Google DeepMind said on 14 October 2026 that it built a scam detection model with GovTech, Singapore's government technology agency. The model is built to run on typical smartphones in Singapore. DeepMind says scams make up over half of reported crime there, and long, convincing conversations now cause up to 80% of losses. Many chat apps use end-to-end encryption, so servers cannot read messages. DeepMind says that is why the check happens on the phone. It tuned its small Gemma model on local scam conversations.
The design has two stages. A light filter screens incoming messages and uses little power. Then the tuned Gemma model reads more of the chat to confirm a threat or clear a false alarm. DeepMind says this split is how it handles the battery limit. It says training excluded personal data. The post does not say how the model reaches encrypted messages. It also does not say how accurate the model is, how often it raises false alarms, or when users get it.
For teams that build or buy messaging or phone software, the open question is what the scan can see, not where it runs. Useful questions for a vendor: where does the software read plain text, how are false alarms measured, and what accuracy evidence exists? DeepMind's post gives no test results.