Privacy · 4 min read · Published May 7, 2026
What data leaves your wrist?
Most "AI assistant on Apple Watch" privacy policies say roughly the same thing. Reading them carefully reveals what's actually different. Here's how to decode them, and what we do.
The shape of any voice-AI request
When you press a microphone button on Apple Watch and ask a question, four things happen, in order, in any third-party AI app:
- Audio capture. The mic records while you speak. The audio data lives in the app's process on the watch.
- Speech-to-text. The audio is turned into text. This usually happens in the cloud (because watch CPUs can't host an STT model that's competitive with cloud ones), but increasingly happens on-device.
- Model inference. The transcribed text is sent to a large-language-model provider, which generates a reply.
- Reply. Text streams back to the watch, optionally turned into audio for playback.
The privacy questions are: which of these steps happen on-device, which go to a third party, and how long any data is kept.
Three questions to ask any AI app
1. Is the audio retained after the reply?
The honest answer for almost every cloud-based AI assistant is: the audio itself is not retained. There's nothing to gain from keeping it once the transcription is done. But "not retained" should be in the privacy policy, in writing. If a policy is silent on retention, it probably means the answer isn't great.
Watch Assistant: "Audio is processed transiently and is not retained on our servers after the response is returned." That's a load-bearing sentence — find it in the privacy policy.
2. Are transcribed prompts used to train AI models?
This is the one that matters most and that policies are most evasive about. An LLM provider will sometimes train on customer data unless you opt out. A consumer app reselling that LLM may or may not opt you out. Look for language like:
- "We do not use your prompts to train AI models." — clear, good.
- "We may use aggregated, anonymised data to improve our service." — squishy. Anonymised prompt text can still be uncomfortably identifying.
- Silence on the topic. — assume the worst.
Watch Assistant: "We do not use your prompts to train AI models." It's in our Terms as a hard contractual commitment.
3. Is conversation history stored on the company's servers?
Some apps store every conversation forever, server-side. Some store nothing — history lives only on your device. Some live in the middle, with a sync option. Check whether deletion in the app actually deletes from the server, or just from your device.
Watch Assistant: Conversation history is on your watch (and your iPhone, when paired). We don't store it server-side. Deletion in-app is real deletion.
Things that should be on-device but often aren't
Two ambient signals that are easy to leak unnecessarily:
- Location. If you ask a location-aware question ("coffee near me"), the assistant needs your coordinates. But it should request a single fix, send the minimal coordinates required, and stop the GPS radio. Apps that just say "Always" location and keep updating in the background are doing it wrong.
- Health and motion data. An AI app has zero reason to read HealthKit unless it's explicitly a health-tracking app. If a third-party Apple Watch AI assistant requests Health permission, that's a red flag.
How to read the policy in 90 seconds
Search the document for these terms:
- "retain" / "retention" — find what's kept and how long.
- "train" — find the AI-training carve-out.
- "sub-processor" / "third party" — see who else handles your data.
- "sell" — should be "we do not sell". If it's not, walk away.
- "advertising" / "tracking" — should be absent or explicitly negative.
If those five words show clean answers, the rest of the policy is almost always fine.
The honest catch
No third-party AI assistant runs entirely on your wrist. To answer "what's the boiling point of water at 3000 metres", something somewhere in the cloud has to actually do the inference. That's what a privacy policy is for: it tells you exactly where the data goes and what's done with it. The trade-off is real and worth thinking about — and we'd rather you read our policy and decide than not.
If you want the short version, our Privacy Policy is one page. If you want a deeper read, the guide to using AI on Apple Watch covers what the experience is like in practice.