Google tweaks Chrome AI privacy wording, insists processing stays on-device

Google has updated the privacy wording in Chrome’s settings related to its on-device AI capabilities, removing a phrase that explicitly stated user data would not be sent to Google servers. The original message assured users that AI models powering features like scam detection ran entirely on their devices without transmitting data to the cloud. The recent change sparked concerns among privacy advocates, who questioned whether this indicated a shift in how Google processes AI interactions. However, Google clarified that the update does not reflect any change in data handling, emphasizing that all AI processing remains on-device. The wording adjustment was noticed amid the rollout of Chrome’s new Prompt API, which allows web pages to interact programmatically with a browser-resident AI model. This development coincided with increased awareness that Chrome has been downloading Google’s 4GB Gemini Nano AI model locally on users’ devices since early implementations in Chrome 126. The presence of this model, which supports security features and developer tools, raised fears that user prompts and responses might be captured by Google, representing a potential privacy regression. Google countered these concerns by highlighting that the model operates offline and that user data is not sent to the cloud. Since 2024, Google has offered Gemini Nano as a lightweight, on-device AI model to enhance security functions such as scam detection. While the model requires local storage space, Google has introduced options for users to disable and remove it via Chrome’s settings, especially on devices with limited resources. This approach reflects Google’s ongoing effort to balance AI functionality with user control and privacy assurances, even as it expands AI integration within its browser ecosystem. The controversy underscores the sensitivity around AI privacy and transparency, particularly as browsers increasingly incorporate advanced AI features. Clear communication about data handling and user options remains critical to maintaining trust, especially when significant AI models are stored and run locally on consumer devices.