Who Actually Hears Your Meetings? AI Notetaker Privacy

TL;DR: Most AI meeting assistants send your audio, your transcript, or both to a cloud you don't control. Otter.ai and Fireflies.ai face class-action suits alleging privacy violations involving recorded voices. A bot that joins the call, a desktop app that streams audio out, and a truly on-device recorder create different privacy situations. This guide maps who hears what, tool by tool, with sources.

Everpaper is our product, and it appears in this guide. Where it does, the claims come from our own privacy documentation, which you can verify. Where competitors appear, the claims come from their privacy pages and court filings, linked inline.

The three architectures

Every AI notetaker on the market in 2026 is one of three things, and the differences matter more than any feature list.

1. A bot joins your call. Otter.ai and Fireflies.ai work this way. Their assistant appears in the participant list of your Zoom, Meet, or Teams call — visible to everyone, auto-joined through your calendar. The audio goes straight to their cloud, the transcript lives in their cloud, and every guest on the call just had their voice captured by a company they never chose.

2. A desktop app streams your audio out. Granola is the honest example here — no bot, nothing joins the call, and its docs say so plainly. But the audio still leaves: its security page names Deepgram and AssemblyAI for transcription and OpenAI and Anthropic for notes, with transcripts stored in a US-hosted AWS VPC. Your computer is the microphone. Their cloud is the tape.

3. Everything stays on the machine. The capture, the transcription, the summaries — all local. This is the category Everpaper was built into: the speech model runs on the Mac's Neural Engine, and the docs' guarantee is short enough to quote: "No provider, no traffic."

Who hears what, tool by tool

Otter.ai. The bot joins as a visible participant and auto-joins via calendar sync. Audio, transcripts, and the bot's screenshots upload to AWS in the US. The privacy policy (effective June 16, 2026) reserves the right to train "proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)." Otter is also the defendant in In re Otter.AI Privacy Litigation — a consolidated class action in the Northern District of California alleging it recorded meetings without all-party consent, in violation of ECPA, CIPA, and BIPA. The motion-to-dismiss hearing was held May 20, 2026 before Judge Eumi K. Lee. Otter claimed 35 million users as of December 2025. (Full details and sources in our Otter.ai vs Everpaper comparison.)

Fireflies.ai. Same bot architecture, same calendar auto-join default, and its own lawsuit: Cruz v. Fireflies.AI, a BIPA voiceprint class action filed in December 2025 in the Central District of Illinois, ongoing as of this writing. The free tier stores 400 minutes of meetings per team on Fireflies' servers. (More in Fireflies.ai vs Everpaper.)

Granola. The best of the cloud group, and worth being precise about. No bot, no stored audio — those are real. But transcription happens at Deepgram or AssemblyAI, notes happen at OpenAI or Anthropic, transcripts live in AWS, and Granola trains its own models on anonymized user data by default on consumer tiers (opt-out in Settings; off by default on Enterprise). Its Series C in March 2026 valued it at $1.5B, so this model has plenty of runway — it just isn't an on-device one. (More in Granola vs Everpaper.)

Limitless. The cautionary tale. Limitless (formerly Rewind) marketed itself on privacy while storing recordings in a TLS/HSM cloud. Meta announced its acquisition on December 5, 2025; the Pendant is no longer sold, desktop and web recording are disabled, and existing customers get the Unlimited plan free with support "throughout 2026." A privacy promise is only as durable as the company making it. (More in Limitless vs Everpaper.)

Reflect. A note app, not a meeting tool, but instructive: its notes are genuinely end-to-end encrypted (XChaCha20, audited by Doyensec), yet its AI features send selected text and audio to OpenAI, Anthropic, and Google — and its privacy policy states those providers "may use prompts we send them as training data." Encryption at rest and privacy in the AI loop are separate questions. (More in Reflect vs Everpaper.)

Everpaper. Microphone and system audio record locally as separate, time-aligned tracks. Transcription runs on-device — Parakeet (parakeet-tdt-0.6b-v3) by default, Whisper or MOSS as alternatives — on the Neural Engine. The transcript, the audio, and the knowledge graph live in ~/Library/Application Support/HindsightNext/ and nowhere else. If you connect a cloud AI provider for summaries, that's behind a toggle that ships off (Allow cloud AI to analyze your journal), and a redaction pass scrubs API keys, tokens, card numbers, SSNs, and OTPs before anything goes out. Connect a local provider instead — Ollama, LM Studio, oMLX, or Apple Foundation Models — and the AI features work with zero traffic off the Mac. That's the whole model, and it's written down in the privacy docs.

The consent problem the industry keeps stepping on

The suits overlap on consent, but their claims are not identical. In re Otter.AI Privacy Litigation alleges ECPA, CIPA, and BIPA violations tied to recording meetings without all-party consent. Cruz v. Fireflies.AI alleges Fireflies collected and stored participants' voiceprints without the notice and written consent BIPA requires. These are allegations, not findings; the Otter motion-to-dismiss hearing took place on May 20, 2026, and the Fireflies case remained ongoing at the research date.

A 2022 Hacker News report described an Otter bot joining two confidential meetings uninvited and emailing attendees an "absurd, inaccurate" summary afterward. That report is an account from one user, separate from the later court allegations.

No-bot tools fix the announcement problem but not the destination problem: your guests' voices still end up at Deepgram, AssemblyAI, OpenAI, or Anthropic, companies none of you are in a contract with.

The on-device path sidesteps both. Everpaper records your Mac's audio with the standard macOS permissions you granted it, the way any local recorder does — the consent conversation is between you and your guests, exactly as it would be with a tape recorder, and no third party enters the room at all. For anything sensitive, Hide my voice and Hide any speaker (⌘4) strip speakers from the record, and the hiding persists across re-analysis.

Five questions to ask any AI notetaker

Before you connect anything to your calendar, get answers to these — from the privacy page, not the landing page.

1. Does a bot join the call? If yes, your meeting guests are being recorded by a vendor they never picked. Consent duties depend on jurisdiction, and the suits above show why that deserves explicit review.
2. Where does the audio go, exactly? "We take security seriously" is not an answer. Named subprocessors (Deepgram, AssemblyAI, OpenAI, Anthropic) with a storage location is an answer.
3. Is the audio stored, or just the transcript? No stored audio (Granola's model) means no playback verification; stored audio (Otter, Fireflies) means a bigger target.
4. Do they train on your data, and is that opt-in or opt-out? Read for the word "de-identified" — it usually sits right next to the permission you'd rather they didn't have.
5. What survives the company? After Meta announced the Limitless acquisition, existing users were promised the Unlimited plan free and support "throughout 2026." Local data under your home folder survives any cap table.

The short version

The safest meeting transcript is the one that never leaves the room. On-device tools make that literal: the audio stays in your home folder, the transcription happens on your chip, and "delete" actually means delete.

Everpaper is in private beta with limited spots, and a founding-license option for early supporters — not a monthly subscription. If the on-device column of this guide is where you want to live, join the waitlist at everpaper.app and read the Meetings and Privacy docs while you wait — they're written to be audited, not admired.