Do AI notetakers train on your data? 13 meeting apps' own policies, side by side
The Morse teamPublished 12 min read
Some do, as of November 2026. Of the 13 AI notetakers we checked, Otter, Fathom, and Granola train their own AI on your meetings by default, using data they call de-identified, and let you opt out. Read AI and Plaud keep their training programs opt-in, though Read AI's policy separately says it uses meeting audio and video to improve its models. Krisp's pricing page says it never trains on your data, though its privacy policy lists improving its own models, with consent. Fireflies, tl;dv, Fellow, Zoom's built-in AI (formerly AI Companion), Microsoft 365 Copilot in Teams, Gemini in Google Meet, and Notion AI Meeting Notes say they don't train on meeting content. Most of them send your transcript to an outside model maker such as OpenAI, Anthropic, or Google, and most say their contracts bar that company from training on your data.
We read each company's own privacy policy, terms, trust center, and help pages on November 17, 2026, and every cell in the table below links to the page it came from. Read AI and Plaud deserve credit for leaving their training programs off until you choose otherwise, and Zoom for naming the model families its AI sends meetings to and offering a setup that uses only Zoom's own models.
Which AI notetakers train on your meetings?
Three of the 13 do by default. The table gives each company's own answer as of November 17, 2026, for personal and small-team plans unless a cell says otherwise.
| Notetaker | Trains its own AI on your meetings by default? | How to opt out | What it says about de-identifying | After you delete | Outside AI models |
|---|---|---|---|---|---|
| Otter | Yes: "training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)" (Otter); Enterprise workspaces "are opted out of AI model training" by default (Otter) | Set Help improve Otter to Don't allow sharing, which covers access "for training and product improvement, including through human review" (Otter) | "A proprietary method to de-identify user data," not described further (Otter) | In Trash for 30 days, then "permanently deleted from our servers" (Otter) | Unnamed "AI Service Provider(s)" that don't train on or store customer data (Otter) |
| Fireflies | No: "We do not use personal information for AI model training" (Fireflies) | None needed; the terms say it "will not use your User Content to train" generative models (Fireflies) | May "derive usage, statistical, and other data" from your content (Fireflies) | Account data deleted "within 30 days" of closing the account (Fireflies) | OpenAI and Anthropic as LLM services, Assembly.ai for transcription (Fireflies), under a "Zero Data Retention" policy (Fireflies) |
| Fathom | Yes, until you opt out: de-identified meeting data trains "our in-house artificial intelligence models" (Fathom) | My Settings; on Team Edition an admin can opt everyone out in Organization Settings (Fathom) | "De-identified data generated from Meeting Content Information" (Fathom) | "Commercially reasonable efforts" to delete within 30 days of an account deletion (Fathom); backups 7 days after that (Fathom) | Anthropic, OpenAI, ElevenLabs, Modal, and other AI providers marked "Forbidden from training AI on Fathom Customer Data" (Fathom); its help center also names Google (Fathom) |
| Granola | Yes on Free and Business: "anonymised data may be used for Granola's own model improvements"; Enterprise is opted out by default (Granola) | Turn off Use my data to improve models for everyone under Settings, then Preferences (Granola) | "De-identified, aggregated data, not Personal Data" (Granola) | Trash empties after 30 days (Granola); data already in its models "will not be removed" (Granola) | Outside providers under "enterprise agreements" that bar training (Granola); not named on its model-training page, but its FAQ gives OpenAI and Anthropic as examples (Granola) |
| Read AI | No: "The default is to opt-out" (Read AI); separately, its policy uses audio and video "to infer certain demographic characteristics about users, which we use to improve our models," without tying that to the setting (Read AI) | Off unless you join its Customer Experience Program in account settings (Read AI) | May disclose "aggregated or de-identified information" (Read AI) | Deleting a report you own removes its data, "and this cannot be undone" (Read AI); audio and video kept "in no case for longer than 2 years" (Read AI) | Anthropic and OpenAI, for "Augmentation of Meeting Data" (Read AI) |
| tl;dv | No, for "foundation models, large language models, or other generative AI models" (tl;dv) | None needed under that policy (tl;dv) | Not addressed for meeting content in its policy (tl;dv) | Free plan recordings kept 3 months; paid plans "until account deletion" (tl;dv) | Anthropic through Google Cloud Vertex AI, plus AssemblyAI and ElevenLabs for transcription (tl;dv) |
| Fellow | No: "Fellow does not use customer data to train AI models" (Fellow) | None needed; admins can switch off outside AI under Workspace Settings, Security, Artificial Intelligence (Fellow) | Uses "Anonymized Platform Usage Data" for analytics (Fellow) | Deleted from its servers early "if you so request" (Fellow) | Anthropic, OpenAI, Deepgram, AssemblyAI, and others, all optional (Fellow) |
| Krisp | Its pricing FAQ says "never" (Krisp); its policy lists "Improving our proprietary AI models" with consent as the legal basis (Krisp) | No setting described on either page (Krisp) | May "anonymize your personal data instead of deleting it" (Krisp) | Kept "for as long as needed to provide Krisp," or anonymized (Krisp) | Unnamed summary providers whose "terms also apply" (Krisp); its list shows Microsoft Azure for "backend support" (Krisp) |
| Zoom's built-in AI (formerly AI Companion) | No: content isn't used "to train Zoom's or its third-party artificial intelligence models" (Zoom) | None needed; the same promise is in its terms (Zoom) | Not claimed (Zoom) | Follows the retention of the Zoom service that made it (Zoom) | Anthropic, OpenAI, and Google Gemini models alongside Zoom's own, with zero data retention outside "trust and safety purposes"; EU and Australian accounts use Zoom-Hosted Models Plus, which keeps data from the model makers (Zoom) |
| Microsoft 365 Copilot in Teams | No, for "foundation LLMs" (Microsoft) | Feedback is optional and managed by admins (Microsoft) | Not claimed (Microsoft) | You can delete your Copilot activity history; admins set retention in Purview (Microsoft) | OpenAI as a subprocessor, and Anthropic models if an admin allows them (Microsoft) |
| Gemini in Google Meet (work and school accounts) | No: content isn't used "for Generative AI model training outside your domain without permission" (Google) | None needed (Google) | Not claimed (Google) | Notes are a Doc in the organizer's Drive (Google); Gemini prompts kept "90 days to indefinite, as determined by admins" (Google) | Google's own Gemini; content isn't shared outside your organization "without your permission" (Google) |
| Notion AI Meeting Notes | No: "By default, your customer data is not used to train models" (Notion) | None needed by default (Notion) | Not claimed (Notion) | Restorable for 30 days, then unrecoverable; embeddings deleted within 60 days (Notion) | Several model providers that keep data "30 days or fewer" on Free, Plus, and Business (Notion); transcription sub-processors "do not store audio" (Notion) |
| Plaud | No, "unless we have obtained your explicit and informed consent in advance" (Plaud) | Stay out of its optional "user experience improvement program" (Plaud) | That program uses "anonymized or aggregated data" (Plaud) | When you deactivate, Plaud will "delete or anonymize" your data (Plaud) | LLM providers its trust center no longer names (it listed GPT, Claude, and Gemini when we read it on November 1, 2026); the policy says they keep data "for a limited period," the trust center says "zero data retention" (Plaud, Plaud) |
Does "we don't train on your data" cover everything?
Not always. Several of the no answers name a kind of model, and the promise stops at that kind.
- Fireflies promises in its terms that it "will not use your User Content to train, retrain, fine-tune or otherwise improve any generative artificial intelligence models." The next sentence lets it "derive usage, statistical, and other data from your User Content" for "analyzing and improving the Services," and the license you grant it is "perpetual and irrevocable" (Fireflies). Its privacy policy goes further and rules out "AI model training" with personal information altogether (Fireflies).
- tl;dv rules out "foundation models, large language models, or other generative AI models" in its policy published September 9, 2026 (tl;dv).
- Microsoft says prompts, responses, and Microsoft Graph data "aren't used to train foundation LLMs." Its page also says Microsoft "may use customer feedback, which is optional, to improve Microsoft Copilot," and keeps that feedback out of the foundation models (Microsoft).
- Google and Notion attach conditions. Google's is "without permission" (Google). Notion's is "by default" (Notion).
- Krisp says on its pricing page that "Your meeting data is never used to train Krisp's AI models" (Krisp). Its privacy policy, last updated September 28, 2026, lists "Improving our proprietary AI models and providing more accurate services" among the uses of your information, with "[consent]" as the legal basis (Krisp). Neither page says where Krisp asks for that consent.
- Fellow's privacy policy, last updated August 18, 2026, never mentions model training (Fellow). Its no-training promise is on its trust center FAQ and its sub-processor page, which says "Customer data is not used to train AI models" (Fellow).
None of this shows any of these companies training on meetings. When you compare notetakers, rely on the sentence in the privacy policy or terms, and check which models it names.
Is de-identified training private?
That depends on a method none of the three companies publishes. Otter 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 the training "is automatic and as such audio recordings and transcripts are not manually reviewed by a human" (Otter). Its privacy policy, effective June 16, 2026, still describes the training data as transcriptions "which may contain Personal Information" (Otter). Fathom and Granola call their training data de-identified without saying how they make it so.
Fathom's help center answers its own question, "Do you train AI models on customer data?", with "No! None of our AI sub-processors (Anthropic, OpenAI, or Google) are contractually permitted to use our users' data to train their AI models." The next sentence reads "Fathom uses de-identified customer data to improve the accuracy of our proprietary AI models in order to improve our service for all users" (Fathom). Granola's FAQ puts both halves in one answer, too: its own models may use anonymised data, and "Third parties like OpenAI and Anthropic are never allowed to train on your data" (Granola). Both answers rule out training by Anthropic, OpenAI, and the other outside providers, then describe training by the notetaker itself.
A product codename, a deal figure, or a described illness can still point to a person or a company after the names are gone. Our pages on whether Otter is private and whether Granola is safe cover each app's other settings.
Does deleting a recording take it out of training?
Not once it's inside a model. Granola's privacy policy, effective September 8, 2026, says de-identified data already built into its models "will not be removed from those models and datasets" (Granola), and its FAQ says it "cannot guarantee" your data wasn't used before you opted out (Granola). Opting out protects the meetings you hold afterward.
Deleting removes the stored copy on each company's own schedule. Otter keeps deleted conversations in Trash for 30 days (Otter), Fathom aims for 30 days after an account deletion but says "we cannot guarantee" it (Fathom), and Krisp may anonymize data rather than delete it (Krisp).
If you were a guest on someone else's call, the recording sits in their account. Read AI lets you ask for deletion of a report you don't own through its Account and Privacy Center (Read AI). Fireflies tells people without an account to "reach out to the meeting host" (Fireflies).
Do OpenAI, Anthropic, or Google see your meeting transcripts?
Usually at least one of them does, because their models write the summary. Zoom lists Anthropic, OpenAI, and Google Gemini models among those its "federated" AI routes requests to, and says "Changes to the models used in federation can occur at any time" (Zoom). Fireflies lists OpenAI and Anthropic as LLM services (Fireflies), Read AI lists the same two (Read AI), Fathom's trust center adds ElevenLabs and Modal to them (Fathom), Fellow lists both as optional AI sub-processors (Fellow), and tl;dv runs Anthropic models through Google Cloud Vertex AI (tl;dv).
Otter (Otter), Fireflies (Fireflies), Fathom (Fathom), Granola (Granola), tl;dv (tl;dv), Fellow (Fellow), Zoom (Zoom), Notion (Notion), and Plaud (Plaud) all say outside providers can't train on your meetings. How long the outside companies keep the data varies more. Zoom says they keep nothing, "with limited exceptions for trust and safety purposes, such as to detect and report CSAM" (Zoom). Notion says its AI providers keep data "for 30 days or fewer" on Free, Plus, and Business plans, while most models on Enterprise keep nothing (Notion). Plaud's own pages disagree. Its privacy policy says data shared with LLM providers "is retained only for a limited period as necessary to provide the service" (Plaud), and its trust center says "All LLMs used in Plaud have zero training and zero data retention policies" (Plaud).
Zoom offers two ways to cut the model makers out. Under Zoom-Hosted Models Plus, which Zoom uses for accounts hosted in the EU or Australia, third-party models run on Amazon Bedrock, and information "is not sent to any third-party AI model providers or service providers except Amazon Bedrock." Zoom-Hosted Models Only drops third-party models altogether, which Zoom says "may result in a reduction of available AI features" (Zoom). Zoom also retired the AI Companion name in a June 24, 2026 post; meeting summaries are now "built right into Zoom Workplace," and ZoomMate is its new agentic product (Zoom). We go further into what Zoom's AI keeps and what Teams' Copilot keeps, and into how the model makers themselves treat your data in which AI companies train on your data.
How do you turn off training in Otter, Fathom, and Granola?
In each app's settings, under a different name: Help improve Otter in Otter's account settings, My Settings in Fathom, and Use my data to improve models for everyone in Granola's preferences (Otter, Fathom, Granola). On Granola's Business plan, "each user needs to opt out individually" (Granola). Otter's pages don't say whether its setting also stops the de-identified training in its privacy policy. The full paths, and the other settings worth changing, are in our pages on Otter, Fathom, and Granola. For Read AI and Plaud, leave their opt-in programs switched off.
How Morse keeps what you do with a transcript out of AI training
Most notetakers above keep a transcript on their servers, and most send it to outside models under contracts you can't read, protected only as far as each policy's wording reaches. Morse is a chat for the work after the call, like writing the follow-up email or working out what a client meant. Those chats are stored in your browser, encrypted, not on Morse's servers.
- Private, for writing the follow-up without a model host keeping it. Every new chat starts here. Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing, and where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
- E2EE, for passages with a colleague's or client's own words in them. Your browser checks the secure enclave's proof that it's genuine, then encrypts the message on your device so only that enclave can read it, and Morse's servers relay text they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE runs open-weight models like GLM, Kimi, and DeepSeek, has no web search or image generation, and comes with Pro at $19 a month.
- Anonymous, to ask GPT, Claude, or Gemini how to word a hard follow-up. Our proxy removes your account, email, and IP address and sends the question under Morse's own account. The provider sees the message without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy, so leave attendees' names out.
Morse doesn't train AI on your chats, sell your data, or show ads (Morse privacy policy). Because we don't store chats, we usually have very little to give authorities. Clearing your browser or deleting your account removes your chats for good, and Morse can't recover them, so keep the original notes wherever your team already keeps them.
Before your next recorded meeting
- Check the training default. On Otter, Fathom, and Granola's personal and Business plans, training is on until you change it.
- Read the policy, not the FAQ. Fathom's help center and Krisp's pricing page give a plainer no than their privacy policies do.
- Ask which models get the transcript. Zoom, Microsoft, Fireflies, Granola, Read AI, Fathom, Fellow, and tl;dv name theirs.
- Treat opting out as forward-only. Granola says data already in its models stays there.
- Sort out consent separately. Our guide to AI notetaker consent laws covers which states need everyone's agreement.
When the call is over, start free on Morse and write the follow-up in a Private chat, or try a chat from the home page without an account.
Common questions
Does opting out of training remove what a notetaker already used?
Not at Granola. Its privacy policy says de-identified data already built into its models 'will not be removed' from them, and its FAQ says it 'cannot guarantee' your data wasn't used before you changed the setting.
Is an Enterprise plan safer when it comes to training?
At Otter and Granola it is. Both say Enterprise workspaces are opted out of training by default, while on Granola's Business plan each person has to turn it off for themselves.
Can I stop someone else's notetaker from training on my voice?
Not from your side of the call, because the host's account settings decide. Read AI lets you ask for deletion of a meeting report you don't own, and Fireflies tells people without an account to ask the meeting host.
Which notetakers name the AI models they send transcripts to?
Of the 13 we read, Zoom, Microsoft, Fireflies, Granola, Read AI, Fathom, Fellow, and tl;dv name outside model makers such as Anthropic, OpenAI, or Google. Otter, Krisp, Notion, and Plaud describe their AI providers on the pages we read without naming them, and Google Meet uses Google's own Gemini.