Gemma on-device fraud detection: Singapore's fine-tuned government AI
- The thread: fine-tuned open models doing on-device government AI work — a consumer-fraud counterweight to the agent-incident record, with privacy carried by the on-device claim rather than a cloud VM.
-
2026-10-14 (first record): Google DeepMind × GovTech Singapore (announcement post, ~10:05 PM ET Oct 13 — https://x.com/GoogleDeepMind/status/2110190101707858039): on-device fraud detection built on the open Gemma model, fine-tuned on local scam tactics, real-time threat identification with a "without compromising personal privacy" frame, against the cited stat that over half of reported crimes in Singapore use sophisticated social engineering. No launch timing, accuracy figures, or GovTech-side statement in the window. Watch: GovTech confirmation, deployment scope, published fine-tuning details, other governments copying the template.
-
2026-10-14 (~12:00 PM ET): the consumer-side counterpoint — The Verge reports Android introducing new scam-defense features, headlined by a new "enhanced protection" setting for Google Messages rolling out in the coming weeks that taps cloud-based Gemini models to evaluate a scammer's intentions in conversations with unrecognized contacts, building on the existing Messages Scam Detection, with additional safeguarding tools alongside (https://www.theverge.com/tech/1010928/android-scam-protection-enhanced-protection-google-messages). Development: the Singapore play's no-egress on-device pitch now sits beside Google's own consumer defense — cloud inference, and announcement-stage rather than shipped. Watch updated: actual rollout dates, whether any Messages detection stays on-device, and the standing AI-generated-scam-script eval question.