DeepMind and GovTech build a Gemma detector for Singapore chat scams

Singapore already spends serious engineering on scams. ScamShield filters suspicious SMS and blocks bad numbers on the handset. GovTech’s SATIS system hunts scam websites at web scale. In May 2026, the Monetary Authority of Singapore opened a proof-of-value with five banks, GovTech and the police to test AI on historical transaction data.

The money still walks out after a conversation. Often that conversation sits inside WhatsApp or Telegram, where end-to-end encryption keeps cloud scanners outside the thread.

On 14 October 2026, Google DeepMind and the Government Technology Agency of Singapore said they had fine-tuned a lightweight Gemma model to watch for that pressure on the phone itself.

DeepMind and GovTech announced a local Gemma fine-tune at SICW

DeepMind published the collaboration on 14 October under its National AI Partnership with the Singapore Government, timed to Singapore International Cyber Week at the Sands Expo and Convention Centre. Authors JJ Geewax and Miguel de Andrés describe a model trained on local scam conversations: impersonations of public institutions, banks and local figures, plus regional schemes.

DeepMind’s crime framing is blunt. Scams make up more than half of Singapore’s reported crime and more than S$4 billion in losses since 2019, the lab writes, with conversational social engineering accounting for up to 80% of those losses. The Straits Times separately reported S$410.6 million lost in the first half of 2026.

Josephine Teo, Minister for Digital Development and Information, told SICW that devices able to recognise manipulation early could help people avoid financial or credential loss while keeping personal messages on the device. She also sketched the multi-stage path many victims travel: social media, messaging, a phone call, then a bank transfer, with no single defender seeing the whole chain.

WhatsApp and Telegram sit outside ScamShield's SMS filter

ScamShield is already installable. Official guidance updated in May 2026 says the app uses an on-device algorithm plus Singapore Police Force blacklist data to filter scam SMS from unknown numbers and block scam calls. Saved contacts are left alone. That is a live product with a permission screen citizens can open today.

Google Messages, separately, already runs on-device scam detection for SMS, MMS and RCS conversations with non-contacts, and Phone by Google offers a related call feature. Those tools cover Google’s own rails.

Straits Times reporting on the DeepMind project names WhatsApp and Telegram as the messaging surfaces the new model is meant to analyse in real time. CNA says Google framed the work for Android devices. DeepMind’s own post talks about phased deployment “within applications” without naming a store listing.

The people who must eventually bind this model are the teams that control those encrypted apps, or whatever government-adjacent Android shell is asked to wrap them. Banks already sit later in Teo’s chain. They are watching for transfers, after the chat has done its work.

The two-stage handset pipeline targets the encrypted-chat gap

DeepMind describes a power-aware pipeline built for phones. A lightweight classifier screens inbound messages at low cost. When something looks suspicious, the fine-tuned Gemma model reads a larger snapshot of the conversation and either confirms a threat or drops a false alarm. A confirmed hit is meant to show a real-time warning telling the user to pause before sharing funds or credentials.

The privacy design is the reason the project exists. Cloud message scanning fails against end-to-end encrypted chats. DeepMind says analysis is meant to stay on-device, that an integrating app may touch conversations only with user consent under that app’s terms, and that neither Google nor the Singapore Government will be able to read those chats through the deployment. Mark Chen, GovTech’s director of trust and safety, framed the same trade-off: catch local scam attempts without opening private communications.

Training happened upstream of that promise. Fine-tuning ran inside a GovTech-managed GCP environment on anonymised datasets drawn from scam reports and intelligence feeds, DeepMind says, with stress tests on synthetic dialogues and local corpora. Those are training-governance claims. A shipped permission model on a handset is a separate artefact.

Place the announcement beside the rest of Singapore’s stack and the novelty sharpens. SATIS, per GovTech’s public page, reviews hundreds of thousands of websites a day and disrupts malicious sites through Google Web Risk and police warnings. Teo’s SICW figures put the daily scan load above 600,000 sites and monthly disruptions around 20,000. MAS’s May proof-of-value asks whether shared historical banking data can flag higher-risk accounts and payments before losses settle. The DeepMind model is aimed at the chat that persuades someone to initiate the payment those systems later see.

Phased rollout still lacks ScamShield's installable bar

Partners say rollout within applications will be carefully phased with the Singapore Government. They have not named the first application, published a launch date, listed eligible devices, or released detection rates, false-positive rates or language coverage. The Straits Times notes it remains unclear whether the model would become a default Android feature. DeepMind says installation details will come later, and that it has “proven the ability” for local real-time detection. That phrase is company language. There is no public benchmark table behind it.

Android focus comes from secondary reporting and Google press materials. DeepMind’s post does not harden an OS exclusivity claim. Treat iPhone coverage as unsettled.

False alarms are the product risk that matches the privacy pitch. A warning inside a private chat that fires too often will be dismissed. A warning that never fires will be trusted past the evidence. ScamShield already publishes how citizens set permissions and where filtered SMS go. Until the Gemma project clears a comparable bar, the SICW announcement is a developed model awaiting a host.

Price the chat layer after the named host app appears

Security and product teams should file the architecture: local fine-tune, two-stage edge inference, consent-gated access aimed at encrypted messengers that ScamShield’s SMS filter and SATIS’s web disruption do not see. Then hold the ship claim until a named application shows consent screens and publishes, or at least measures, false-alarm behaviour.

Banks can keep treating chat-side warnings as an earlier signal that might eventually complement MAS’s transaction work. Other governments DeepMind says it may approach later can ask for the same missing packet before copying the slide.

Singapore has already built installable and background layers for SMS, calls and scam sites. The chat layer is the open slot. DeepMind and GovTech described how a Gemma model could fill it on the handset. The next public fact that matters is which app opens first, and how often that warning is wrong.