Singapore’s scam-warning model reads conversations on the phone, not in the cloud

What changed

Google DeepMind and Singapore’s GovTech announced an on-device model designed to identify conversational scams. Rather than send private messages to a server for analysis, the system processes the conversation on a compatible phone and can surface a warning when it detects manipulation before a person shares credentials or money.

The team fine-tuned a lightweight Gemma model on local scam conversations, including impersonations of public institutions, banks and local figures. The system uses a two-stage pipeline: a low-power screen first, then a larger conversation analysis when patterns look suspicious. The goal is to catch social engineering that simple keyword filters miss while preserving end-to-end encrypted message privacy.

What the result means

The announcement frames this as a delivered tool, but does not publish a model accuracy or false-positive rate. That is the key next metric to watch: a warning needs to catch real scams without training people to ignore alerts. The approach is also specific to Singapore’s scam language and tactics, so performance elsewhere would require local data and evaluation.