“Offline” AI should show when it phones home

I’m increasingly wary of AI features marketed as “offline” when that apparently means “try locally, then quietly send the request elsewhere if the laptop struggles.” That fallback may be reasonable: a cloud model can be faster than a weak CPU, preserve battery, or produce a better result. But the privacy boundary changes completely, and the interface should acknowledge it.

I’d like a per-request indicator showing local or cloud processing, the destination, and why fallback happened, plus a policy choice: local-only, ask first, or allow cloud fallback. “Offline-capable” is otherwise too slippery for a privacy claim. Developers can expose readiness and routing checks; power users should be able to verify or block destinations with firewall logging and rules rather than trusting a label.

Should this disclosure be mandatory, or is a clear settings toggle sufficient? I’m happy to be argued out of this—especially by people using current AI PCs or local-model software.

The indicator is useful, but “why fallback happened” may be noisy or misleading. Hardware readiness can change between requests, and the real reason may be a timeout or service policy. I’d require destination plus local/cloud status, then make detailed diagnostics available to developers. The hard control is local-only versus permitted fallback.

I agree on disclosure, but I doubt most users want a prompt every time the NPU is busy. An obvious status icon and a first-run choice probably beats consent fatigue. The default should still be honest: if cloud fallback is enabled, don’t call the feature simply offline.

A toggle is not enough if the application can reinterpret it after an update. Local-only should be enforceable outside the app, with outbound rules and logs. Otherwise the setting is just a promise made by the component that benefits from sending the data.

Speaking as the person who has to ship the feature: graceful fallback prevents a lot of ugly failures, especially on mixed hardware. I’d expose an event and destination in logs, show a small cloud badge, and let advanced users set policy. Mandatory per-request prompts would make ordinary users disable the whole feature.

I’m not convinced “cloud” always means worse privacy or “local” means safe. A local model can still be untrusted software, and a cloud request may be tightly minimized. Disclosure is still worthwhile, but routing alone shouldn’t become a security verdict.