TL;DR: An AI dictation app has up to three things that can leave your machine: your audio, the text transcribed from it, and usage data about how you use the app. Most privacy marketing only talks about the first. Before you speak prompts, Slack replies or client emails into any tool, find out where each of the three goes, whether it's stored, and whether there's a mode where none of it leaves your device.
Why should developers care where dictation goes?
Once voice becomes how you prompt Cursor, Claude Code or ChatGPT, your dictation tool hears everything:
- internal project and customer names
- descriptions of security bugs before they're fixed
- architecture details, file paths, sometimes a pasted-from-memory API key
- Slack messages about colleagues, emails to clients
A keyboard never sees any of this as data. A dictation app does. That makes it worth the same scrutiny you'd give a browser extension that can read every page.
What data does an AI dictation app actually handle?
Break it into layers. Each one can be handled differently, even within the same app.
| Layer | What it is | Why it matters |
|---|---|---|
| Audio | The raw recording of your voice | The richest data. It can reveal identity, mood, background conversations |
| Voiceprint | Biometric characteristics of your voice | Treated as biometric data under laws like Illinois' BIPA and the GDPR |
| Transcribed text | What you said, as words | Contains everything sensitive you said, just not your voice |
| AI cleanup | A language model rewriting the text | Often a separate cloud call, sometimes to a third-party model provider |
| Analytics | Events about app usage | Should describe how you use features, never what you said |
The common trap: an app says "on-device transcription", which is true, but then sends the transcript to a cloud LLM for the cleanup step. Your audio stayed home; your words didn't. That might be fine for you. You should just know it.
The 7 questions to ask any dictation app
- Where is speech transcribed? On-device, the vendor's cloud, or a third-party speech API?
- Does audio ever leave the device? Including for "quality improvement", debugging or support?
- If text goes to the cloud, who processes it? The vendor, or a model provider like OpenAI, Anthropic, Groq or Cerebras? That provider's privacy policy applies too.
- Is anything stored server-side? Audio, transcripts, outputs, and for how long?
- Is your content used for training? By the vendor and by any model provider in the chain.
- Is there a fully local mode? One switch that guarantees nothing leaves the machine, for the moments that need it.
- What does analytics collect? Feature usage is normal. Transcript content in analytics is not.
If a privacy policy can't answer these, that's your answer.
How do popular Mac dictation apps handle privacy?
Based on each vendor's own site and docs as of October 2026. Policies change, so verify before relying on this for anything regulated.
| App | Where speech is processed |
|---|---|
| Apple Dictation | On-device for many languages on supported hardware; otherwise Apple's servers |
| Wispr Flow | Cloud (offers SOC 2, ISO 27001, HIPAA with a BAA on business plans) |
| Superwhisper | On-device or cloud, your choice of model |
| VoiceInk | On-device; optional cloud enhancement sends text only |
| Willow | Cloud, with an on-device offline mode |
| Aqua Voice | Cloud; Privacy Mode stores nothing on Aqua's servers |
| Monologue | Cloud and offline transcription; no audio or transcripts saved on its servers |
| Superquill | On-device transcription; only text goes to the cloud, or nothing in Private mode |
There's no single "most private" answer. A cloud tool with a signed BAA and SOC 2 can be the right call for a healthcare team, and a fully local tool can be the right call for a solo dev who never wants to think about it. Match the tool to your threat model.
How Superquill handles it (and what it doesn't claim)
Disclosure: I build Superquill, a voice-to-text app for Mac in private beta. Here's how it answers the seven questions, including the parts that aren't "nothing ever leaves":
- Audio and voiceprints never leave your Mac. Speech recognition runs locally. Recordings aren't uploaded, transmitted or stored by us or any provider.
- In standard mode, the transcribed text is sent to Cerebras for the AI step that turns rambling speech into a short message or prompt. Text only, never audio. Cerebras' own privacy policy applies to that text, and we link to it.
- Private mode keeps everything on-device. Turn it on and the text isn't sent to Cerebras or any other cloud service. Toggle it for the sensitive stuff, like security bugs or client details.
- We don't train our own models on your content.
- Analytics (PostHog) are for feature usage, like which features you use and app version, not what you said or wrote.
The design goal was simple: the most sensitive layer (your voice) never leaves, the next layer (text) leaves only when you allow it and only to one named provider, and there's a one-switch way to keep everything local. The full list of providers and what each one receives is in the privacy policy.
If you dictate prompts to coding agents, see my guide to voice prompting for Cursor and Claude Code.
Practical habits for dictating safely
- Never dictate secrets. API keys, passwords, tokens. Not into any tool, local or cloud. Type them, or better, reference an env var.
- Use a local mode for sensitive work. Incident write-ups, vulnerability details, anything under NDA.
- Read the cleanup provider's policy, not just the app's. The app's promise only covers the app.
- Check your company policy. Many teams already have rules for sending code or customer data to third-party AI. Dictation falls under them.
- Re-check after updates. Features like "cloud sync" or "history" can quietly change what's stored.
FAQ
Is voice dictation safe for confidential work?
It can be, if speech is transcribed on-device and you use a mode that keeps text local too. Cloud dictation can also be acceptable with the right agreements (SOC 2, a BAA, a no-retention policy), but you should confirm those exist before dictating confidential data.
Do dictation apps store my voice recordings?
Some do, some don't. On-device apps like Superquill, VoiceInk and local Superwhisper models don't need to send audio anywhere. For cloud apps, check the privacy policy for audio retention and whether recordings are used to improve models.
What is the difference between on-device transcription and on-device processing?
On-device transcription means your audio is turned into text locally. The text may still be sent to a cloud language model for cleanup or rewriting. Fully on-device processing means both steps happen locally and nothing leaves the machine.
Is my dictation used to train AI models?
It depends on the vendor and on any model provider they use. Look for an explicit statement covering both. Superquill does not use your content to train its own models.
Which dictation app works offline on a Mac?
Apple Dictation (for supported languages), Superwhisper with local models, VoiceInk, Willow's offline mode, and Superquill in Private mode can all work without sending your speech to the cloud.
What's your line for AI tools at work: anything goes, local only, or approved vendors only? Curious how other teams are handling this. 👇
Top comments (0)