Google DeepMind and Singapore's Government Technology Agency (GovTech) have teamed up to build a scam-detection model by fine-tuning the lightweight Gemma. It analyzes conversations directly on the phone and, when it spots something suspicious, shows a real-time alert on the device itself. Messages are not uploaded to the cloud. The project was shown at Singapore International Cyber Week (SICW) 2026, and DeepMind described it in a post on October 14. It falls under the two sides' "National AI Partnership" framework.
Singapore has real pressure behind this effort. According to figures DeepMind cites, scams account for more than half of all reported crime in Singapore, with cumulative losses since 2019 exceeding S$4 billion, and sophisticated social-engineering scams account for as much as 80% of the losses. Common tactics include impersonating public agencies, banks and local acquaintances, and investment scams and fake job offers are part of the mix.
Two-stage detection
The model works in two tiers. The first is a lightweight, fast screening classifier that picks out suspicious conversations. Only after a hit is the conversation passed to the second tier, the fine-tuned Gemma, which analyzes it across multiple turns to decide whether manipulative tactics are at work. DeepMind says the design is meant to save battery, so the phone does not have to run a larger model on every message.
Privacy is the point the scheme leans on most in its public messaging. The model reads on-device conversations only with the user's consent and in line with the terms of use of the app involved. According to both sides, neither the Singapore government nor Google can see users' conversations through this deployment. Fine-tuning took place in a Google Cloud environment managed by GovTech, with the training data kept isolated. It used an anonymized dataset compiled from scam reports and intelligence sources, with personally identifiable information removed; testing relied on synthetic conversations and local-language material.
GovTech's Director of Trust and Safety, Mark Chen, put it this way:
"Fighting scams shouldn't mean sacrificing personal privacy."
What has not been disclosed
The article gives neither the model's parameter count nor any metrics such as accuracy or false-positive rate. There is no rollout date either; it says only that the effort will proceed "in phases and with caution." Which apps will carry it, and whether it will be folded into GovTech's existing ScamShield anti-scam app, has not been stated officially and cannot be verified for now. The two sides say they will explore bringing similar technology to other countries.
The false-positive rate matters a great deal for a product like this. If alerts pop up too often, users quickly switch them off; if a high-value scam slips through even once, reputation takes the hit from the other side. Singapore's government has long been willing to publish anti-scam data, so whether it can release hit rates and false-positive rates after the phased launch could be the first solid material for outsiders to judge how good the project really is.
A view from China
Users in China are no strangers to anti-scam alerts. The National Anti-Fraud Center app and the call blocking offered by carriers and phone makers, judging from public descriptions, mostly rest on number tagging, warning lists and databases of scam-linked URLs. They judge who is calling and where a link points. Singapore's approach judges what the conversation itself says, which is content-level recognition aimed squarely at social-engineering scams that use clean numbers and polished scripts.
Content-level recognition cannot avoid privacy concerns, and the reason the two sides give is exactly privacy, which is why the model was pushed down to the device. Phone makers in China have been promoting on-device large models for the past year or two, and running a model of a few hundred million to several billion parameters on a phone is no longer unusual, so the technical conditions are in place. The hard part is authorization and the boundaries of responsibility: who has the right to read chat content, and who is accountable when the judgment is wrong. Singapore's division of labor ties the right to read to user consent and app terms, with the government and Google supplying only the model. That approach is a useful reference for similar projects in China.
Sources: Google DeepMind official blog, CocoLoop, GovTech Singapore public materials; the partners, the two-stage detection flow and the basis of the scam-loss statistics were checked.