Privacy-FocusedDictationonMac:WhattoActuallyAsk

Search for a privacy-focused dictation app and you will find a dozen products all using the same vocabulary — secure, private, encrypted, your data is yours. Almost none of it distinguishes between products, because almost none of it answers the only question that matters: does my audio leave this machine, and if so, who receives it?

This is a guide to asking that question well, including where Warp lands on it.

The split that actually matters

Dictation tools divide cleanly into two architectures, and nearly every meaningful privacy property follows from which one a product uses.

On-device tools run a speech model locally. Audio never touches a network. The tradeoff is model size — what fits and runs fast on a laptop is meaningfully weaker than what runs in a datacenter, particularly for accented speech, technical vocabulary, and anything multilingual.

Cloud-routed tools stream audio to a provider. You get stronger models, better handling of hard audio, and capabilities like live translation that are impractical on-device today. The cost is that your voice — including whatever you happened to say near the microphone — reaches a third party.

Neither is universally correct. But a product that will not tell you which one it is has answered the question.

Where Warp sits

Warp is cloud-routed, and it is worth saying that plainly rather than in the passive voice.

• Dictation audio leaves your Mac. While you hold the shortcut, audio streams to a speech provider for real-time transcription.
• Translated and rewritten text leaves too. Live translation and the selection assistant route text to an inference provider.
• Configuration stays local. Shortcuts and preference metadata are stored on the device.
• Capture is explicit. There is no always-on listening mode for core usage — audio is captured while you hold the trigger and not otherwise.

If your work involves material that cannot go to a third-party processor — regulated health or financial data, privileged legal material, client work under a confidentiality obligation that names sub-processors — then Warp is not the right tool for that material, and no configuration changes that. Apple's on-device dictation is the better answer, and saying so costs us nothing. The full routing breakdown is on the privacy page.

Questions worth asking any vendor

"Does audio leave the device?"

Not "is it encrypted" — everything is encrypted in transit, and it tells you nothing about who is on the other end. Watch for answers that describe transport security when you asked about destination. That substitution is usually deliberate.

"Which third parties receive it?"

Most dictation products do not run their own speech models. They call somebody else's. That subprocessor's retention policy is your retention policy, and a vendor who cannot name theirs has not thought about it.

"How long is it retained, and is retention separate from training?"

"We do not train on your data" is the most commonly offered reassurance and one of the least informative, because it says nothing about storage. Ask about both, separately.

"What is captured when I am not dictating?"

A tool with microphone access can, in principle, listen whenever it likes. Ask what triggers capture and how you can tell it is happening. A visible listening indicator is not a nicety — it is the only thing standing between you and having to trust a claim you cannot verify.

"What does Accessibility permission let this app see?"

Accessibility is broader than microphone access and gets discussed less. It permits reading the contents of other applications' text fields and typing into them. For a dictation tool that is load-bearing and legitimate — it is how text gets delivered — but you should know you granted it and why.

Habits that help regardless of tool

• Do not dictate secrets. Passwords, keys, and tokens should be typed or pasted from a manager, never spoken into any transcription pipeline.
• Watch the indicator. Build the reflex of knowing whether the microphone is live.
• Separate contexts. If some of your work cannot go to a cloud processor, use a different tool for that work rather than trusting yourself to remember mid-sentence.
• Re-check after updates. Permission grants and provider routing can change between versions.

Frequently asked questions

Is Warp’s dictation processed on-device?

No. Warp streams audio to a cloud speech provider for real-time transcription, and routes translation and selection-assistant requests to a separate inference provider. Shortcuts and preference metadata stay local. If on-device processing is a requirement for you, Apple Dictation’s offline mode on Apple Silicon is the appropriate tool.

What is the most useful question to ask a dictation vendor?

Ask where the audio goes, not whether it is secure. "Encrypted in transit" is true of essentially every product and tells you nothing about whether your voice reaches a third party. The question that discriminates is: does audio leave the device, and to whom.

Does "we do not train on your data" mean the data is not stored?

No, and the two get conflated constantly. Training and retention are separate policies. A provider can decline to train on your audio while still retaining it for a period, and can retain nothing while a vendor in front of it logs everything.