechmark records, transcribes, diarizes, and summarizes entirely on your Mac. We don't train on your meetings because we never receive them — there's no account, no upload, and no server-side copy. The original audio stays next to the transcript as a local file. One purchase, $79 per Mac, no subscription.

One honest exception, which we'd rather you hear from us: if you connect your own Google Gemini API key for summaries, that transcript goes to Google under your account — and on Gemini's free tier, Google may use it to improve its models and human reviewers may read it. Turning on billing changes those terms, and the fully local summarization option sends nothing anywhere. We walk through the trade-off in the Gemini key guide.

If you want the wider field graded on the same axes — where audio is processed, who keeps the recording, what it costs — see the best private meeting notetakers for Mac, or the head-to-head with Otter.

FAQ

Do AI notetakers train on your meetings?

Some do. As of August 2026, Otter's privacy policy says it trains its proprietary AI on de-identified audio recordings and transcriptions. Fathom trains its own models on de-identified customer data with an opt-out, and Granola uses anonymised data on Free and Business plans by default. Fireflies and Zoom state they do not train on customer content.

Does Otter.ai train on my recordings?

Yes, by its own account. Otter's privacy policy (last updated June 2026) lists "training our proprietary AI technology on de-identified audio recordings and on transcriptions" among its uses of your data. Otter says the de-identification is automatic and that recordings are not manually reviewed by a human.

How do I stop an AI notetaker from training on my meetings?

It depends on the tool. In Granola, open Settings → Preferences → Data & sharing and disable "Use my data to improve models for everyone". In Fathom, opt out in account settings; Team Edition admins can disable it org-wide. In Otter, go to Account Settings → Meetings → Feedback and Training and set Help improve Otter to Don't allow sharing.

Is "we don't train on your data" the same as private?

No. A no-training promise says one thing about how your transcripts are used; it says nothing about where they are stored, who can subpoena them, or what happens after an acquisition or policy change. A cloud notetaker that never trains still holds a copy of your meeting on its servers.

Which meeting notetakers can't train on your meetings at all?

Ones that never receive your audio. Speechmark, Meetily, and MacWhisper transcribe and summarize on your Mac, so there is no server-side copy to train on — the guarantee comes from the architecture rather than a policy that can be revised.

See how Speechmark works → · Pricing →