Google DeepMind and Singapore's GovTech launch fraud detection that runs directly on the phone

Google DeepMind and Singapore's GovTech launch fraud detection that runs directly on the phone

Google DeepMind and Singapore's GovTech have announced an AI model that catches manipulative scam conversations in real time – entirely on the device. Alongside the announcement comes a claim of a template for how protection can be reconciled with end-to-end encryption, and a series of open questions.

The announcement

Google DeepMind and the Government Technology Agency of Singapore (GovTech) have developed a scam detection model that runs directly on the smartphone, according to a blog post by JJ Geewax and Miguel de Andrés at Google DeepMind, published on October 14, 2026. The collaboration is part of what is described as the National AI Partnership with Singapore's government, and the model is, according to the company, optimised to run on typical Singaporean smartphones. The presentation took place at Singapore International Cyber Week (SICW 2026).

The core of the tool is that analysis happens locally: messages are to be screened for manipulative conversational patterns without the content being sent to a cloud service. When the model detects manipulative patterns, it triggers a real-time warning on the user's device advising the user to pause and exercise caution before sharing sensitive information or money.

Why Singapore

The rationale for the effort is an extensive scam problem, as presented in the blog post: scams are said to account for more than half of reported crimes in Singapore and to have caused more than S$4 billion in losses since 2019. At the same time, the tactics are said to have shifted from simple phishing to complex social manipulation, in which attackers build trust through convincing conversations before asking for money. Such scam calls are now said to account for up to 80 per cent of all losses.

It is worth noting that these figures come solely from Google DeepMind's own blog post. Underlying official crime statistics are not included in the material underlying this article, and there is no independent secondary coverage of the announcement. The figures should therefore be read as the company's and the agency's own framing of the problem, not as independently verified facts.

How the two-stage pipeline works

Technically, the blog post describes a two-stage architecture:

1. First stage: A lightweight and efficient classification screen for incoming messages. This is to use minimal power and is built to catch potential threats without draining the battery – an important consideration for a model intended to run continuously on an ordinary phone.

2. Second stage: Contextual evaluation with DeepMind's fine-tuned Gemma model, which analyses a larger excerpt of the conversation to confirm a threat or discard a false alarm. The point of two stages is to save resources: only messages the classifier flags are passed on to the heavier language model.

The fine-tuning is said to have been done on local scam conversations – from impersonations of public institutions, banks and local figures, to regional scam methods that are spreading. According to the company, this gives the model the ability to distinguish harmless social chit-chat from genuine manipulation, which is a crucial distinction for reducing the number of false warnings in everyday conversations.

Privacy and governance – as claims

The entire vendor account rests on the processing taking place on the device, but the company also explains how the development itself was reportedly governed:

• "Sovereign Sandbox": The fine-tuning is said to have taken place in GovTech's self-managed Google Cloud Platform environment, with isolated training data.
• No citizen data: The training is said to have strictly excluded private or personally identifiable information about citizens, using anonymised datasets derived from scam messages and intelligence sources.
• Stress testing: The model is said to have been tested against synthetic dialogues and local text corpora.

The same caveat applies here: these are the company's own accounts, and no independent audit has been presented of either the data basis or the governance mechanisms. Readers assessing the privacy claims should therefore treat them as promises from the developers, not as documented properties.

A template for messaging platforms?

The strategic point of the announcement is broader than Singapore. The blog post presents the collaboration as a model for messaging platforms that must balance user protection against end-to-end encryption – a long-running debate in which cloud-based scanning of message content has met strong resistance from encryption experts and privacy organisations.

Since the model runs locally, the argument goes, the content never leaves the phone, and no third party – neither platform nor government – gains access to the conversation. Mark Chen, Director for Trust and Safety at GovTech Singapore, is quoted in the blog post: "Fighting scams shouldn't mean sacrificing privacy. Together with Google DeepMind, we're bringing advanced AI protection right onto citizens' devices. This lets us catch highly targeted, localised scam attempts in real time, while keeping personal communications completely private."

The quote is translated from English; the original is available in the blog post. It is GovTech itself, through Chen, that advances the claim that communications are kept "completely private" – a formulation that depends on the entire pipeline actually running on the device in production, something that cannot be verified from the available information.

Open questions

Several of the most practical details are missing from the announcement:

• Deployment timeline: The blog post does not say when or in what form the model will become available to users.
• Platform integration: No messaging platforms or operating systems are named as partners, and it is unclear whether the tool is to be integrated into existing apps or delivered in some other way.
• Measured accuracy: No figures are given for hit rates or false-positive frequency – key metrics for any system that will warn users about their own conversations.
• User control: It is unclear whether users can opt the feature out or in, and how such settings would work.
• Division of roles: The blog post refers to GovTech as the partner, while the post's URL refers to MDDI (Ministry of Digital Development and Information). The exact division of responsibility between the agency and the ministry is not specified in the material.

All in all, the announcement currently rests on the parties' own account. Until there is independent documentation of accuracy, actual deployment and privacy practices, the most reasonable reading is of the announcement as what it is: a technical direction – a small model on the device, fine-tuned locally, in response to a major societal problem – presented by the actors behind it.