"Opt out any time" is not the same sentence as "we don't."

Six meeting tools make six different promises about what happens to a recording after the meeting itself is over, and the promises are easy to mistake for each other if nobody reads past the marketing page. Some state outright that customer data is never used to train anything. Others offer a setting that turns training off, which is a real control and also an admission that the setting starts in the other position. And at least one routes audio through a third-party partner whose own training policy depends on which plan a customer happens to be on. None of this makes a tool untrustworthy — a written opt-out is still a real commitment, and a company that documents the mechanism is doing something rarer than staying silent. It does mean the six sentences below are worth reading exactly as written, on each vendor's own site, checked today rather than assumed from a general sense of "everybody says privacy now."

The reason this specific question gets asked more often now than it used to is not mysterious: a language model that writes a meeting summary has to be trained on something, and the obvious, cheapest source of training data for a company that already sits between millions of meetings and their transcripts is the meetings themselves. That does not make training on customer audio wrong by itself — plenty of software improves by learning from how it is used, and a de-identification process is a real mitigation, not a fig leaf. It does mean the question "does this company train on my calls" is worth asking directly, in those words, rather than inferring the answer from how much a vendor talks about privacy elsewhere on its site. A page that mentions encryption six times can still train on your audio; the two facts live in different paragraphs of the same policy.

tl;dv — "no customer data is used to train the AI"

tl;dv's own security page states this outright, and describes the mechanism rather than just the conclusion: metadata shared with its AI partner, Anthropic, is anonymised — email, company name, personal names stripped — and meetings are chunked into pieces and reordered before anything leaves tl;dv's own systems, so the partner processing a summary cannot reassemble which segments belonged to the same conversation. Its Free tier runs $0, with Pro and Business tiers priced in euros on its pricing page — worth converting before comparing to dollar figures elsewhere. An account can also choose whether processing is hosted in Europe or the US, a control aimed squarely at organisations with a specific regional data-residency requirement rather than at the training question alone. Full comparison here.

Fireflies — a stated line, plus zero data retention

Fireflies' own site states that customer data is never used for AI training or for any purpose outside direct business needs, and pairs that with a stated zero-data-retention policy — meaning the commitment is not only about training but about how long anything is kept at all. Its plans run from free through Pro and Business, detailed on its pricing page checked today. A bot still joins every call to capture it; there is no bot-free recording mode, so the no-training commitment applies to a recording that still had to reach Fireflies' servers first. Zero data retention is worth reading as its own promise rather than a footnote to the training one: it means the company states it does not keep a copy indefinitely, which is a separate question from what it does with a copy while it briefly exists. Full comparison here.

Granola — opt-out, on every tier, on by default

Granola's pricing page, checked today, states that every tier — Basic through Enterprise — can "opt out of model training at any time." Read plainly, that means training is the default state and the opt-out is a setting a user or an organisation has to find. Enterprise is the tier that can flip the setting for an entire team at once rather than one account at a time. Granola is priced from free through $14 a user a month on Business to $35 on Enterprise, and captures the computer's own audio without a bot joining the call — a real strength that sits alongside, not instead of, the training question. Nothing on Granola's pricing page describes what "training" actually improves or how long a recording is retained before or after the opt-out is applied, which is the kind of detail worth asking support for directly if it matters to a specific team's policy. Full comparison here.

Otter — de-identified, not skipped

Otter's own privacy page states that it uses a proprietary method to de-identify user data before using it to train its models, so that an individual cannot be identified in the result — a claim about the output rather than a claim that training does not happen. One clear exception is stated: imported documents, such as files brought in from Google Workspace, are excluded from training entirely. No opt-out setting for the audio-based training is described on that page. Otter's paid tiers run $8.33 to $19.99 a month annually, per its pricing page today. What Otter's page does not state, and this page will not guess at, is whether an individual account can request exclusion from training the way Granola's opt-out or an enterprise data-processing agreement might allow — the de-identification commitment is described as applying broadly rather than as a toggle a single user controls. Full comparison here.

Notta — depends on the plan, and the mechanism runs through a partner

Notta's own privacy documentation states that the service uses third-party partners' APIs for speech recognition, and that depending on the plan a customer is on, those partners may use the customer's audio data for recognition training. No settings page describing an opt-out from that specific mechanism was found on the policy checked today. Notta's pricing runs from a free tier through Pro at $8.17 a month and Business at $16.67, both billed annually, per its pricing page. It offers both a bot-based and a bot-free recording mode, which is a separate axis from the training question entirely — a bot-free recording still routes through the same third-party speech-recognition partners once the audio reaches Notta, so choosing bot-free does not by itself change the training answer. Full comparison here.

MacWhisper — nothing leaves the Mac, so there is nothing to train on

MacWhisper Pro is €64, paid once, confirmed on its own site today, and its transcription runs entirely on the Mac using local models — meaning the training question the five tools above each answer with a policy simply does not arise, because no audio or transcript is transmitted anywhere to begin with. Its own summarisation step is the exception: it connects to an AI provider the user chooses, and once that connection is made, whatever policy that provider states applies to the summary step specifically — which means MacWhisper's own no-training answer is only as strong as whichever provider a user connects it to, unlike the five hosted tools above where one company's policy covers the whole pipeline. Full comparison here.

Coii AudioNotes — no training, because nothing is uploaded at all

Coii AudioNotes is $19, paid once, for three of your own Macs, with a 30-day trial and no account. The recording, the transcript, and the summary — written by a language model bundled inside the app rather than a provider reached over a connection — are all produced on the Mac and none of them are sent anywhere. There is no training policy to read because there is no data path for training to happen on.

Licence activation and a launch-time clock check are the only things that touch a network at all. Full comparison here and here for the two closest written commitments among the hosted tools above.

Reading a training policy without getting it wrong

Three different sentences are doing three different jobs here, and it is worth keeping them apart on purpose. "We do not train on your data," stated by tl;dv and Fireflies, is a claim that the practice does not happen at all. "Opt out any time," stated by Granola, is a claim that a control exists, paired with the fact that the control starts off. "We de-identify before training," stated by Otter, is a claim about the later step in a pipeline that still begins with training. None of the three is dishonest, and a reader who only skims for the word "privacy" on a marketing page will not be able to tell them apart — which is exactly why each one here is quoted from where it was actually written, not from how it is usually summarised.

Why a settings page is not the same as a guarantee

Even a genuine opt-out has a timing problem worth understanding. Switching a setting off stops new recordings from being used going forward; it says nothing, on any of the pages checked here, about whether recordings made before the switch was found were already used. For a team that has been on a hosted tool for a year before anyone asked this question, the honest answer to "was any of our data used" may simply be unknowable from outside the vendor — which is exactly the situation a policy of never uploading the recording in the first place avoids by construction, and the situation a written no-training commitment from day one, like tl;dv's or Fireflies', avoids by never having the ambiguity to begin with.

The table

Which to actually try

If a written "we do not train on customer data" is the specific requirement, tl;dv and Fireflies are the two names here that state it outright today — read both pages yourself, since a policy is the kind of thing worth confirming rather than trusting to a summary, including this one. If the requirement goes further than any policy and the recording simply should never reach a server capable of training on it, MacWhisper and Coii AudioNotes are the only two names on this page where that is true regardless of what any settings page says, because there is nothing transmitted for a setting to govern. Either way, the useful habit coming out of this page is not picking a winner and trusting it forever — it is checking the actual policy again the next time a vendor updates its terms, since a written commitment today is not a promise it stays written the same way next year. A calendar reminder to reread the one policy page that actually matters, once or twice a year, costs less time than this whole comparison took to write and stays accurate long after any single article checked on a specific date in September 2026 has gone stale. The six pages linked above are worth bookmarking directly for that reason — not this summary of them, but the vendor's own words, which is where the answer actually lives and where it will change first if it ever does. None of this is a reason to distrust every vendor by default — several of the sentences quoted above are genuinely strong commitments, written in plain language rather than buried where nobody reads them — it is only a reason to keep checking rather than treating any single sentence, from this page or any other, as settled for good.