A warning before the transfer

Singapore's GovTech and Google DeepMind are developing AI warnings intended to spot scam conversations on a user's own device. Minister Josephine Teo confirmed the collaboration in her October 14 address at Singapore International Cyber Week. Google published technical details the same day.

The intended intervention comes before someone hands over money or credentials: identify signs of manipulation and ask the user to pause. The minister presented it as a further step in Singapore's defences, alongside existing systems that find malicious websites. Those systems' operating figures are not performance results for this new model.

Two checks, kept local

Google says it adapted Gemma, its lightweight model family, to local scam patterns. A small classifier first screens incoming messages. A second stage examines more of the conversation to decide whether to warn or dismiss the suspicion. The design aims to limit how often the heavier analysis runs.

The distinction matters because a suspicious request can look harmless in isolation. Context may reveal impersonation or a gradual attempt to win trust. Whether the model reliably makes that distinction remains a question for testing, not something the architecture alone proves.

What the privacy promise covers

According to Google, processing will run on-device, and conversation access will require consent under the app's terms. It says neither Google nor Singapore's government will gain access to chats through this deployment. That is a claim about this detector, not a blanket assurance about every part of an app or phone.

Training and everyday use are also different stages. Google says fine-tuning happened in a GovTech-managed cloud environment using anonymised scam-report and intelligence data, excluding private or personally identifiable information. Local processing does not mean the model was trained on the handset.

What users still need to know

Google describes a phased rollout within applications. Its announcement does not give a release date, named participating apps, supported device list, measured battery use or accuracy results. It also leaves false alarms and missed scams unquantified.

Those details will determine the practical value of the project. Too many unnecessary warnings can make a safeguard difficult to use; missed manipulation can leave users exposed. For now, the verified development is a jointly announced approach to local scam detection, not a generally available service or proof that financial losses have fallen.

Sources

  1. Singapore MDDI: Josephine Teo's SICW opening addressOctober 14 government confirmation, especially paragraph 14. Existing SATIS figures describe a separate defence, not this model's measured effectiveness.
  2. Google DeepMind: on-device scam-detection collaborationOctober 14 company account of Gemma fine-tuning, two-stage processing, consent, training governance and planned phased rollout. Technical and privacy assurances are attributed; no quantitative accuracy claim is adopted.