Google DeepMind and Singapore’s Government Technology Agency have developed a model designed to recognise scam conversations on a user’s device. Announced on 14 October, it is intended for phased integration into applications, rather than immediate availability as a standalone consumer app.[1]
The collaboration addresses a change in scam tactics. A persuasive scam may build trust across many messages, explain away warning signs and delay the request for money until the victim has accepted the story. A detector that judges one sentence at a time can miss that progression.[1][2]
Two stages divide speed from judgement
The announced architecture uses a lightweight classifier as a first filter. When that stage detects enough risk to justify a closer look, a fine-tuned, lightweight Gemma model examines the conversation in more context. Google says the analysis runs on the device, reducing the need to send private conversation data to a remote service.[1]
The proposed detection path
| Stage | Job | Why it matters |
|---|---|---|
| 1. Fast classifier | Screen the exchange for possible scam signals | Most conversations do not require the cost of deeper analysisSource 1 |
| 2. Contextual Gemma model | Interpret the suspicious conversation as a sequence | Pressure, impersonation and changing requests can be assessed togetherSource 1 |
| 3. User warning | Intervene before sensitive information or money is sent | The useful moment is before the irreversible actionSource 1 |
Local processing changes the privacy trade-off
With on-device processing, the conversation can be analysed where it already resides. Google says deployment will require user consent under the host application’s terms, and that neither Google nor the government will gain access to conversations through this model. The engineering constraint is fitting useful contextual analysis within a phone’s memory, computation and battery budget.[1]
Google says the training environment was a GovTech-managed sovereign sandbox and that the work used anonymised scam reports and intelligence feeds rather than private citizen personally identifiable information. The announcement also says the system was tested with synthetic dialogues and local corpora, but it does not publish an accuracy figure, checkpoint or general release date.[1]
It joins a layered defence rather than replacing one
Singapore already operates SATIS, a separate website-defence system. In her 14 October speech, Minister Josephine Teo said it scans more than 600,000 sites daily and disrupts around 20,000 malicious sites monthly. GovTech’s product page, last updated on 12 May, still gives the earlier figures of more than 400,000 and 15,000 respectively. Neither set measures the new conversational model.[2][3]
The distinction reveals why anti-scam systems are multiplying. One tool can block infrastructure; another can filter calls or messages; an on-device model can recognise the social-engineering pattern even when the contact channel or link is new. Each covers a different point in the chain.
The proposed intervention sits before the payment or disclosure. A suspicious link can be replaced and a phone number changed; a conversation may still reveal impersonation, pressure or a manufactured reason to transfer money. Detecting that pattern locally is the distinctive contribution this project aims to make.
Sources & further reading
- Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection Google DeepMind · 14 Oct 2026 · Accessed 15 Oct 2026
- Opening Address by Minister Josephine Teo at the Singapore International Cyber Week (Day 2) on 14 Oct 2026 Cyber Security Agency of Singapore · 14 Oct 2026 · Accessed 15 Oct 2026
- Scam Analytics and Tactical Intervention System (SATIS) GovTech Singapore · 12 May 2026 · Accessed 15 Oct 2026