DeepMind and Singapore Build On-Device AI to Catch Scams Without Sending Data to Cloud

Google DeepMind said in a blog post on Oct. 14, 2026 that it co-developed an on-device scam detection AI model with Singapore’s Government Technology Agency (GovTech Singapore). The model runs on smartphones, analyzing conversations on the device, not on cloud servers. It shows a real-time warning urging users to pause before sending money or sharing sensitive information.

The system only accesses conversations with user consent. Neither the Singapore government nor Google can view them. This is a key design choice. Scam detection often requires reading messages or listening to calls, but that raises privacy concerns. The new model keeps everything local.

How the model works

The AI is built by fine-tuning a lightweight Gemma model on local scam conversation data. That data includes impersonation of public agencies, banks, local celebrities, and locally trending tactics. Training data excludes personal identifiers and uses only anonymized scam-report data. So the model learns patterns without knowing who the victims are.

The model uses a two-stage structure for battery efficiency. A lightweight classifier runs first. If it flags a conversation as potentially suspicious, a fine-tuned Gemma model makes the final determination. This two-step approach avoids running the heavier model on every message, which would drain the phone’s battery.

Fine-tuning was done in a Google Cloud environment managed by GovTech Singapore. The cloud is used for training, not for runtime analysis.

Privacy as a feature

GovTech director of Trust and Safety Mark Chen said fighting scams should not mean sacrificing privacy. That is the core principle here. Users get scam warnings without handing over their chat logs or call recordings. The system only activates with explicit permission.

This is a notable shift from cloud-based AI assistants that often process data remotely. On-device processing means no network latency, no risk of data interception during upload, and no server-side storage of sensitive conversations.

Rollout and next steps

The rollout into apps will happen in phases after consultation with the Singapore government. That suggests a cautious, staged deployment. The model is trained on local Singaporean scam tactics, so it is not immediately portable to other regions without retraining on localized data. But the architecture is generic.

Technical details

The lightweight classifier runs first and looks for patterns common in scams. The fine-tuned Gemma model is larger but still fits on a modern smartphone. Gemma is Google’s family of open-weight models designed for on-device use without a constant internet connection.

Battery efficiency is a real concern. Running a large language model on every notification would be costly. The two-stage design mitigates that. The first stage filters out the vast majority of benign conversations. Only a small fraction triggers the second, more expensive stage.

What this means for users

If the system is active, users may see a warning during a suspicious conversation urging them to pause before sending money or sharing sensitive information. The system advises rather than blocks, giving the user the final decision. That is a deliberate choice, since blocking could cause false positives that frustrate users.

Broader implications

This project shows a path for AI safety that does not require mass surveillance. Many scam detection tools rely on analyzing data in the cloud, which creates privacy risks. This on-device approach offers a middle ground: AI that is smart enough to catch scams but respectful enough to keep private conversations on the phone.

It also sets a precedent for government-industry collaboration. GovTech provided the scam data and domain expertise. DeepMind provided the model and engineering. The result is a tool that neither could build alone.

Final take

This is a practical, privacy-preserving use of AI. It does not claim to stop all scams. It just gives people a moment to think. That moment can be enough to prevent a fraudulent transfer or a leaked password. The fact that it runs entirely on the device is a strong selling point in an era of data breaches and cloud leaks.

The rollout will be watched closely. If it works well, expect to see similar tools from other tech companies and governments. If it fails, the failure will likely be in false positives or missed detection, not in privacy. That is a good trade-off to test.