Google DeepMind and Singapore build AI scam detection that can work on a device
On this page
Announcement date: October 14, 2026
Google DeepMind and Singapore’s Government Technology Agency announced work on an AI model for scam detection on October 14, 2026. The project focuses on a practical problem: scams often depend on a convincing conversation, not just an obviously suspicious link.
Google says the team adapted its lightweight Gemma model using local scam conversations. The aim is to recognise patterns such as impersonation of banks, public institutions and local figures, while allowing the model to run on a device.
What does on-device AI mean?
Many AI tools send a request to a remote server, which processes it and returns an answer. On-device AI performs the relevant model processing locally, on hardware such as a phone. That can change how much information needs to leave the device.
It is still sensible to read the privacy information for the finished app. A locally running model is one part of a system. Other features may have their own data handling, permissions or reporting options.
Why a local scam needs local examples
A scammer may imitate a familiar service, use a local expression or mention a process that sounds normal in one country. A general model may not recognise why that combination should raise a concern.
Fine-tuning means taking an existing model and adapting it with more focused examples. In this project, Google says those examples include locally relevant deception techniques. The announcement describes a model and collaboration, rather than proving that every person can already switch it on in every phone.
How a warning might help
Consider a fictional message from someone claiming to be a manager. They ask you to move the conversation to a private channel and urgently send money for an unexpected expense. No single phrase proves a scam. The combination of pressure, impersonation and an unusual request is what deserves attention.
An AI warning could give a person time to pause. That pause is useful only if the person has a clear next step. For example, they could contact the supposed sender through a number or channel they already trust, rather than replying using details in the suspicious message.
A warning is not a verdict
Detection systems can miss a scam or flag an innocent message. A false positive is a warning about something harmless. A false negative is a failure to warn about something harmful. Both matter when evaluating a protective tool.
For a beginner, the practical lesson is to avoid treating silence from a detector as proof that a message is safe. Equally, a warning should lead to checking the situation, rather than assuming an accusation is established.
When comparing products, look for an explanation of what was tested, which languages and situations were covered, and how the app handles a warning. A broad claim that a tool uses AI tells you very little on its own.
QueryFun’s take
The promising idea here is focused AI that understands a particular setting and can process information locally. The important questions will be how the finished applications perform and how clearly they help people respond. For now, this is a development to watch, with the limits of the announcement kept in view.
Our AI privacy guide explains more about checking what a tool can access.
Original source: Google DeepMind’s announcement with GovTech Singapore. Company statements are attributed to their source. Practical examples and commentary are QueryFun’s explanation, not an independent product test.