Google DeepMind and the Singapore government unveiled a new approach to scam detection on 14 October 2026: instead of sending the contents of someone’s private conversation to a server, a small AI model can look for warning signs directly on the person’s phone. The idea behind these AI scam warnings is strikingly simple. A conversation that sounds reassuring can gradually become an attempt to obtain money, bank details or credentials. The software is designed to recognise that change and encourage the recipient to pause before responding. Google DeepMind's announcement describes a collaboration with Singapore’s Government Technology Agency, GovTech, rather than a feature already installed on every handset.
This matters because many scams no longer start with an obviously suspicious link. An impostor may patiently build trust, claim to be from a bank or government service, and only later introduce the request that causes harm. A warning at the point of persuasion could be more useful than a filter that only examines individual messages for familiar fraud phrases. It is also a test of whether useful fraud detection can be built without turning private communications into a permanent source of surveillance.
AI scam warnings look for the story behind a message
A traditional suspicious-message filter may scan for bad links, strange senders or known scam wording. This project is intended to assess the developing conversation instead. Google says its researchers adapted a lightweight Gemma model using examples of local scam conversations, including people impersonating public institutions and banks. The system looks for manipulative intent, not simply a prohibited word. That distinction matters because a perfectly ordinary sentence can be part of a dishonest interaction when it follows a carefully manufactured story.
The design uses stages to reduce the processing burden on a phone. Not every message needs the same level of analysis. More intensive checks can be directed towards conversations that look suspicious. This is important for a tool that might run throughout the day: a security feature which drains the battery, freezes a messaging app or repeatedly interrupts harmless chats risks being switched off. The announced design is an engineering approach, however, not published evidence that every phone will experience negligible battery use.
It has been tailored to Singapore’s scam patterns. That specificity is a strength for the intended pilot setting and a caution for anyone assuming the same model will immediately recognise every fraud in Britain or another country. Criminals adapt their language, preferred platforms and impersonation techniques to local circumstances. The company says it tested the model with synthetic dialogues and local materials, but has not established that this system has already prevented a measured number of losses in widespread public use.
Why keeping the conversation on the phone changes the argument
People want protection from criminals, but they also want the freedom to speak privately. End-to-end encryption prevents a messaging provider from casually reading what participants say. Sending those conversations to an external fraud-detection service would raise a new set of privacy questions, particularly when messages concern relationships, money or personal circumstances. Google says the proposed system processes the material on the device, so its protective analysis does not require the conversation to be uploaded to a cloud service.
There are practical limits to that promise. On-device processing describes where analysis takes place; it does not by itself tell users how long information is retained, how warnings appear in each app or exactly which protections an eventual commercial integration would offer. Those details depend on developers who incorporate the model. The sensible consumer question is not just whether a feature claims to use AI, but what its privacy documentation says it accesses, records and shares.
Singapore’s authorities are treating fraud as a public-safety problem rather than merely an inconvenience for technology firms. At Singapore International Cyber Week, minister Josephine Teo confirmed that GovTech was collaborating with Google DeepMind on real-time scam detection on personal devices. She also described broader national anti-scam work. The new model should be understood as one component of that effort, not as proof that existing fraud-prevention systems have suddenly become obsolete.
A warning is an invitation to verify, not a verdict
A model examining conversations can make mistakes. It may miss sophisticated manipulation, especially if fraudsters change tactics, or flag an unusual but legitimate request. A warning is useful when it gives someone time to verify a claim independently; it is harmful if the product presents a machine-generated suspicion as certain evidence that another person is a criminal. The announced material does not establish a reliable global false-positive or false-negative rate for consumer deployment.
For an individual, the old protections therefore remain relevant. A supposed bank employee who requests an urgent transfer should be checked using a contact number obtained independently, not one supplied within the conversation. Unexpected requests for verification codes deserve particular caution. These habits complement a possible on-device warning and would still matter when the software fails to recognise an elaborate story.
LiveAIWire has examined how voice-cloning scams can exploit familiar voices and how fraudulent dating profiles build false trust. Both illustrate why the decisive signal in a scam may be the developing human relationship, rather than a single suspicious word. A locally aware detector offers one possible response, although the risk of misplaced trust in automated warnings should not be ignored.
The promise of this project is not that a phone will become an infallible lie detector. It is that a private message might be analysed privately, with a useful interruption at the moment a manipulative request arrives. Whether the promise holds will depend on transparent testing, restrained warning design and what real users encounter once any service reaches their devices.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.