Google DeepMind and Singapore Build On-Device Gemma Tool to Catch Scam Chats

Google DeepMind and GovTech Singapore have built an AI tool that watches for conversational scams inside private chats, running the detection entirely on the user's device rather than on remote servers. The system uses lightweight versions of Gemma, Google's open model family, to analyze message threads locally. Neither party disclosed a rollout date or how many users the tool will reach.

Scam messages have moved past bad grammar and fake prize notices. Modern fraud arrives as a plausible conversation — a fake recruiter, a romance approach, a "friend" asking for a transfer — that unfolds across days and only turns suspicious late in the exchange. Filtering those threads by keyword fails, and cloud-based scanning means handing private messages to a third party, which is a hard sell in most jurisdictions.

That is the problem the on-device design addresses. Running Gemma on the handset means no message leaves the phone, so there is no data egress to a server and no central store of citizens' conversations to breach or subpoena. The tradeoff is size: a model small enough to sit on a mid-range phone has less reasoning capacity than a cloud model, so the detection problem becomes one of catching the right signals with a fraction of the compute.

The partnership follows a pattern Google has pushed for the past two years, putting Gemini Nano on Pixel devices for tasks like call screening and text summarization. Singapore has run parallel experiments through GovTech, its public-sector technology arm, including scam-detection work and the AI Verify testing framework. Pairing a government that wants a deployable public-safety tool with a model family it can inspect and run locally is a natural fit for both sides.

For the AI industry, the interesting part is the deployment shape rather than the model itself. Gemma is open-weight, which lets a government audit and adapt it without a commercial license, and the on-device constraint sidesteps the data-residency objections that have slowed similar public-sector AI projects in Europe. If the approach holds up, it becomes a template for other regulators who want AI enforcement without building a surveillance pipeline.

What to watch is whether the tool actually catches scams at a useful rate. Conversational fraud is adversarial — scammers adapt to whatever filter stands in front of them — and no one has published false-positive numbers for an on-device detector of this kind. Google DeepMind and GovTech have not said when results will be shared, or whether the model weights and evaluation data will be released for outside testing.