Chrome: How private is local AI in Google's browser?
Current version of the browser contains a local AI model that will not only download data in the future but also be allowed to communicate with Google's servers.
(Image: Google / heise online / dmk)
Google is integrating a local AI model into its browser, which is intended to perform functions such as fraud detection more privacy-friendly. Furthermore, a dedicated Javascript API (application programming interface) will allow websites to interact with the model. Now, a change in the AI settings in Chrome 148 raises questions: Does the local AI phone home and reveal private data to the search engine giant?
A few days ago, various media reported on the chaos surrounding unsolicited data downloads, and now Redditors have compared the local AI configuration settings between Chrome 147 and the recently released Chrome 148. They discovered that a short but important half-sentence is missing.
Videos by heise
Your data to Google servers?
In Chrome 147, it still read: “[Link auf https://webmachinelearning.github.io/prompt-api/][..] use AI models that run directly on your device without sending your data to Google servers,” but the last half-sentence is missing in the current Chrome version. We were able to confirm this on an editorial device running macOS; the option can be found in the browser's system settings (chrome://settings/system).
The purpose of the change is unclear. Google might actually want to reserve the right to enrich the local AI with results from cloud-based LLMs, or it could be preparation for the “Prompt API,” which is apparently being rapidly integrated into the browser in Mountain View.
Skeptics can disable the local AI model, but Google notes in the Chrome Help: While this saves storage space, it limits AI-based features. These could include assistance with writing or rewriting, fraud warnings, organization of browser tabs, or summarization of web pages.
(cku)
