Do AI note takers train on your interview recordings? What 13 vendors say, and how to opt out

Some do and some don't, and the answer can change with your plan and settings. As of October 2026, Otter, Fathom and Granola say on their own sites that they use de-identified or anonymized customer data to improve their own AI models; Fathom and Granola let you switch that off in settings, and Otter offers an organization-wide opt-out on Enterprise. Metaview, BrightHire, Fireflies, Zoom and Microsoft Teams' intelligent recap say they do not use customer content to train AI models, tl;dv rules out generative models, Greenhouse says "not currently" for its Notetaker, and most vendors that describe their AI providers say those providers are barred from training on your data.

The table below quotes each vendor's own privacy, trust or help page, with the setting a recruiting team needs to change where one exists. Storage, retention and zero data retention are separate questions, covered in zero data retention for AI note takers and where AI note takers store interview data.

Which AI note takers train on customer data, and which let you opt out?

"Its own models" means the note-taker vendor's in-house models. "Third-party providers" means the AI model and speech-to-text companies the vendor sends your data to. Where a page does not say, the cell says so.

Vendor wording changes. Before you rely on a row, open the linked page and check its date against the one shown here.

What does "de-identified" or "anonymized" training mean for an interview?

It means the vendor says it removes or masks details that identify people before the data trains a model. The vendors above do not publish how they do it, and an interview is harder to anonymize than a sales call. A candidate says where they worked, which school they went to, which city they live in, the size of the team they ran and the project they are proudest of. Remove the name and the combination can still point to one person.

Read the exact words, too. Otter's policy covers transcriptions it says "may contain Personal Information", which is broader than de-identified audio. Greenhouse separates personal data, which it says it will not use for training even if its policy changes, from anonymized data it may use under the customer agreement. Ask each vendor three things: how transcripts are de-identified, whether people review the data, and whether your organization's data can be excluded entirely.

Who can switch training off: the recruiter, the admin or only sales?

Where an opt-out exists, it sits at one of three levels, and the level decides whether one setting covers your whole team.

• Per user. Fathom's My Settings opt-out and Granola's Settings toggle are set by each person. One recruiter who forgets leaves their candidates' interviews in the training pool.
• Per organization or workspace. Fathom says Team Edition organizations can opt out all users from Organization Settings, and Sembly's toggle sits in workspace settings. This is the level a recruiting team wants.
• By contract. Otter's Enterprise page says to contact your account manager to keep your organization's data out of training, and Granola's Enterprise users start with training off. Get the confirmation in writing with its effective date.

A five-step check for a recruiting team

1. List every note taker in use, including free personal accounts that recruiters or hiring managers set up themselves. A personal account follows that person's settings, not the company's.
2. Find the training setting for each using the table above, and switch it off at the highest level the product allows.
3. Screenshot the setting with the date and keep it with your vendor file.
4. Ask about data already collected. Does the opt-out cover interviews recorded before the switch, or only new ones?
5. Recheck every six months and whenever the vendor emails a privacy policy update.

Does an opt-out cover candidates who never signed up?

Candidates are not the vendor's customers, so they have no settings of their own. Their interview is processed under the account of the recruiter or company that ran the note taker, and your setting is the only one that applies. That is why the organization-level switch matters: a candidate cannot opt out of training on a tool they never chose.

If you hire in the EU or UK, the vendor is normally your processor, and Article 28(3)(a) of the GDPR requires a processor to process personal data only on the controller's documented instructions. Training a vendor's own models is not something most recruiting contracts instruct, so it needs to be dealt with expressly; the DPA guide for recruiting software covers the clauses. In California, the privacy agency's FAQ confirms that the CCPA's protections include job applicants. This is general information, not legal advice.

What should be in writing?

A settings toggle can change with a product update; a contract term cannot without notice. Add these points to your security review, which slots into section 4 of the AI note taker security questionnaire:

1. Whether the vendor uses your content, including transcripts, summaries and audio, to train or improve any model, its own or anyone else's.
2. Whether "anonymized", "de-identified" or "aggregated" data is carved out, and what each word means in their process.
3. Which AI and speech-to-text providers receive your data, and the contract term that stops each from training on it.
4. The level of your opt-out (user, organization, contract), its effective date and whether it covers data collected before it.
5. Notice before any change to the training terms, and the right to delete your data if you do not accept the change.

An example clause to adapt with counsel, not a recommended standard:

Model training. Supplier shall not use Customer Data, including
interview audio, transcripts, notes, summaries and scores, or any
data derived from them, whether or not de-identified, to train,
fine-tune or improve any machine learning model, except models
used solely to provide the Services to Customer. Supplier shall
bind each subprocessor to the same restriction and shall give
Customer [30] days' written notice of any change to this clause.

What should you tell candidates?

Only what is true for your setup. If your vendor says it does not train on customer content, or you have switched training off at the organization level and have it in writing, one line in your notice is enough: "Our note-taking tool's provider does not use interview content to train its AI models." If you cannot confirm that, leave the line out rather than guess. Wording for the rest of the notice is in the candidate privacy notice template, and reusing your own recordings to train interviewers is a different question, covered in can you use interview recordings for training.

Keeping less data also shrinks the question. We make Interview Signal, which keeps transcripts, quotes, candidate names, notes and submittals on the recruiter's computer, sends audio to a speech-to-text provider in short chunks during the call without retaining it, and keeps only counts and durations on its server; its security page lists the providers so you can ask them the same questions.

Which questions does this page not answer?

• How long each vendor keeps recordings. Training and retention are separate; a no-training vendor can still keep audio for years. See where AI note takers store interview data.
• What the AI model provider keeps. That is zero data retention, explained in its own guide.
• Whether a recording is legal at all. Consent comes first, under the recording law where you and the candidate are, whatever the vendor's training policy.

Not legal advice. Vendor policies are quoted from their own pages as of October 2026 and change often; check each linked page before you rely on it.