Google and Singapore design scam detection to stay on the phone

An offer of work can start like an ordinary conversation. The concerning request may arrive several messages later. Google DeepMind and Singapore’s GovTech announced on October 14, 2026 a scam-detection collaboration designed to analyze that local context on the phone. Original technical and rollout explanation.

Why the surrounding messages matter

The described system first uses a lightweight screener, then passes suspicious patterns to a fine-tuned Gemma model for broader conversation analysis. Its deployment design uses on-device processing and user consent; rollout into applications will be phased. These are the partners’ descriptions, not an independent audit.

For a creator evaluating such a tool, the useful question is what happens across a sequence. Consider a normal project enquiry, then the same enquiry followed by an unexpected demand for sensitive information. An illustrative comparison would inspect the warning on each sequence and the ordinary enquiry it should leave alone. This is a proposed evaluation, not a test we performed.

That second check matters: a detector that warns on routine enquiries could become something people stop paying attention to. A reassuring result on one example would not establish dependable protection either.

Wait for the actual application details

The announcement does not identify a universal public app, a device list, a release date or Malaysia access. It describes future application integration. Before looking for a download, wait for a named product and its concrete rollout information.

When those details arrive, inspect what the application asks to read and how its consent controls work. On-device analysis addresses where inference happens; you still need to understand the surrounding application's behavior. No model result can substitute for that product-level check.

Our AI output evaluation guide explains why a convincing example and a repeatable evaluation answer different questions.