Google DeepMind and Singapore's Government Technology Agency (GovTech) announced a model on October 14 that looks for manipulation in message conversations on a phone. Its design gives routine messages a quick first check and sends potentially suspicious ones to a fine-tuned Gemma model for a closer look. The partners describe a way to warn someone during a conversation without sending that conversation to a server for analysis. They have not announced a consumer release or published detection results.

The big change

  • What changed: DeepMind and GovTech described a two-stage, on-device scam detector: a lightweight classifier screens incoming messages, and a fine-tuned Gemma model examines more of a conversation when the first stage raises concern. A detected threat prompts a warning on the phone.
  • Why it matters: The design aims to examine conversational manipulation inside private messaging while keeping inference on the device. Its first stage also limits how often the more demanding contextual model has to run. Those are design choices described by the partners, not independently measured privacy, battery or fraud-prevention outcomes.
  • What to watch: A release route, supported devices and published tests of missed scams, false alarms, speed and power use would show how this model performs outside the partners' reported development work. The announcement supplies none of those details.

A quick screen before the conversation analysis

In DeepMind's account of the project, a lightweight classifier screens incoming messages with minimal power. A message it judges benign ends there. A possible threat moves to a fine-tuned Gemma model, which evaluates a larger snapshot of the conversation and either discards the concern or produces a scam alert. The warning advises the person to pause before sharing money or sensitive information.

The second stage matters because a single message may not reveal the manipulation developing across a conversation. DeepMind says the model was tuned on local scam conversations involving impersonation of public institutions, banks and local figures, as well as regional schemes. It does not identify the exact Gemma version or size, the first classifier's architecture, how many messages reach the second stage, or the threshold for showing a warning. Those missing details make the power and accuracy trade-off impossible to evaluate from the announcement alone.

DeepMind positions the system for messaging services that use end-to-end encryption. In that setting, a service cannot simply inspect encrypted message contents on its servers. The partners say their model processes the conversation on the device. That is a specific claim about the proposed inference path; the post does not give an independent privacy audit or a specification of what an eventual app might collect through other features.

Cloud training and local inference are different steps

The project's training did involve cloud infrastructure. DeepMind says fine-tuning took place in a Google Cloud environment managed and maintained by GovTech, with the training data isolated there. It says private and personally identifiable information was excluded, and that the datasets were anonymized from scam reports and intelligence feeds. The partners also report tests against synthetic dialogues and local corpora. The post does not provide the data preparation method, test protocol or numerical results, so these remain the partners' descriptions of their process.

That development environment should not be confused with the proposed day-to-day operation. The reported design runs message screening and Gemma's contextual inference locally on a phone, then displays the warning there. Cloud-based fine-tuning does not, by itself, imply that a user's live conversation is sent to GovTech's training environment. Equally, the partners' on-device claim does not establish the data practices of an unannounced consumer integration.

DeepMind says the work was presented at Singapore International Cyber Week and that messaging platforms and operating-system vendors could integrate the models. It does not name an integrating service, app or operating-system release, list supported phones, offer the tuned weights, give a price, or state a rollout date. Its reported stress tests are not an independent field evaluation. The useful advance here is a clearly described architecture for examining scam conversations locally; whether people receive a reliable warning in practice remains an open question.

Sources & further reading

  • Google DeepMind's announcement, October 14, 2026: the two-stage flow, Gemma tuning, proposed on-device warning, GovTech-managed cloud training and the partners' governance and testing claims. This is the project team's account, without public accuracy, latency, battery or rollout data.
  • Singapore MDDI's National AI Partnership announcement, May 20, 2026: background on the Google–Singapore collaboration. It does not document the scam model's performance or availability.