Google's opt-in Messages setting has Gemini judge what unknown senders want

In the example Google used, a text from an unknown sender pretends to show the recipient a photo of them with their dog. Told it has the wrong number, the sender tries to keep the exchange going and asks the recipient's name ^[3]. As Engadget describes it, the exchange has no link and no request for money ^[3].

The new Messages setting is built to read threads like that one. "Scammers often disguise their true motives behind slow-burn conversations," Google said in its update announcement ^[4]. According to The Verge, the setting sends conversations with an unrecognized sender to cloud-based Gemini models that evaluate the sender's intentions ^[2].

Google describes the cloud step as an addition. In the same statement it said: "This secure cloud analysis augments and extends our existing on-device detection, allowing us to identify novel and sophisticated attacks, actively uncovering and stopping abusers who try to hide their tracks." ^[12] Intent reading is a second layer over the existing detection. Because the setting is opt-in ^[1], anyone who never opens it stays on the first layer alone.

The pitch is software that understands what a stranger wants. The thing being done is a toggle that a person has to find and switch on ^[1]. The Verge reports that the same mode also runs the Gemini Nano check of unfamiliar URLs before they are opened, so the link check comes with the cloud step ^[10].

The one usage figure in either report belongs to a different feature. Google said Android has blocked over three million fake calls since launching its bank-call verification in June, built with banks in Brazil ^[9]. The reports do not include an accuracy or false-positive rate for the message analysis, or say whether an administrator can set the toggle on managed phones.

We'd turn it on for any group of phones that regularly hears from numbers outside the contacts list, and accept that those threads go to Google's cloud on the strength of Google's own privacy description. Two questions sort every group of phones: how often it hears from unknown numbers, and whether your rules allow its messages to be processed in Google's cloud.

- Frequent unknown senders, cloud processing allowed: switch it on, and ask users to report what it flags and what it gets wrong. - Frequent unknown senders, cloud processing not allowed: leave it off and rely on the existing on-device detection. There is no Gemini check to substitute. - Rare unknown senders, cloud processing allowed: little to gain, so switch it on only if the group asks. - Rare unknown senders, cloud processing not allowed: leave it off.