Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection

Summary

Google DeepMind and Singapore's MDDI and GovTech fine-tuned a lightweight Gemma model on local scam conversations to detect scams on the user's device. A small classifier screens incoming messages and passes suspicious ones to the Gemma model for multi-turn analysis, which shows a warning when it detects manipulation. Announced at Singapore International Cyber Week 2026; not yet available.

Key points

• Base model: a lightweight Gemma; the exact variant, size, and fine-tuning method are not given.
• Training data: anonymised scam reports and intelligence feeds (impersonation of institutions, banks, local figures), with no personal data; trained in GovTech's managed Google Cloud environment.
• Two-stage pipeline: cheap classifier first, then the Gemma model only for flagged conversations.
• Runs on-device with user consent; neither Google nor the government can see the conversations.
• No accuracy figures published; the model was stress-tested on synthetic dialogues and local corpora. Rollout into apps will be phased.

Why it matters

A clear example of the pattern "small open model, fine-tuned on domain data, run on device" for privacy-sensitive classification, with a cascade to keep costs low.