Back to Blog
Industry Insight•2026-11-16•12 min read

Does My AI Note Taker Train on My Data? 6 Vendors Compared

Does My AI Note Taker Train on My Data? 6 Vendors Compared
TL
Team Laxis
Laxis Team @ Laxis

Does my AI note taker train on my data? It depends on the vendor, and on some plans it does unless you switch it off. Several popular note takers use customer meetings, usually in de-identified form, to improve their own models, while others say they never do. Below is what six vendors publish about it today, and ten questions to put to any of them before you install, because by the time you find the setting, earlier meetings may already have been used.

Granola is the example people are reading about this autumn. Its security and privacy FAQ says that on the Free and Business plans, by default, anonymised data may be used for its own model improvements, that Business users opt out one by one, and that it can't guarantee data wasn't used before you changed the setting. A proposed class action, Chamberlain v. Granola, alleges that this default, together with recording people who weren't told, breaks California and federal privacy law; nothing has been proven, and our article on the Otter and Granola lawsuits tracks where the case stands. The lesson for buyers doesn't depend on the outcome: defaults matter, and an opt-out may not reach backwards.

Whether everyone had to agree to being recorded in the first place is a separate question, covered in our guide to recording consent laws across states and countries (general information, not legal advice). This piece covers the other half of AI note taker privacy: what the vendor does with your meetings once it has them.

What "we use your data to improve our models" really means

Every cloud note taker has to process your meeting to transcribe and summarise it; that's the service. Training is different: using your meeting to improve the product for every customer. Privacy policies often blur the two.

De-identified doesn't mean anonymous, not in a meeting. Vendors that train usually say they strip identifiers first. Otter's privacy policy, for example, says it trains its proprietary AI on de-identified audio recordings and on transcriptions "which may contain Personal Information". Removing names from metadata doesn't remove them from what people said. The plaintiffs suing Otter question whether its de-identification works.

Opt-outs usually work forward, not backward. Switching a setting off stops future use; it doesn't recall what has already gone in.

A trained model can't simply forget. Granola's own privacy policy says that once training is complete, it is not technically possible to isolate or extract specific data from the resulting model. That's how machine learning generally works, which is why the default matters more than the opt-out.

The people you meet with don't get a switch. Training settings belong to the account holder. A client or candidate on the far side of the call has no account, no setting and often no idea which tool is running. Whether they had to agree to being recorded at all depends on where they are.

More than one company may be involved. Most note takers send audio to outside speech and language-model providers. Several vendors say publicly that those providers can't train on customer data. If yours doesn't, ask.

The 10 questions to ask any AI note-taker vendor

Paste these into an email to the vendor's sales or security team. A trustworthy vendor can answer each in a sentence or two.

  1. Do you train AI models on my meeting content, and if so, is that on by default or opt-in? "We don't" is the cleanest answer; opt-in is the next best. "On by default, switch it off in settings" means every meeting before you found the switch may already be in.
  2. Does one opt-out cover my whole workspace? On some plans each person has to opt out separately, which in a team of forty means forty people remembering. Look for an admin-level control.
  3. Is opting out retroactive, and can you remove my data from a model you've already trained? From anyone who trains, expect an honest no to the second half, said up front.
  4. Do your AI and transcription sub-processors train on my data, and does the contract say so? Ask for the sub-processor list and the clause itself, not a line from a landing page.
  5. Where is my data processed and stored? Country and cloud provider. If you need EU or other regional residency, find out which plan offers it, if any.
  6. How long do you keep recordings, transcripts and notes, and what happens when I cancel? "Indefinitely, unless you configure a policy" is common. Ask how long deletion takes after you close an account, and whether a free plan hides older meetings or deletes them.
  7. Who can see my notes by default, and are share links public? Private by default is what you want. Then check whether an "anyone with the link" option exists, who can switch it on, and whether admins can restrict it.
  8. Do you create voiceprints, and is that off by default? Voiceprints can be biometric identifiers under laws such as Illinois's Biometric Information Privacy Act. Off by default, with consent before anyone's voice is enrolled, is the answer to look for.
  9. What certifications do you hold, and what do they cover? SOC 2 Type II tests controls over months; Type I checks their design at one point in time. Make sure the certificate is the vendor's own, not just its cloud host's, and ask about a BAA if you handle health information.
  10. What control do people in my meetings who don't use your product have? Can a participant ask for their data to be deleted, and how? Often the honest answer is "through the person who recorded you", which is worth knowing before a client asks.

Quick tip: get the answers in writing and keep the email. Privacy pages change; most of the competitor policies we cite below were revised during 2026, and a dated reply is your record of what you were told when you signed up.

Curious how we answer them? Question 1 is answered in writing: Laxis will never use your data to train AI models without your consent. Install Laxis and look through every setting before your next call. 14-day Premium trial, no credit card.

How six note takers answer, in their own words

We checked each vendor's own privacy policy, help center or security page on November 16, 2026, and recorded what each one publishes. "Not published" means we didn't find an answer on those pages, not that the answer is bad.

VendorTrains its own models on your meetings?What it says about its AI providers trainingRetention and deletionVoiceprintsCertifications
LaxisNot without your consent. Its Privacy Policy (Section 3) says recordings, transcripts, notes and dictation aren't used to train AI models without consent, and training is never on by defaultIts pledge covers what Laxis does, not its providers; subprocessor list on requestDelete items any time; deleting your account deletes your content immediately, backups included; custom retention on Enterprise; Basic archives meetings older than 30 days rather than deleting themOnly if you turn on Voice Recognition, which is off by defaultSOC 2 Type I (per its privacy policy); not HIPAA, no BAA
GranolaYes, anonymised data, by default on Free and Business; users opt out individually; Enterprise is opted out by default; no guarantee about data used before opting outSays OpenAI and Anthropic are never allowed to train on your data, under enterprise agreementsNotes and transcripts kept indefinitely by default; any user can set transcripts to auto-delete, and Enterprise admins can set a workspace policy; audio deleted after transcriptionNot publishedSOC 2 Type II; HIPAA with a BAA on Enterprise (desktop apps)
Otter.aiYes. Trains its proprietary AI on de-identified audio and transcriptions; its privacy pages don't state a default or a self-serve opt-outSays no customer data is used to train its AI service providers' modelsNo fixed period: kept as long as needed for the policy's purposes or as required by law; deleted conversations leave the Trash after 30 daysGenerates speaker identification information; lists Speaker Identification / Biometric Information as a categorySOC 2 Type II; HIPAA compliance as a paid Enterprise add-on
FirefliesNo. Says it doesn't use personal information or meeting data to train AI modelsSays its AI vendors are contractually barred from training on it and keep no meeting content after processingKept while the account is active; deleted within 30 days of closing it; custom retention on EnterpriseService providers may derive voice characteristics, which Fireflies says aren't used to identify anyone; its policy has an Illinois biometric section with a retention scheduleSOC 2 Type II, GDPR; HIPAA-ready on Enterprise with Private Storage and a signed BAA
FathomDepends on your settings: may use de-identified meeting data to train in-house models unless you opt out in account settings; default not statedSays it doesn't authorise OpenAI, Anthropic, Google or others to train on itKept until you delete your account; aims to delete within 30 days of the requestOnly for iOS users who opt inSOC 2 Type II, HIPAA (standard BAA on its Team plan), GDPR-aligned
tl;dvNo. Doesn't use customer content to train foundation, large language or other generative modelsSays its AI providers don't use customer content to train their general-purpose models3 months on free accounts, until account deletion on paid ones; stored in the EU, with AI processing in the EU or US depending on the hosting location you chooseNot publishedSOC 2 Type II, GDPR

Sources, checked November 16, 2026: Laxis privacy policy (updated November 10, 2026), No-Training Pledge, help center security page and pricing page; Granola security, privacy and data FAQ, model training page and privacy policy (effective September 8, 2026); Otter privacy policy (effective June 16, 2026), privacy and security page and pricing; Fireflies privacy policy (updated September 28, 2026) and security page; Fathom privacy policy (updated August 16, 2026) and pricing; tl;dv privacy policy (September 9, 2026) and security page.

Credit where it's due. Fireflies and tl;dv say plainly that customer meetings don't train their models. Granola, Fathom, Otter, Fireflies and tl;dv all publicly limit how their outside AI providers can train on your data, though the exact promise varies (the table has each one); it's worth asking to see the clause. Granola deletes audio after transcription, and Fathom takes voiceprints only from iOS users who opt in. Several hold stronger certifications than we do today: SOC 2 Type II and HIPAA programs, where Laxis states SOC 2 Type I and offers no BAA. For full write-ups, see our detailed Otter.ai review and our ranked list of 2026's AI note takers.

Red flags and good signs in a vendor's answers

  • Good sign: a one-word answer to question 1 that also appears in the privacy policy, not just in a blog post.
  • Good sign: admin controls for training, retention and sharing that cover the whole workspace.
  • Good sign: a published deletion window, such as 30 days after an account is closed.
  • Red flag: "we may use data to improve our services" with no word on whether that's on by default or how to stop it.
  • Red flag: silence on whether opting out reaches meetings you've already recorded.
  • Red flag: certifications that belong to the data center rather than the company you're buying from.

Regulated firms have more to check than this list. Our guides for lawyers choosing an AI note taker and financial advisors choosing an AI note taker cover privilege, confidentiality and record-keeping.

Where Laxis fits

Here's how Laxis answers the ten questions, limits included.

  • Your meetings are not training data. Laxis will never use your data to train AI models without your consent. Section 3 of our Privacy Policy puts that in writing for recordings, transcripts, notes and dictation: training is never on by default, never a condition of using Laxis, and any future program would be opt-in. The reasoning is in our No-Training Pledge. That answers questions 1 to 3.
  • No bot by default, plus a nudge to disclose. With the Laxis AI meeting assistant, capture happens on your own computer, and before you record the app prompts you to let the other participants know. Prefer the tool to tell people for you? The Laxis Chrome extension lets participants know a recording is running, and the optional AI Assistant Bot joins as a named, visible participant.
  • Private until you share, and encrypted. According to the Laxis help center, meetings are private to you until you share them with specific people, a folder, your workspace or anyone with the link (workspace admins decide whether new meetings are shared with the workspace by default), and encryption covers data both at rest and in transit. Our privacy policy states SOC 2 Type I compliance. Those cover questions 7 and 9.
  • Voiceprints only if you ask for them. Also per the help center, Voice Recognition ships switched off, and switching it off later deletes every enrolled voice across your devices (question 8).
  • Retention you can plan around. Enterprise plans get custom data retention policies and data residency. On the free Basic plan, meetings older than 30 days are archived, not deleted, and upgrading unlocks them. You can delete items whenever you like, and deleting your account deletes your content immediately, backups included (question 6).

Where Laxis falls short: audio is processed in Laxis's cloud rather than on your device, and Laxis isn't HIPAA compliant or able to sign a BAA, so keep patient information elsewhere. Our help center doesn't name a default storage region, so ask about question 5 if it matters to you. And our pledge covers what Laxis does, not the terms of our AI and transcription providers; the subprocessor list is available on request, so put question 4 to us in writing, exactly as you would to any other vendor.

The bottom line

Ten questions can feel like a lot for a note-taking app. But a meeting tool hears more of your company than almost any software you buy: pricing, hiring, strategy, and what people say when they think it's just the room.

Meeting notes that aren't training data. No bot by default, private until you share, and never used to train AI without your consent. 14-day Premium trial, no credit card. Try Laxis free

Frequently asked questions

Does my AI note taker train on my data?

It depends on the vendor, and some train on customer meetings by default. Granola says anonymised data may be used for model improvements by default on its Free and Business plans, and Otter says it trains on de-identified recordings and transcripts, while Fireflies and tl;dv say they don't train on customer meeting content. Check the privacy policy for a stated default and an opt-out.

Is Granola safe to use?

Granola has solid security basics, but check its training default before you rely on it. It holds SOC 2 Type II, deletes audio after transcription and says OpenAI and Anthropic are never allowed to train on your data. On Free and Business plans, anonymised data may be used for its own models unless you opt out, and an unproven class action is pending.

Can I get my data removed from an AI model after it has been trained?

Usually not: once a model has been trained, specific data generally can't be pulled back out of it. Granola's documentation says so directly, and the point applies to machine learning in general. Opting out of training typically protects only future meetings, which is why the default setting matters more than the opt-out. Choose a vendor that doesn't train, or switch training off on day one.

Do AI providers like OpenAI train on my meeting transcripts?

That depends on the note-taker vendor's contract with them, which you usually can't see. Granola, Fathom, Otter, Fireflies and tl;dv each publicly limit how their outside AI providers use customer meeting data for training, though the wording differs: tl;dv's promise, for example, covers the providers' general-purpose models. A vendor's promise about its own models doesn't automatically cover its sub-processors, so ask every vendor, including us, for the sub-processor list and the relevant clause in writing.

Does Laxis train AI models on my meetings?

Laxis will never use your data to train AI models without your consent. Its Privacy Policy says recordings, transcripts, notes and dictation are not used for training unless you opt in, and training is never on by default. Laxis records without a bot by default, reminds you to tell participants and keeps meetings private until you share them. It is not HIPAA compliant.

What happens to my recordings when I cancel an AI note taker?

It varies: some vendors delete your data within a set period after you close your account, while others keep it until you delete it yourself. Fireflies says it deletes account data within 30 days of closure, and Fathom aims to delete within 30 days of a request. On Laxis, dropping to the free Basic plan archives older meetings rather than deleting them, while deleting your account deletes your content immediately.

Related Product

Laxis Meeting Assistant

Automate your notes today.

  • Bot-free notes for Zoom, Google Meet, Teams and in-person meetings
  • AI summaries, action items and follow-ups in 60+ languages
  • Mac, Windows, iPhone and Android, with a free plan

Start for free, no credit card required.

Learn More ▷