Google DeepMind and Singapore’s GovTech introduce scam-detection AI that runs on phones

Google DeepMind and Singapore’s Government Technology Agency, GovTech, have introduced a scam-detection model designed to examine conversations directly on smartphones. The October 14 announcement describes AI that warns people about suspected manipulation before they share money or sensitive information. Its privacy approach is to keep conversation analysis on the device; public deployment has not been announced.

Screen the message, then examine the conversation

Presented at Singapore International Cyber Week 2026, the project tackles a limit of cloud-based threat scanning. End-to-end encryption stops many messaging platforms from inspecting private conversations, DeepMind says. This project checks them on the phone instead, with the goal of spotting manipulation while keeping conversation data on the user’s device.

That creates a different constraint: the detector must work within a smartphone’s power budget. DeepMind says the model is optimized for typical smartphones in Singapore and identifies battery efficiency as a critical concern. Rather than give every incoming message the same larger-model analysis, the system divides the work into two stages.

First, a lightweight classifier—a model that screens messages for potential threats—checks incoming messages using minimal power, according to DeepMind. A fine-tuned Gemma model then evaluates a larger snapshot of the conversation. This second stage is intended to confirm a threat or dismiss a false alarm, bringing surrounding conversational context into the decision.

When the model detects manipulative patterns, it triggers a real-time warning on the device. The warning advises the user to pause and exercise caution before sharing sensitive information or funds. The described intervention is a warning to the person, giving them a chance to reconsider before financial or credential exposure.

Singapore’s scam conversations shape the model

DeepMind says generic models can miss local scams. It adapted its lightweight Gemma model with local scam conversations, including impersonations of public institutions, banks and local figures, and trending regional schemes. The goal is to tell ordinary social chatter from real manipulation, not flag messages just because they are unfamiliar.

DeepMind describes a shift from simple phishing toward complex social engineering: attackers use convincing conversations to manipulate victims before requesting money. It says these scams now account for up to 80% of scam losses. The larger conversation snapshot is therefore a deliberate part of the detector’s design, rather than a check confined to one isolated incoming message.

The announcement puts the scale of Singapore’s problem at more than S$4 billion in losses since 2019. Scams also make up over half of the country’s reported crimes, according to DeepMind. Those figures describe the problem the collaboration targets; they are not measures of money saved or scams prevented by this model.

Cloud training, local analysis

On-device operation does not mean the model was trained on citizens’ phones. DeepMind’s training safeguards describe a separate development environment and restrictions on the information used to adapt Gemma:

• Fine-tuning took place inside a GovTech-managed Google Cloud environment, which kept the training data isolated.
• Training excluded private or personally identifiable information. It used anonymized datasets derived from scam reports and intelligence feeds.
• Testing used synthetic dialogues and local collections of text to cover different scam categories and evolving impersonation tactics.

DeepMind says messaging platforms and operating-system vendors can integrate the lightweight models directly. That positions the collaboration as technology for those services to incorporate, rather than an announced consumer app. The announcement gives no public rollout date or measured detection results, leaving its effectiveness in everyday conversations an open question.