Local AI Privacy: “Runs on Your Device” Is Only the First Question
Use this checklist to verify where an AI task runs, what files it accesses, which optional cloud features it uses, and how telemetry works.
“Local AI” can mean that model inference happens on your computer. It does not automatically mean that every part of the application is offline, no data leaves the device, or the software has no access beyond the prompt you enter. Treat local execution as one fact to verify, then map the rest of the data path.
Separate the model from the surrounding service
Ask what runs locally: the model download, the inference request, file indexing, speech processing, embeddings, account sign-in, update checks and optional web search can be separate components. An application may offer a local model while still providing cloud-based features you can enable later.
Read the vendor’s documentation and settings for the exact version. Avoid relying on a marketing label or a screenshot of a toggle. Documented behavior can change with a release, model provider or account tier.
Check access before you give it a folder
An assistant that can read a folder may expose relevant file contents to its local index or model workflow. Limit access to the smallest folder needed, use copies or test files for an unfamiliar application, and avoid treating a “local” label as permission to point the tool at an entire home directory or work share.
The same rule applies to browser extensions and connectors. A local model does not make a connected cloud account local. Review permissions separately, especially when a tool can retrieve or modify content.
Look for outbound features and telemetry controls
Check the privacy policy, diagnostics setting, update mechanism and network-related preferences. Some applications explain telemetry and offer a setting; others may not. A network connection alone does not prove prompts are uploaded, but it is a reason to read the documented behavior rather than assume.
If the work is regulated, confidential or organizationally sensitive, follow the organization’s approved tools and data-handling process. This guide is a consumer checklist, not a compliance determination.
Keep the hardware question separate
Local capability also depends on the model size, memory, software and supported accelerator. A CPU, GPU or NPU label does not establish that a specific model will run privately or well. Use Better Things Daily’s CPU/GPU/NPU explainer to understand the hardware roles, then confirm the chosen application’s current requirements.\n\n## Sources\n\n- https://learn.microsoft.com/windows/ai/\n- https://www.nist.gov/privacy-framework\n\n*Cover photo: Huy Phan (@huyphan2602) / Unsplash License. Contextual photograph; not a product test or installation instruction.*
This article may contain affiliate links. If you purchase through them, Findory may earn a commission at no extra cost to you. Read our disclosure.
