Your Phone May Soon Spot Scams Without Uploading Chats

Your phone may soon warn you that the person texting you is running a scam, without sending a single word of your chats to the cloud. Google DeepMind and the Government Technology Agency of Singapore have built a privacy-preserving model designed to spot scams right on the handset, according to DeepMind. The tool, announced this week, watches for manipulative conversation patterns and flashes a real-time warning before anyone hands over money or login details.

The stakes in Singapore are high. Scams now account for more than half of the city-state's reported crimes, and losses have passed four billion Singapore dollars since 2019, according to DeepMind's announcement. The hardest cases are no longer crude phishing links. Attackers now run long, convincing conversations, from fake job offers to bogus investment tips and impersonations of officials, and social engineering of that kind now drives up to four-fifths of all losses.

How the two-stage check works

The system works in two stages, built for battery life as much as accuracy. A lightweight classifier screens incoming messages with minimal power, looking for anything that smells like a threat. Only when something looks risky does the second stage wake up: a fine-tuned version of DeepMind's compact Gemma model reads a larger slice of the conversation to confirm the threat or dismiss the false alarm.

The tuning is deliberately local. The model was trained on anonymised scam reports and intelligence feeds covering impersonations of public institutions, banks and local figures, plus regional schemes trending in Singapore. Generic models often miss those local tricks, DeepMind said, while the tuned version is meant to spot scams while telling ordinary social chatter apart from genuine manipulation.

Consent is baked into the design. When the model ships inside apps, it will only be able to look at on-device conversations with the user's permission, under that app's terms of use. Because everything runs on the handset, DeepMind said, neither the Singapore government nor Google will be able to reach users' conversations through the deployment. Development took place inside a GovTech-managed cloud environment, and training used no private or personally identifiable information.

A phased rollout, starting in Singapore

The project was presented at Singapore International Cyber Week, and the rollout will be phased, developed in partnership with the Singapore government. DeepMind frames the work as a template for messaging platforms caught between two duties: protecting users and respecting end-to-end encryption. Traditional scam filters scan messages on company servers, but encryption means even the platform running the app cannot read what passes through it. Moving the analysis onto the phone itself gets around that wall, since the model reads the conversation where it already lives.

The team also says it wants to explore similar deployments in other countries, and promises more detail soon. That ambition matters because scam tactics travel fast: a scheme that works in one market is usually copied in others within weeks. A model trained to spot scams in one country would need fresh local tuning everywhere it goes, which is why the Singapore partnership, with its access to local scam reports, was the starting point rather than a finished global product.

What still has to be proven

For all the promise, the announcement leaves the most important questions unanswered. There is no public release date, no published detection rate, and no published false-positive rate. No list of supported devices or messaging apps has been shared, and a tool built to spot scams will only help people if app makers actually build it in. Until independent researchers can test the system, trust rests largely on the makers' own claims. Scepticism toward AI safety tools is healthy: even Google's own detector for AI fakes comes with fine print that limits what a clean result proves.

That is the tension at the heart of the project. A tool that reads your private messages to protect you is only as welcome as its promise that nobody else reads them. Mark Chen, director of trust and safety at GovTech Singapore, put the pitch this way: "Fighting scams shouldn't mean sacrificing personal privacy." The design, with on-device processing and consent gates, is built to honour that line. Whether it holds up in practice will depend on details that have not been published yet, including how the warning looks, how often it cries wolf, and who audits the model over time. The question of who bears responsibility when scams succeed has already sparked debate over whether platforms, not victims, should carry the blame.

For now, the practical advice stays the same. A system that can spot scams can buy you a pause, and the pause is the point: the model is designed to prompt people to stop and think before sharing sensitive information or funds. No detector, on-device or otherwise, removes the need for that pause. If a stranger's job offer, investment tip or official-looking message is pushing you to act fast, the safest move is still to slow down, verify through a separate channel, and avoid sending money because a chat told you to. Young jobseekers should be especially careful: fake job scams are already targeting Gen Z, and the fake-recruiter playbook is exactly the kind of long con this tool is meant to catch.