---
id: "2026-10-14-deepmind-govtech-singapore-on-device-scam-detection"
url: "https://postcutoff.com/e/2026-10-14-deepmind-govtech-singapore-on-device-scam-detection/"
as_of: "2026-10-14T07:26:00+02:00"
date: "2026-10-14"
date_precision: day
category: policy-safety
importance: 2
confidence: high
status: [Confirmed]
sources: 1
editor: human editor
human_review: null
version: null
---

As of: 2026-10-14 07:26 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-14-deepmind-govtech-singapore-on-device-scam-detection/

# Google DeepMind and Singapore's GovTech build an on-device scam-detection model from a fine-tuned Gemma

On Oct 14, 2026, Google DeepMind and Singapore's Government Technology Agency (GovTech) announced an on-device scam-detection model built for typical smartphones in Singapore. A light classifier screens incoming messages, then a lightweight Gemma model fine-tuned on local scam conversations checks a longer slice of the conversation and warns the user in real time. Data stays on the phone, which matters because end-to-end encrypted chats cannot be scanned on servers. No app name, launch date or accuracy figure was given.

## Key facts
- Announced in a Google DeepMind blog post (Oct 14, 2026, Responsibility & Safety; authors JJ Geewax and Miguel de Andrés) and presented at Singapore International Cyber Week 2026; part of Google DeepMind's National AI Partnership with the Singapore Government
- Two stages: a lightweight classifier screens incoming messages 'with minimal power use'; a fine-tuned lightweight Gemma model then analyses a larger snapshot of the conversation to confirm or dismiss the threat. On detection the user gets a real-time warning to pause before sharing sensitive information or money
- Privacy: conversation data is processed on the device, not on servers; DeepMind notes that server-side scanning cannot inspect end-to-end encrypted platforms
- Training: fine-tuning ran in GovTech's managed Google Cloud environment with isolated training data, using anonymised scam reports and intelligence data and excluding personal data; the model was stress-tested on synthetic dialogues and local corpora
- Context given by DeepMind: scams are over half of Singapore's reported crimes, losses exceed S$4 billion since 2019, and social-engineering scams account for up to 80% of losses
- GovTech's Director of Trust and Safety: 'Fighting scams shouldn't mean sacrificing personal privacy.'
- Not stated: the app or channel it will ship in, a launch date, supported languages or accuracy numbers

## What happened

Google DeepMind and GovTech Singapore described a scam detector that runs on the phone itself. A cheap first-stage classifier
looks at incoming messages; when something looks suspicious, a small Gemma model fine-tuned on Singaporean scam conversations reads
more of the conversation and decides whether to warn the user. The work was shown at Singapore International Cyber Week and is part
of DeepMind's national partnership with Singapore, which earlier produced a "double-blind" model evaluation with Singapore's AI
Safety Institute.

## Why it matters

It is an example of a government using a small open-weight model on devices to fight AI-era fraud without reading people's messages
on a server. How well it works is unknown: the post gives no accuracy, false-positive rate or rollout date.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 167 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 106 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 197 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 136 days after its cutoff

## Sources

1. [Google DeepMind: Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams](https://deepmind.google/blog/on-device-ai-scam-detection-mddi/) (deepmind.google, official)

## Changes

- 2026-10-14 (filed): Created from the DeepMind post

## Related

- 2026-08-27: [Google DeepMind pilots the first 'double-blind' evaluation of a proprietary frontier model with Singapore's AISI and MLCommons](https://postcutoff.com/e/2026-08-27-deepmind-double-blind-ai-evaluations/index.md)