Google DeepMind and the government agency GovTech Singapore have unveiled a model that detects scam conversations directly on the user's device and warns against rushing to transfer money or share personal data. The system is based on a fine-tuned open-source Gemma model, showcased on October 14, 2026, at Singapore International Cyber Week. All computations are performed locally, so messages are not sent to the cloud: neither Google nor the Singaporean government can read them through this deployment.

image
image
image

What happened

On October 14, 2026, at Singapore International Cyber Week, Google DeepMind and GovTech Singapore demonstrated a scam dialogue detection system that runs entirely on the user's device. The open-source Gemma model here is fine-tuned on anonymized local fraud reports: impersonation of government agencies and banks, and regional schemes. Fine-tuning took place within the 'Sovereign Sandbox'—a GovTech-managed GCP environment, and the announcement was published in a DeepMind blog post jointly with MDDI. The architecture is two-stage: a lightweight, low-power classifier screens the entire incoming message stream, and only when it flags a conversation as suspicious does Gemma engage to analyze a longer dialogue fragment and confirm or clear the alert.

Context

The scale of the problem in Singapore justifies the choice of domain: according to DeepMind, fraud accounts for more than half of registered crimes in the country, losses since 2019 have exceeded 4 billion Singapore dollars, and up to 80% of the damage comes from long-term trust-building schemes. Such long manipulative dialogues are poorly recognized by classic keyword filters; they require analysis of the extended context of the conversation. The project brought three elements together in a single product cycle: open model weights, national data in anonymized form, and a sovereign cloud environment for fine-tuning controlled by a government agency. This combination enables a deployment compatible with end-to-end encryption, as inference is performed on the device.

Why this matters for the industry

For the industry, this is the first public example of a national government fine-tuning an open LLM in its own cloud 'sovereign' environment and deploying it on citizens' devices—a working template for government edge-AI. The 'lightweight classifier plus heavy model on trigger' cascade solves the main problem of constant on-device inference, namely battery drain: the main message stream is handled by the classifier, and Gemma wakes up only upon suspicion. The same framework can be transferred to anti-fraud in chats, email triage, and other local tasks, and the scheme of fine-tuning an open model on national data is reproducible by other countries and messenger developers. The system cannot be used 'as is': APIs and SDKs have not been presented, and only a documented reference architecture example is currently available. However, DeepMind has already stated that it is considering similar projects in other countries—a signal of where the open model market is shifting: toward local inference and government fine-tuning templates.

Why this matters for users

There is no direct effect for readers at the moment: specific applications and implementation timelines have not been named, and the launch will be phased and in collaboration with Singaporean authorities. If the feature appears in messengers, a warning about manipulative conversations will appear locally, without sending messages to the cloud, and neither Google nor the government will be able to read the conversation through this deployment. How the display of warnings and consent for dialogue analysis will be structured is currently unknown. It is worth watching which applications will feature this protection and how it will work in conjunction with conversation encryption.

What is currently unknown / limitations

There are no public metrics for the system: accuracy, false positive rates, and power consumption have not been disclosed, so confidence in its effectiveness is not methodologically justified—the architectural and product intent is confirmed, but the model's proven capabilities are not. The verifiability of local inference depends on the implementation of specific applications, and sources do not describe how a user can verify that the conversation truly does not leave the device.

Sources

Author

Look at AI, editorial team