Google DeepMind Builds Two-Step Model Against Scams in Singapore's Messaging Apps
Google DeepMind and GovTech Singapore have announced a scam-detection model that runs entirely on the user's smartphone, with no message data sent to the cloud. The system is the first technical deliverable from the country's national AI partnership with Google DeepMind — but it still lacks a launch date and any performance figures.
Google DeepMind and Singapore's government technology agency, GovTech Singapore, have announced a privacy-friendly scam-detection model based on Google's open Gemma model. The system was presented on 14 October 2026 at Singapore International Cyber Week by Minister Josephine Teo, and is described as the first technical deliverable from the national AI partnership between the two parties, announced in May 2026 (MSN).
What sets the solution apart from existing scam protection is that the entire analysis takes place on the device itself: the system analyses suspicious messages in real time without transferring conversation data to any server.
The background: a substantial scam problem
The announcement comes against a backdrop of significant figures: Singapore lost 410.6 million Singapore dollars (around 321 million US dollars) to scams in the first half of 2026 — a 17.9 percent year-on-year decline that, according to the reporting, reflects intensified countermeasures by the authorities but remains a large sum. Scams account for more than half of recorded crime in the country, and cumulative losses since 2019 exceed S$4 billion (MSN). The figures come from secondary reporting of Singaporean authorities' statistics, not from the underlying report itself.
How the system works: two steps
Technically, the solution is built on a two-step pipeline. The first step is a low-energy classifier that screens every incoming message with minimal power consumption. Only when this flags a potentially suspicious message is the second step activated: a fine-tuned Gemma model ingests a larger excerpt of the entire conversation, including earlier message turns, to perform an analysis in context across multiple messages (MSN; Cryptonomist).
The second step assesses the context either to confirm a threat or to dismiss a false alarm. The architecture means the heavy model does not continuously drain the battery — it only switches on when the lightweight classifier has first reacted.
The focus on multi-turn conversations is significant: many scam attempts, particularly so-called social engineering, unfold gradually across several messages, and are harder to expose in a single message than in context.
Training and privacy
The model is fine-tuned in a GovTech-managed Google Cloud environment, where the training data is kept segregated. Google says that private and personally identifiable information about citizens was kept out of the training, and that anonymised datasets drawn from scam reports and intelligence sources were used instead (Marketing-Interactive).
As for access after deployment, Google states that rollout will happen gradually and be developed together with the Singaporean government, and that the model will only be able to access conversations on the device with the user's consent and in accordance with the terms of the app it is integrated into. According to Google, this means neither Google nor the government gains access to users' conversations — but these are the company's own statements, not independently verified facts.
"Fighting scams shouldn't mean sacrificing privacy. The collaboration with Google DeepMind brings advanced AI protection directly onto citizens' devices. This lets us catch highly targeted, localised scam attempts in real time while personal communications remain private," Mark Chen, director of trust and safety at GovTech Singapore, told Marketing-Interactive. JJ Geewax of Google DeepMind also made statements positioning the project as real-time on-device scam detection with privacy preserved.
What remains unclear
Despite the concrete technical description, major things are missing. Minister Josephine Teo gave no launch date for the Gemma-based tool at the announcement, and did not share performance figures either — neither detection rates nor false-positive rates (Crypto Briefing). This means that, as of today, there is no public way to assess how effective the system actually is.
It is also unresolved how distribution will happen in practice: the rollout depends on messaging platforms integrating the model, and the details of that have not been settled. In other words, the tool has been announced, but not deployed or made available to users.
Why it matters
At the same time, the partnership points to something bigger than a single tool. The on-device approach addresses a well-known dilemma in scam prevention: effective detection often requires access to conversation content, while that access collides with privacy. By analysing messages where they already are — on the user's own phone — the need to send data off the device is removed, while the model can still be trained on locally rooted scam patterns, described as "highly targeted, localised scam attempts".
If the model proves to work without an unreasonably high rate of false alarms, the approach could become a template for similar initiatives in other countries with large scam problems. But that is a conditional conclusion: without published performance figures and a concrete rollout plan, it remains a wait-and-see assessment based on what Google and GovTech themselves have made public.