On-Device AI in Browsers: Why "Local" Doesn't Automatically Mean Private

Chrome and Edge have both shipped AI features — translation, summarization, writing assistance — that run using small language models downloaded and executed directly on the device, rather than sending text to a cloud server for every request. That's a genuine architectural shift, and it carries a real privacy improvement over the alternative — with one important caveat worth understanding before treating "local" as a synonym for "private."

What's Actually Running Locally Right Now

Chrome's stable built-in AI APIs — Translator, Language Detector, and Summarizer — run on Gemini Nano, a compact model Google built specifically for on-device use, while more experimental capabilities like Writer, Rewriter, and a general Prompt API are available on an opt-in basis. Edge has followed a similar path, shipping a local model (originally Phi Silica, with a newer Aion-1.0-Instruct model in developer preview) alongside Language Detector and Translator APIs that developers can call directly, supporting over 145 languages without a cloud round-trip.

The Real Privacy Benefit

When inference happens on-device, the actual text being summarized, translated, or rewritten doesn't need to leave the machine to reach a remote server — that's the core structural advantage over a cloud-based AI feature, which by design requires transmitting the input somewhere else to be processed. For anyone concerned about text passing through a third party's servers, this is a meaningful, real improvement, not a marketing distinction without a difference.

"Local" describes where the computation happens, not what the surrounding website does with the text before or after that computation. A page could still capture keystrokes, log form submissions, or send analytics on what was typed — none of that involves the AI model at all, and switching to on-device inference doesn't change any of it. The privacy benefit is specific to the inference step, not a guarantee about everything else happening on the page.

Why This Distinction Matters in Practice

A website using Chrome's on-device Summarizer API could, independently of that API, still run its own tracking scripts, save form data to its own servers, or log analytics events tied to what a user typed — the local AI model has no relationship to any of that, since it only sees the text at the moment it's asked to process it. Local inference is a genuine privacy feature for one specific step in a pipeline; treating it as a guarantee about the entire page's behavior overstates what it actually does.

The Hardware Reality Behind "On-Device"

Chrome's built-in AI documentation lists a minimum threshold of roughly 22GB of free disk space to enable on-device features — developers have clarified this is a storage threshold requirement rather than the actual size of the downloaded models, which are considerably smaller, but it still means on-device AI isn't universally available on lower-storage devices. The practical pattern most production sites follow is attempting local inference first, checking hardware and API availability, and falling back to a cloud endpoint when either is missing — meaning "on-device" isn't yet a guarantee for every user on every device, even on a supported browser.

What Turning It Off Actually Does

Both Chrome and Edge let users disable on-device AI entirely through browser settings, which deletes any already-downloaded local models and stops future background downloads until re-enabled — a meaningful control for anyone who'd rather not have gigabytes of model weights cached locally at all, independent of any privacy consideration about how the models are used.

What This Means for Choosing a Tool

On-device browser AI is a genuine, structural improvement for the specific moment text is being processed by that particular model — but it doesn't say anything about what a given website does with that text before or after, which is a separate question worth checking independently. A tool built around not transmitting text to any server at all, for any purpose, addresses that broader question directly rather than relying on one processing step being local while everything else on the page remains unverified.

---

Browser-native AI moving inference on-device is a real, structural privacy improvement over sending every request to a cloud server — Chrome and Edge have both built genuine local model infrastructure in 2026, not just a marketing label. The caveat worth remembering is that "local" describes one specific step in a pipeline, not a guarantee about the website surrounding it — a distinction that matters more than it might first appear.