Google DeepMind and Singapore's GovTech ship an on-device scam detection model: fine-tuned from Gemma, a two-stage pipeline, and neither the government nor Google can read users' conversations

Start with the scale of the problem. The post says scams such as investment fraud, fake job offers and impersonation of public officials make up over half of Singapore's reported crimes and account for more than S$4 billion in losses since 2019. It also notes that scam tactics have shifted from simple phishing to complex social engineering, where attackers manipulate victims with convincing conversations before requesting money, a category that now accounts for up to 80% of all losses. Google explains why generic models are not enough: they often struggle with local deception techniques, which makes detection harder. So the model is Google's lightweight Gemma fine-tuned on local scam conversations covering impersonations of public institutions, banks and local figures, plus trending regional schemes, which Google says ensures the fine-tuned model can distinguish benign social chatter from genuine manipulation when deployed in applications citizens install.

The technical route is specific: a two-stage pipeline. Stage one is a lightweight, efficient classifier that screens inbound messages using minimal power and focuses on catching potential threats. Stage two does in-context analysis with the fine-tuned Gemma model, reading a larger snapshot of the conversation to confirm a threat or discard a false alarm. Google notes that running a complex scam classifier on modern smartphones is no easy task and that battery efficiency is a critical concern, which is precisely why the pipeline is split in two. When the model detects manipulative conversational patterns, it triggers a real-time warning directly on the user's device advising them to pause and exercise caution before sharing sensitive information or funds. Unlike traditional threat detection, which relies on server-side scanning of messages in the cloud, this model processes conversation data on-device; Google says that once integrated into applications it allows faster analysis of conversation patterns to detect manipulative intent and helps citizens act before financial or credential exposure can occur.

The governance section deserves its own paragraph, because it breaks the privacy promise into three checkable commitments. The first is a Sovereign Sandbox: fine-tuning took place within GovTech Singapore's managed and maintained GCP environment, keeping training data isolated. The second is no citizen data: training strictly excluded private or personally identifiable information and used anonymised datasets derived from scam reports and intelligence feeds. The third is stress-testing: the model was tested against synthetic dialogues and local corpora to cover diverse scam categories and evolving impersonation tactics. Google also says the roll-out within applications will be carefully phased and developed in partnership with the Singapore Government, and that once deployed the model will only be able to access on-device conversations with the user's consent, in accordance with the terms of use of that application. The post is bylined JJ Geewax and Miguel de Andrés, was presented at Singapore International Cyber Week (SICW 2026), and forms part of Google's National AI Partnership with the Singapore Government. Mark Chen, Director of Trust and Safety at GovTech Singapore, says fighting scams should not mean sacrificing personal privacy, and that bringing advanced AI protection directly onto citizens' devices lets the team catch highly targeted, local scam attempts in real time while keeping personal communications private.

A few things readers should keep in mind. First, Google has not published accuracy, false-positive or recall figures for the model, nor the number of users involved in testing, so its real-world miss rate and nuisance rate cannot be judged from this post. Second, the post frames this as an interim result of the collaboration with Singapore and says the team is exploring the potential for similar technology to be deployed in other countries, which means there is no availability timeline for anywhere outside Singapore today. Third, the model is a capability offered to application integrators such as messaging platforms rather than a standalone product a user can download right now; whether it can truly land on end-to-end encrypted messaging platforms is described as a model for platforms balancing protection with encryption rather than a finished answer. Finally, every fact here comes from Google DeepMind's post and the sources below, with no speculation added.