Chrome built-in AI: what runs locally and what to check first

Separate the model from the website

Chrome documents built-in AI APIs using models that run on supported desktop and laptop systems. Its current guidance lists device requirements, an initial model download and APIs at different availability stages. It says using the model itself sends no data to Google or a third party. [1]

That statement describes the model operation. It does not establish the behavior of every website that might use it. A page can have its own account system, synchronization, analytics or remote processing. Before entering a document, look for an explanation of the complete feature rather than treating the words “local AI” as a privacy guarantee.

Check availability before buying hardware

Start with the exact task: summarizing a page, rewriting text or translating a passage. Check the documentation for the relevant feature and browser version. A demonstration working on one computer does not establish that every phone, laptop or browser can run it.

Do not buy extra memory or a new computer solely because a promotional page says it supports AI. Compare the specific application's requirements with your current system and ask whether the feature is available to ordinary users or only in a test. Keep the source link and review date with that comparison, since software availability can change independently of your hardware.

As checked on October 10, 2026, Google lists at least 22 GB of free space for its foundation-model APIs, plus either a GPU with more than 4 GB of VRAM or a CPU with at least four cores and 16 GB of RAM. Audio input in the Prompt API requires a GPU. These are eligibility conditions, not performance guarantees; the separate Translator and Language Detector APIs have their own scope. [1]

Try a document you can safely verify

Our suggested first test uses a short text you wrote yourself and that contains no private information. Include a date, a name and a clear instruction. Ask the tool to perform one task, then compare the output with the original. For a summary, check whether it preserved the important conditions rather than merely sounding fluent.

Repeat with a passage containing an exception or a negative statement. This is a way to learn the tool's limits for your use, not a general accuracy benchmark. Keep original and output separate so that a rewritten sentence is not later mistaken for a direct quotation.

Ask what happens when local processing is unavailable

Before depending on a service, look for its explanation of what happens if the model is not ready or your device is unsupported. Does it show an error, disable the feature or offer a remote alternative? Those outcomes have different implications for the information you are about to submit.

A useful interface should make that choice understandable. If it does not, start with non-sensitive material and ask the provider for clarification. This guide does not claim that every Chrome AI feature works offline or that every site using Chrome is private. It offers a checklist for evaluating a particular feature and the application around it.