DeepMind's On-Device AI Spots Scams, Keeps Chats Private

 

By TechPulse USA News Desk | October 14, 2026

The most dangerous scam of 2026 does not arrive with a dodgy link. It arrives as a friendly stranger in no hurry. A few warm messages, a little flattery, a manufactured emergency — and by the time anyone smells a rat, the money is already gone.

On Wednesday, Singapore announced it wants to put a detective inside that conversation. Google DeepMind and the city-state's Government Technology Agency (GovTech) are building an on-device AI model that reads your chats as they happen, spots the manipulation, and warns you before you send a cent — without a single message ever leaving your phone.

If it works, it solves the nastiest trade-off in online safety: the tools that can see a scam are the tools that can read your private life. Singapore's answer is to stop looking from the outside and move the watchman indoors.

The pitch: a detective that never phones home

The collaboration was announced on October 14 on the sidelines of the Singapore International Cyber Week, by Josephine Teo, Singapore's Minister for Digital Development and Information. Teo, who also carries the cybersecurity and Smart Nation portfolios, framed the project as the defensive half of an AI arms race the scammers are currently winning. Speaking at the Sands Expo and Convention Centre, she said the fight against scams has become harder to win because of artificial intelligence.

Her argument for the new model was simple. "If our devices can recognise signs of manipulation early, we can help potential victims avoid financial or credential loss, while keeping their personal messages on their devices," Teo said, according to Channel News Asia. And she rejected the idea of retreating from the technology: "Our response should not be to turn away from AI because scammers misuse it."

The project is part of Google's National AI Partnership with the Singapore Government. The technical case was laid out the same day in a Google DeepMind blog post by JJ Geewax, the lab's Director of Applied AI, and Miguel de Andrés, a Group Product Manager.

Here is the part that matters. Traditional scam detection works by scanning messages on a server somewhere in the cloud. That breaks the moment a conversation is end-to-end encrypted — which is to say, most conversations that matter. WhatsApp, Telegram and their peers cannot hand over chats they themselves cannot read. So DeepMind and GovTech moved the analysis onto the handset itself. The model runs entirely on the device, so the encryption stays intact and nobody's inbox gets vacuumed up to a data centre.

Photo: Singapore's Marina Bay skyline at night. The city-state has lost more than S$4 billion to scams since 2019 — and is now betting on on-device AI to fight back. Credit: Bijay Chaurasia / Wikimedia Commons, CC BY-SA 4.0.

How the two-stage pipeline works

The engineering problem is brutal. A model clever enough to tell grooming from gossip is usually too heavy, too hot and too thirsty for a phone battery. The team's answer is a two-stage pipeline.

Stage one is a lightweight, efficient classifier that screens inbound messages using minimal power. Its job is triage: catch anything that smells like a threat, and let the rest pass untouched. Stage two only wakes up when stage one flags something. Then the fine-tuned Gemma model — Google's lightweight open model family — examines a larger snapshot of the conversation and decides whether it is a genuine threat or a false alarm.

What makes it Singaporean is the training. Generic models, the DeepMind authors write, often struggle with local deception techniques. So the Gemma model was fine-tuned on local scam conversations: impersonations of public institutions, banks and local figures, plus trending regional schemes. The training data came from anonymised datasets drawn from scam reports and intelligence feeds, and the model was stress-tested against synthetic dialogues built to reflect local speech patterns. No private or personally identifiable information went into training — a rule the team calls "No Citizen Data" — and the fine-tuning happened inside GovTech's own managed Google Cloud environment, a "Sovereign Sandbox" that keeps training data isolated.

When the model does smell a scam, it triggers a real-time warning directly on the user's screen, urging the person to pause before sharing sensitive information or sending funds. The warning is the whole product. Scams are defeated in the ten seconds before a transfer, not in the post-mortem afterwards.

Why Singapore, why now

The numbers explain the urgency. Scams make up over half of Singapore's reported crimes, according to the DeepMind post, and have accounted for more than S$4 billion in losses since 2019. The tactics have shifted from crude phishing to slow-burn social engineering — attackers who groom victims with convincing conversations before ever asking for money. DeepMind says such scams now account for up to 80% of all losses.

The bleeding has not stopped this year. The Straits Times reports that S$410.6 million was lost to scammers in the first six months of 2026 alone, with e-commerce, phishing, investment, job and government-official impersonation scams the most common. The scam has become Singapore's most successful criminal industry, and it is industrialising fast.

Teo walked her audience through the modern scam's journey, and it is worth repeating because it explains why the old defences keep missing. A victim first meets the scammer on social media, drifts to a messaging app, takes a phone call, and finally transfers money through a bank. Four systems, four chances to intervene — and no single defender sees the whole chain. "It is therefore not enough to make every organisation better at stopping scams," she said. "We must also connect our defences so that a single signal detected on one part of the ecosystem can trigger action elsewhere."

Singapore is not starting from zero. The Scam Analytics and Tactical Intervention System (SATIS), run by GovTech and the Singapore Police Force, uses AI to scan and assess more than 600,000 websites every day, each in seconds, and disrupts around 20,000 malicious websites a month, according to Channel News Asia. The new on-device model is the missing piece at the end of that chain: the one that watches the conversation itself, where the victim actually gets talked into it.

Photo: A man checks messages on his phone in a café. The new model analyses exactly these kinds of private exchanges — entirely on the device. Credit: kargaltsev / Wikimedia Commons, CC BY 2.0.

The privacy bargain — and its limits

The privacy pitch is the project's strongest card, and its authors play it hard. When deployed, the model will only be able to access on-device conversations with the user's consent, under the terms of the app it is built into. And because it runs entirely on-device, DeepMind says, neither the Singapore Government nor Google will be able to access users' conversations through its deployment.

Mark Chen, GovTech's Director of Trust and Safety, put it bluntly: "Fighting scams shouldn't mean sacrificing personal privacy. Working with Google DeepMind, we're bringing advanced AI protection directly onto citizens' devices."

That is a genuinely elegant bargain — if it holds. Consent is doing a lot of work in that sentence. A model that can read your chats "with your consent" is one toggle, one pre-ticked box, or one unread terms-of-service update away from reading your chats without you ever really agreeing. The history of consent screens is not encouraging. And "neither government nor Google can access your conversations" is a claim about this deployment, not a law of physics: the same architecture, shipped by a different vendor under a different government, could answer a very different question.

There is a second, subtler risk. A detector that lives on your phone is also a censor that lives on your phone. Today's target is scammers impersonating bank officers. The mechanism — on-device scanning of private conversation, with warnings that interrupt you mid-chat — is content-neutral. Singapore's governance here looks careful: sovereign sandbox, anonymised training data, stress-testing. But the template DeepMind wants to export will land in countries with thinner safeguards and stronger appetites for surveillance. The question is never just "does this tool work?" It is "who gets to retrain it, and for what?"

What we still don't know

For all the polish of the announcement, the two numbers that would let anyone judge this system are missing. No release date has been announced. It is not even clear whether the model will be a default feature of the Android operating system or something users install. And no performance figures — no detection rate, no false-positive rate — have been published. We checked the DeepMind blog post, the Channel News Asia and Straits Times reports, and Crypto Briefing's write-up against each other, and all four agree on the shape of the system and the absence of a scoreboard.

That absence is the whole ballgame. A scam detector that cries wolf too often will be muted within a week; people learn to tap past warnings the way they tap past cookie banners. A detector that misses too much is decoration. And a detector trained on today's scripts will rot fast, because scammers adapt in weeks, not years. DeepMind says the model will undergo continuous training as tactics evolve. It will need to: the half-life of a scam playbook is brutally short.

Distribution is the other open question. A model only helps if it is inside the apps people actually use, and that requires the cooperation of messaging platforms and device makers. DeepMind is pitching the project as a template for exactly that — lightweight models that platforms and operating-system vendors can integrate without breaking encryption. Whether Meta, Apple or Telegram pick it up is a story for next year, not this one.

Here is our position, plainly: Singapore has built the most honest version of this idea anyone has yet attempted, and the honesty is the point. On-device processing, consent-gated access, anonymised training data, a sovereign sandbox — that is what a scam detector looks like when its designers actually take privacy seriously instead of treating it as a slide in the deck. But an honest design is not a proven one. Until DeepMind and GovTech publish detection rates, false-positive rates and independent audit results, this remains a promising prototype with a press release, not a proven shield. Governments should not get to grade their own homework on a tool that reads citizens' chats. Publish the scoreboard.

Why this matters beyond Singapore

Do not mistake this for a local story. Every country with a smartphone has the same wound: encrypted chats that protect dissidents and grandparents alike, and scammers who exploit the fact that nobody can look inside them. If Singapore's model works — if a phone can spot a con artist without becoming a listening device — every government on earth will want one, and every platform will face the question of why it has not shipped one.

The timing is pointed, too. A day earlier, Singapore demanded real-time AI incident transparency from the AI giants — a reminder that this city-state is not content to be a testbed; it wants to set the rules. Read our earlier report on Singapore's demand for real-time AI incident transparency. The same week has already supplied the counter-examples for why that pressure exists, including OpenAI's agents escaping their testing environment. Singapore is playing both sides of the AI ledger at once: demanding accountability from the labs, and building with them.

Google, for its part, also said on October 14 that it would give the Singapore Government access to its most advanced AI cybersecurity tools to help defenders find, verify and fix vulnerabilities. The scam detector is the citizen-facing half of a much bigger bet: that AI's best defence is more AI, pointed in the right direction.

Whether that bet pays off will be decided in the dullest possible place: the false-positive rate, the install base, the update cadence. Watch those, not the launch video. If the scoreboard ever appears, we will be first to read it.

Sources: Google DeepMind blog post by JJ Geewax and Miguel de Andrés; Channel News Asia; The Straits Times; Crypto Briefing; Unite.AI.

TechPulse USA is an independent publication covering global tech and AI news for readers in the USA, Europe, Canada and Australia. We verify every fact against primary and reputable sources before publication.

DeepMind's scam detector: your questions answered

What exactly did DeepMind and GovTech announce? An on-device AI model that reads your chats as they happen, spots manipulation, and warns you before you send money — without a single message ever leaving your phone. It was announced on October 14 at Singapore International Cyber Week by Minister Josephine Teo.

How does the two-stage pipeline work? Stage one is a lightweight classifier that screens inbound messages using minimal power. Only when it flags something does stage two wake up: a fine-tuned Gemma model examines a larger snapshot of the conversation and decides whether it is a genuine threat or a false alarm.

Will it work on WhatsApp and Telegram? That is the point of putting the model on the device: end-to-end encrypted chats cannot be scanned on a server, so the analysis has to happen on the handset. But the model only helps if it is inside the apps people use, which requires the cooperation of messaging platforms and device makers. No integrations have been announced.

When can I get it, and will it be on Android by default? No release date has been announced, and it is not clear whether the model will be a default Android feature or something users install. The Straits Times reports the model targets Android users; whether it becomes an Android default is still unannounced.

How will anyone know it works? By the scoreboard that does not exist yet: no detection rate and no false-positive rate have been published. Britain is already stress-testing Google's Gemini models for safety — the same independent scrutiny this detector will need before governments can grade their own homework. And it lands in the same week as Google's flagship Gemini 4 launch, a reminder of how fast the lab's models are moving from research to products.