Reads Your Chats, Promises Nothing Leaves the Phone

Google DeepMind and Singapore's government technology agency, GovTech, announced a phone-based AI that scans private chats for scam tactics, and say no chat data is sent to the cloud. ^[1]

The announcement came on October 14, 2026, as part of what DeepMind calls its National AI Partnership with Singapore. The model is meant to run on typical smartphones in the city-state and to process conversations on the device. ^[1]

That last point is the one that matters. A tool that reads your messages to protect you is only as welcome as its promise that nobody else reads them. The promise comes from the announcement. The material reviewed contains no independent check of it, and no published figures for how often the detector catches scams or raises false alarms. ^[1][2]

Two stages, one small model

The design is built around economy. A lightweight classifier, a simple filter that sorts messages into likely harmless and worth a second look, first screens incoming messages while using little power. Most chats, presumably, never get past it. ^[1]

When something looks suspicious, a second stage takes over. This is a fine-tuned version of Gemma, meaning a model retrained on extra examples so it learns a specific job. Gemma is Google's family of models that developers can use to build their own AI applications. The retrained model examines a larger snapshot of the conversation, and its job is to confirm a likely scam or throw out a false alarm. ^[1]

Think of a smoke alarm that, instead of shrieking at every burnt slice of toast, wakes a more careful inspector before anyone is told to leave the building. The first stage is cheap and always on. The second is slower and looks at context.

Context is the point. A filter that judges one message at a time sees only a single line, and a scammer's manipulation can unfold across an exchange. So the second stage looks at more of the conversation before deciding. ^[1]

If it sees manipulative patterns, the announcement says, the user gets a real-time warning on screen, advising them to pause before sharing sensitive information or money. ^[1]

Why Singapore

DeepMind puts the scale of the problem in stark terms. It says scams account for over half of Singapore's reported crimes and more than S$4 billion in losses since 2019. It adds that social-engineering scams, where a fraudster manipulates someone through a convincing interaction, account for up to 80% of losses.

These are the company's figures. ^[1]

The model was fine-tuned on local material, which is what a generic scam filter would lack. DeepMind says the examples include impersonations of public institutions, banks and local figures, as well as regional schemes. ^[1]

DeepMind says the fine-tuning took place in a GovTech-managed Google Cloud environment. It says training excluded private or personally identifiable information and drew on anonymised datasets derived from scam reports and intelligence feeds. Testing, it says, used synthetic dialogues and local collections of scam examples. ^[1]

The privacy promise

GovTech's Director of Trust and Safety, Mark Chen, framed the project as a refusal of a trade-off. "Fighting scams shouldn't mean sacrificing personal privacy," he said, adding that the aim is to put protection on citizens' devices while keeping personal communications private. ^[1]

The technical idea behind the pitch is sound in outline. Many messaging apps use end-to-end encryption, which means only the people in a conversation can read it. That limits what the service can scan on its own servers. Analysis on the phone, after a message has been decrypted for the user anyway, offers a way to warn people without asking the service to read their chats.

But a design that keeps data on the device is a promise about software, and software changes. Whether the claim holds depends on how the tool is built, how it is wired into apps and operating systems, and what it does when it updates or reports back on its own performance. Training in the cloud, as DeepMind describes it, is a separate matter from running on the phone, and the announcement's privacy claim is about the second.

DeepMind says messaging platforms and operating-system vendors could integrate the lightweight models locally. It does not say any of them has. ^[1]

What has not been shown

The announcement gives no public launch date. It does not name the messaging apps or phone models that would support the tool, and it publishes no detection or false-alarm rates. A secondary report on the announcement says the same, and that no performance figures were provided. Neither source includes an independent test of the privacy claim or the detector's accuracy. ^[1][2]

The false-alarm rate deserves as much attention as the catch rate. A scam detector that cries wolf teaches people to swipe its warnings away, and one that sits on every private conversation will be judged on how often it interrupts honest ones.

What to watch

The practical questions are plain. When will the tool reach users, and on which apps and phones? Will independent evaluators be allowed to test both how well it detects scams and whether message content stays on the device, including during updates and diagnostic reporting? Can users see what the model inspects, how long that information is kept, and switch the feature off? And who oversees updates as scammers change their scripts?

Until those are answered in public, the strongest thing Singapore and DeepMind have offered is a sentence. It is a good sentence, and it is theirs.