Google DeepMind has partnered with Singapore’s Government Technology Agency (GovTech) to develop an artificial intelligence tool designed to detect scam conversations directly on smartphones. Announced on October 14, 2026, the initiative aims to identify suspicious messages in real time while keeping users’ private conversations on their devices.
The project addresses a growing challenge as scammers increasingly use convincing conversations, impersonation and social engineering to persuade victims to transfer money or disclose sensitive information.
How Google’s AI scam detection tool works
The system uses a lightweight model from Google’s Gemma family, fine-tuned on local scam conversations to recognize suspicious patterns. These include messages impersonating public institutions, banks and local figures, as well as emerging regional fraud schemes.
Rather than relying on a single model to examine every message, the technology follows a two-stage process. A lightweight classifier initially screens incoming messages for potential threats. If it identifies suspicious activity, a more capable fine-tuned Gemma model analyzes a broader portion of the conversation to determine whether it may be a scam.
When the system detects manipulative patterns, it is designed to display a warning on the user’s device, encouraging them to pause before sharing sensitive information or transferring money.
Privacy remains a key priority
One of the project’s main features is its on-device processing approach. Unlike conventional systems that may rely on server-side analysis, the model is designed to assess conversations locally without transmitting their contents to external cloud servers for detection.
This approach could help messaging platforms introduce additional scam protection while preserving end-to-end encryption. The model is intended to access conversations only with users’ consent and in accordance with the relevant applications’ terms.
Google DeepMind and GovTech also say that the model was fine-tuned in a controlled cloud environment managed by GovTech. Its training used anonymized datasets derived from scam reports and intelligence feeds, excluding private and personally identifiable information.
Why the technology matters
Scammers increasingly rely on social engineering rather than simply sending suspicious links. They may build trust over several messages, impersonate legitimate organizations or create urgency before asking victims to take action.
By examining conversational context, an AI model could help identify suspicious behavior that traditional filters based on known phone numbers or malicious links might miss. However, the effectiveness of the new system will depend on its ability to distinguish genuine scams from legitimate conversations and minimize false warnings.
The initiative also highlights the growing role of smaller AI models that can perform useful tasks directly on smartphones, reducing the need to send sensitive information to remote servers.
When will users get access?
The announcement describes a collaboration to develop and integrate the technology, rather than a broadly available consumer product. Google DeepMind and GovTech have not announced a confirmed public rollout date or named all participating messaging applications.
The project is initially tailored to Singapore’s scam landscape, although its approach could eventually inform similar protections in other markets.
As AI-driven fraud becomes more sophisticated, on-device detection could offer a useful additional layer of protection. Its real-world impact will depend on deployment, detection accuracy and how effectively the technology adapts to changing scam tactics.
Also read: Google launches Gemini 4 Argon AI model.
