AI app privacy audits, from their own documents.

Fathom trains its in-house models on de-identified data from your meetings unless you switch that off, and none of its documents state which position a new account starts in. Against that it publishes a 30-name subprocessor list, contractual no-training terms with its AI vendors, and encryption in transit and at rest.

Read the audit

A bot joins the call, and everything it captures is processed in the United States across a 17-name subprocessor list, six of which would see meeting content. The defaults arrive at their widest.

Read the audit

Granola captures meeting audio on your computer, with no bot in the call, then transcribes and summarizes it in its own cloud. Training is on by default outside Enterprise, and the opt-out is one self-serve toggle.

Read the audit

Used for noise cancellation alone, Krisp says no audiovisual data leaves the device. Turn on the AI Meeting Assistant and, per the Privacy Policy, recordings, transcripts and summaries may be stored on Krisp servers, auto-share ships on, and the no-training promise is split across documents that each cover something different.

Read the audit

Notion states in its AI terms that it does not use your content to train the models behind Notion AI, with carve-outs for feedback and permission. The same promise is written five times across five documents, and the subject or the scope changes each time, while the binding DPA carries no training clause at all.

Read the audit

Every recording goes to Otter's cloud and is stored on AWS in the United States, and outside Enterprise no way to opt out of AI training is documented. Otter also publishes real security work: SOC 2 Type 2, encryption, and 2FA.

Read the audit

Training on your meetings is opt-out by default and the bot is built to be noticed, both documented by Read AI. Meeting reports still auto-share with everyone invited by default, and the Privacy Policy lists categories of personal information Read AI shares and sells to marketing and advertising partners while its marketing page says it sells nothing.

Read the audit

A dictation app that can run entirely on your device, with a written no-training promise and no documented sharing surface. The qualifiers: cloud modes exist, Super Mode reads by default, the policy is dated 2024.

Read the audit

tl;dv promises never to train AI on customer data, and since the July 1, 2026 version that promise is a clause in the Privacy Policy itself, written more broadly than either marketing page. The gaps sit elsewhere: tl;dv's own sources state its SOC 2 status three ways, and an August 2026 researcher disclosure about meeting metadata is contested by the vendor.

Read the audit

Cloud dictation with no offline mode, and training on your dictation is the default outside Enterprise. Privacy Mode turns it off on every plan, free included, against a certification claim the help center does not support.

Read the audit