🏠 Everyday life AI

Google DeepMind's New On-Device AI Is Built to Catch Chat Scams

A partnership with Singapore's GovTech uses lightweight Gemma models to spot scam patterns in private messages — without sending your data anywhere.

4 min read · October 14, 2026

This article was written by AI. It passed automated fact and quality checks, and our editor spot-checks published articles every day.

See AI World HQ first on GoogleAdd us as a preferred source and our stories show up higher in Google’s Top Stories.Add on Google

Picture this: you're scrolling through WhatsApp or Telegram, and a message pops up. It looks like your bank, a delivery service, or even your kid asking for help. Your heart rate jumps. Is it real? Most scam protection today catches bad links or known scam numbers. But what about the long, friendly chat where a scammer slowly builds your trust before asking for money? That's exactly the kind of scam a new tool from Google DeepMind and GovTech Singapore is built to catch.

What's actually being announced

Google DeepMind and GovTech Singapore (the agency that builds government tech for Singapore) are working together on an AI tool that sits on your phone and watches private chat conversations in real time. When a chat starts looking like the opening moves of a scam — fake delivery notices, "hi mom I changed my number" tricks, romance setups — the tool flags it before you send money or click a link.

The headline technical choice: it runs on-device, meaning everything happens on your phone. Your messages never leave to be checked somewhere else.

How "on-device" actually works

An on-device AI is a model that lives on your phone or computer and does its thinking right there, instead of calling out to a company's servers (the big remote computers that power most AI tools you use today).

The model powering this tool is from the Gemma family — Google's family of compact AI models designed to run on everyday hardware. The version being used is described as lightweight, which means it doesn't need a supercomputer or a fast internet connection to work. It runs on a regular phone, which matters because most scam messages hit phones first.

Why "zero cloud data egress" is the whole point

Cloud data egress is a fancy way of saying "your data leaving your device and traveling to someone else's computer." When you use most AI tools, your message goes from your phone, across the internet, into a company's servers, gets analyzed there, and a reply comes back. Along the way, a copy of your message lives on that server.

"Zero cloud data egress" means: the model does all the analysis on your phone. Nothing gets uploaded. Your messages stay on your device, full stop.

For a tool that reads your private conversations, that is not a small detail. A scam detector that read every message you sent and stored a copy on a company server would create a privacy problem worse than the scams themselves.

What this means for you

If you get a lot of chat messages: this on-device approach is the direction the industry is moving. The kind of protection described here — local AI watching for scam patterns without uploading your messages — is what you'll likely see built into phones and chat apps over the next few years.
If a scam has ever caught you off guard: you're not alone. Conversational scams (the long, chat-based kind) are one of the fastest-growing fraud types, partly because traditional spam filters miss them. Awareness is still your first line of defense — slow down when anyone asks for money, codes, or login info, even if they "sound" like someone you know.
If you're privacy-conscious: on-device is the standard to look for. When any tool offers to "protect" you by reading your messages, ask whether the analysis happens on your device or on a company's server. On-device is the better answer.
If you're just curious about AI: this is a great example of AI doing a narrow, useful job. It doesn't need to be the smartest model in the world — it just needs to recognize "this conversation looks like a scam opening" well enough to warn you.

Wrap-up

Google DeepMind and Singapore's GovTech are showing what scam protection can look like when privacy is built in from day one: a lightweight AI that lives on your phone and reads your messages without sending them anywhere. You can't download this tool today — it's a partnership project for Singapore — but the approach it demonstrates is worth paying attention to, because it's almost certainly coming to a chat app near you.

For now, the single safest habit you can build is also the oldest: pause before you act on any message that asks for money, codes, or login info, no matter how real it sounds.

#ai-news#scam-protection#gemma#on-device-ai#deepmind#privacy

Keep reading

📘 How to Spot an AI Scam Call, Fake Email, or Deepfake Voice Today Three AI-powered scams happening right now and the one-second checks that catch each one 📘 How to Spot AI-Generated Scam Messages Before You Click or Reply A practical five-minute checklist for checking suspicious texts, emails and DMs with your usual AI assistant 📘 How to check if a suspicious message is a scam by pasting it into an AI A 5-minute way to get a fast second opinion on a sketchy email or text before you click anything 📰 Small Language Models Explained: The Lighter AI on Your Phone What an SLM is, why companies and phones are using them, and what that means for you 📘 Getting Started with Mistral's Le Chat: Your Private AI Guide Discover how to confidently use Europe's leading private AI assistant for daily writing and organizing tasks, with privacy in mind. 📰 Local AI Is Getting Faster: What On-Device Models Mean for You A plain-language guide to running AI directly on your computer or phone, and why it suddenly matters in 2026

Was this helpful?

✦ Generated by AI in AI World HQ's automated newsroom, from official sources. Checked by automated fact and quality gates, and our editor spot-checks published articles every day. Spot a mistake? Use the buttons above.

← Back to all stories