“Local AI” is a useful description only when you know which part is local. The model, chat interface, document search, speech service and optional web tools can be separate components. A local model can sit inside an application that also calls external services. Treat privacy as a route to inspect rather than a badge to trust.

Identify where inference happens

Ollama supports local and cloud models. Its API documentation lists separate local-server and direct-cloud addresses, and a local installation can also be signed in to use cloud-hosted models. Seeing a local application window is therefore not sufficient evidence that the request stays on your computer.

Ollama’s cloud documentation says cloud inference runs on its service. Its current cloud FAQ says prompts and responses are processed transiently, not stored or logged as content, and not used for training; it also describes limited account and usage metadata. Those are provider policy statements, not a claim that cloud inference happens inside your home.

For each app you use, find the equivalent documentation. Record the selected model, provider and any enabled online feature. If the interface hides those choices, use harmless sample text until you understand them. A convincing answer cannot tell you which network route produced it.

Look beyond the model

Make a short inventory of the other components. Where is chat history saved? Are uploaded documents copied into an index? Does speech recognition use an external API? Can the app search the web, run a tool or send an error report? These questions do not mean every app behaves badly; they help you find the settings that actually govern your information.

Check the operating system and backup arrangements too. A locally saved conversation can be copied by a cloud backup or visible to another account with access to the same files. Deleting a chat from one interface does not establish that every copy has been removed.

Test with disposable information

  • Create a short document containing invented names and sample figures.
  • Run the intended task with that document before introducing real information.
  • Inspect the app’s model and service settings and its documented history locations.
  • Check what remains after closing and reopening the application.
  • Review optional online tools separately before enabling them.

Disconnecting a network can help establish whether a particular operation needs a connection, but it does not prove what the app stored or transmitted earlier. Prefer documented controls and an understandable setup to assumptions based on an offline test alone.

Choose the boundary that fits the material

For ordinary public drafts, a cloud service may fit your needs. For private records, reduce the information used, remove unnecessary identifiers and choose tools whose behaviour you understand. Avoid putting passwords or recovery secrets into experimental prompts. Local processing can reduce the need to upload information, but maintaining that boundary still depends on the surrounding software and the way it is configured.

Sources & further reading

Checked 17 Nov 2026. Source-based reporting and guidance; individual test results are identified where available.

Found something that needs correcting?

Tell us the article title, the detail and a supporting source.

Corrections & contact →