On-device AI scam detection lands where Singapore lost S$913m in a year

Google DeepMind and GovTech Singapore showed an on-device scam detector. A light classifier screens incoming messages on minimal power, and a fine-tuned Gemma model reads the whole multi-turn conversation when the first stage flags something. Nothing leaves the handset, the training set came from anonymised scam reports with personal data stripped, and the fine-tuning ran inside GovTech's own managed cloud environment.

“The model runs on the device, and neither Google nor the government reads the conversation”

— Google DeepMind, On-device scam detection announcement

What Singapore lost

Singapore police counted 37,308 scam cases in 2025 with losses of S$913.1m, down 17.9% from S$1,112.4m a year earlier. That averages S$24,475 per case in a country of six million people, and the fall came after two years of bank controls, call blocking and a national reporting app rather than from any model.

Crypto carried S$182.2m of those losses, 20% of the total, and tether, ether and bitcoin made up 91.7% of that slice. Investment scams drove 38.4% of the crypto losses, about S$70m. The scam types the model covers read like that list: investment fraud, fake job offers, impersonation of officials, bank impersonation and the social engineering that holds them together.

Why the window matters

Recovery rates explain the push toward earlier detection. The Anti-Scam Command clawed back S$140.5m: S$117.7m in fiat, which is 16.1% of fiat losses, and S$22.8m in crypto, which is 12.5% of crypto losses. Money that moved on-chain came back less often.

Impersonation of government officials is the category Google names first and the one that grew fastest, with losses up 60.5% to S$242.9m, a quarter of everything Singapore lost to scams last year.

Police and partners averted at least S$348m by reaching people mid-scam, 38.1% of what the country actually lost. The on-device model aims at the same window and arrives earlier, without a human on the line. Amazon put a similar check into its shopping assistant in September, and the US and UK signed a joint pact against scam compounds the week before that.

What settles after launch

Three things settle after this ships. Whether the 2026 brief shows fewer message-initiated cases. Whether Google publishes a false-positive rate. Whether any of it reaches phones outside Singapore. A classifier that marks a real conversation as a scam costs the user something as well, and nobody has published that number.

Informational material, not investment advice. The loss figures come from the Singapore Police annual brief for 2025, and the detection model has no published accuracy or false-positive rate yet.