Google DeepMind announced on October 14, 2026 that it has been working with the Government Technology Agency of Singapore (GovTech Singapore), as part of a National AI Partnership with the Singapore Government, to build a fine-tuned, on-device scam detection model. According to the DeepMind blog post, authored by JJ Geewax and Miguel de Andrés, 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. The post states that scam tactics have shifted from simple phishing to social engineering, in which attackers manipulate victims through convincing conversations before requesting money, and that such scams now account for up to 80% of all losses. To address local deception techniques that the company says generic models struggle with, Google's lightweight Gemma model was fine-tuned on local scam conversations, including impersonations of public institutions, banks and local figures. DeepMind says the model can process conversation data directly on-device rather than relying on server-side cloud scanning, allowing faster analysis of conversation patterns to detect manipulative intent. The company states the rollout within applications will be carefully phased and developed in partnership with the Singapore Government.
- Scams account for more than S$4 billion in losses in Singapore since 2019, per DeepMind
- Google's Gemma model was fine-tuned on local scam conversations and runs on-device rather than in the cloud
- The project is part of Google DeepMind's National AI Partnership with the Singapore Government; rollout will be phased
What it means for you
This is a government-scale security project, not something you'll buy or deploy. The useful takeaway for anyone running an organization is the pattern it confirms: scams have moved from obvious phishing links to patient, conversational manipulation — someone impersonating a bank, a vendor, or a senior colleague and talking a person into moving money. That's a people-and-process risk, and no model your staff can use today changes it.
Try this
Pick your highest-risk money action — approving a payment, changing vendor bank details — and confirm there's a verification step that doesn't rely on a single message or call (e.g. a callback to a known number). Map where that step is missing.
Who should care
Whoever owns payments, accounts payable, and vendor onboarding, plus anyone who can authorize a transfer or a bank-detail change.
Skip this if
You already require independent, out-of-band verification for every payment and detail change, and your team knows impersonation scams target conversations, not just inboxes.
Sources: Google DeepMind — read the original