Does an AI note taker train AI on your recordings?

The short answer

It varies by vendor, and the default matters more than the option. Most tools that do train on customer content let you turn it off somewhere — but a setting you never found isn't protection, and on some plans the opt-out simply isn't offered.

One concrete example from this category: Notta uses customer recordings and transcripts to improve its models by default, and the opt-out is an Enterprise-plan feature. If you're on a Pro or Business plan, that's the deal you agreed to.

Noter AI doesn't train on your content. Your recordings and transcripts aren't used to train models, they're encrypted in transit and at rest, and you can delete them at any time.

Almost every mainstream option processes your audio in the cloud, including Noter AI. Processing and training are different things, and conflating them is the most common mistake people make when they try to evaluate this.

"Training on your data" vs "processing your data" — not the same thing

Getting this distinction right is most of the work.

Processing is what has to happen for the product to function. Your audio is sent to a server, a speech model turns it into text, a language model writes the summary, and the result comes back to you. Every mainstream meeting tool does this — none of the well-known ones run a full transcription and summarisation stack offline on your phone. It's a transfer of data with a purpose and an end.

Training is when your content is retained and used as material to improve the vendor's models for everyone. Your meeting stops being a thing you asked them to process and becomes part of what the product is built from. It's harder to reverse: deleting your account doesn't necessarily unwind a model that has already learned from your data.

Retention sits between the two, and it's the one people forget to ask about. How long does the audio live after the transcript exists? Granola, for example, deletes audio once transcription is done. Others keep it indefinitely by default, which is fine until it isn't.

A tool can be entirely honest about all three and still not be the right fit for a conversation covered by a client NDA. The point isn't to find a vendor with nothing in its terms — it's to know which of the three you agreed to.

How to check any tool in five minutes

1. 1Open the vendor's privacy policy and search the page for "train", "machine learning", and "improve our services". That last phrase is the one that most often carries the training permission.
2. 2Work out whether training is on by default or off by default. If the policy says you may opt out, assume you're currently opted in.
3. 3Check whether the opt-out is plan-gated. A control that exists only on an Enterprise tier isn't available to you on a personal subscription.
4. 4Find the retention period for audio specifically, not just for transcripts. They're often governed by different sentences.
5. 5Check who else processes the data — the speech and language model providers behind the product are subprocessors, and their terms matter too.
6. 6Confirm there's a real delete path: individual notes, and the whole account, on demand rather than by emailing support.
7. 7If the answers matter for your work, take a screenshot of the policy page with the date. Terms change.

Where the main meeting tools stand

This table covers what is publicly documented and verifiable. Where a vendor's current position isn't something this page can confirm, it says so rather than guessing — check their policy yourself before you rely on it.

Why this matters more for meetings than for most software

• You're consenting on other people's behalf. You agreed to the terms. The client on the other end of the call didn't, and they're in the recording too.
• Meetings are where the unguarded version of everything gets said. Pricing floors, staffing problems, the real reason a deal stalled. It's a higher-value corpus than almost anything else your company produces.
• NDAs generally don't carve out your note-taking app. If a conversation is covered, feeding it to a vendor that trains on customer content is a contractual question, not just a preference.
• Regulated work has harder rules. Health, legal, financial and HR conversations frequently can't be processed by whatever tool you personally liked best.
• Deletion doesn't run backwards. You can delete a recording. You can't retroactively remove its influence on a model that already trained on it — which is why the default setting matters more than the delete button.

How Noter AI handles it

No training on your content. Recordings and transcripts aren't used to train models — not by default, not on a lower tier, not as something you have to go and switch off.

Encrypted in transit and at rest. Audio is captured on your device and sent for transcription and AI summaries; the processing is cloud-based, and this page isn't going to pretend otherwise. There's no fully offline transcription mode in Noter AI, or in any of the mainstream alternatives.

Sign in with Apple, so you can use the app without handing over your real email address.

Delete anytime — individual notes or your entire account and everything in it.

That combination is a reasonable answer to give a client who asks what happens to the recording, which is ultimately the practical test: can you answer the question honestly in the meeting, without checking.

If you need more than a policy promise

For genuinely sensitive conversations, the strongest control isn't a vendor setting — it's not creating the recording.

Take performance reviews, legal discussions, and anything involving someone's personal circumstances by hand. Delete recordings once the summary has served its purpose. And for regulated work, get the tool through whatever review process your organisation uses for handling client data, rather than expensing a subscription and hoping the terms are fine.

Everything else — the standing team sync, the project kickoff, the supplier call — is exactly what these tools are for, and worrying about it equally is how people end up ignoring the question entirely.