On-device AI privacy
On-device AI is more private because nothing leaves the device. When the model runs locally, the prompt you type, the documents or photos it reads and the answer it writes all stay on your phone or laptop. There is no request to a server, so there is no server log, no stored conversation and no copy in a data centre for an attacker, an insider or a legal demand to reach.
Apple put the principle in one line when it introduced Private Cloud Compute: data that exists only on user devices is by definition disaggregated and not subject to any centralized point of attack. A breach of one phone exposes one person. A breach of a cloud AI service can expose everyone who used it.
Google describes its on-device model the same way. Gemini Nano runs in Android’s AICore system service and lets apps deliver generative AI without needing a network connection or sending data to the cloud. Google says AICore isolates each request and does not store a record of the input or the output after processing, and that AICore has no direct internet access of its own.
On-device AI data leaks
The model can be local while the app is not. On-device inference protects the step where the model reads your data; it does not control what the app does before or after.
- Results. An app can run the model locally and then upload the summary, the tags or the extracted fields to its server.
- Cloud fallback. Some features run small requests on the device and send larger ones to a cloud model. Apple, for example, sends some requests to Private Cloud Compute, which it designed so personal data is not retained after the response and so researchers can verify its claims.
- Telemetry and analytics. Usage events, crash reports and third-party SDKs can carry fragments of what you did.
- Model downloads and updates. The device fetches models from a server, which reveals that the feature is in use even if no prompt is sent.
- Sync and backup. Chat history saved to a cloud backup leaves the device even if the model never did.
OWASP lists sensitive information disclosure among the top risks for LLM applications, and every path above is a way for it to happen even when the model itself runs locally.
AI feature privacy checks
Checking an AI feature’s privacy means asking where each piece runs and where each piece of data goes.
- Turn off the network and try it. If the feature still works, the model runs on the device. If it fails or degrades, part of it runs in the cloud.
- Read the privacy label and policy. Look for what is collected, whether prompts or outputs are stored, and whether they are used for training.
- Look for a cloud fallback setting. If the app offers a way to keep processing on the device, turn it on.
- Check telemetry settings. Turn off analytics you do not need.
- For your own apps, decide field by field. Keep prompts, documents and outputs on the device unless a feature needs them elsewhere, and send the smallest form that works.
Local inference on a nearby device
When one device cannot run the model, a device nearby can run it instead of the cloud. A phone may be too small for the model a task needs, but a laptop, a workstation or a gateway box on the same site often is not. Sending the request to that device keeps the prompt on your local network, among devices you control.
Offline Protocol is built for exactly this. A device with a loaded model can advertise an inference capability, and nearby agents and apps discover it and send classification or generation requests over Bluetooth LE, Wi-Fi or local IP, multi-hop across the mesh, with no cloud access. Each device proves itself with an Ed25519 key it generated, and once two devices have a confirmed MLS session, request and response bodies are sealed inside it end to end. The first request to a provider a device has never met travels signed but readable by the devices that carry it, so the docs say to keep secrets out of first-contact requests and check the encrypted flag on each response. One model can serve a whole team on site, and the prompts never leave the site.