Google DeepMind and GovTech Introduce On-Device Conversation Scam Detection for Android

Gemma scam detection is a new project from Google DeepMind and Singapore’s GovTech, introduced on October 14, 2026. It is designed to analyze suspicious conversations directly on Android devices and display a warning to users when a possible scam is detected. Public availability, supported apps, and the installation method have not yet been announced.

The project uses a fine-tuned lightweight version of the Gemma model. Its stated goal is to detect conversation scams without sending message content to external servers. This could be particularly important for chat services with end-to-end encryption, where on-device analysis may work without centrally scanning messages.

Gemma scam detection works in two stages

According to Google DeepMind, the system has a two-stage architecture. First, an efficient classifier selects messages that may indicate a scam. The Gemma model then evaluates the broader conversation context. This approach is intended to reduce computing requirements, since the more complex model does not need to process every message.

Scams carried out through messages often do not originate from a single sentence. Attackers may gradually persuade victims to provide sensitive information, transfer money, or install unknown software. Evaluating the broader conversation context could therefore identify social engineering patterns that simple filtering of individual messages may not recognize.

Channel NewsAsia, citing Google, reported that the model is intended for Android and is supposed to display an alert directly on the device when a scam is detected. However, it has not been confirmed whether it will be part of the Android system, a feature of specific communication apps, or a solution made available through device manufacturers.

On-device processing and training data

Google says that conversation content is intended to remain on the device. The model therefore is not supposed to send messages to external servers for evaluation. This approach could enable scam protection in encrypted chats without centrally scanning message content.

According to Google, private or personal data belonging to citizens was not used for training. The company says it used anonymized data from scam reports and intelligence sources. However, claims about privacy protection and effectiveness currently come primarily from the organizations involved in the project.

No independently verified data has been published on the model’s accuracy, the number of false positives, or results from real-world use. Without these metrics, it is not possible to assess how reliably the system can distinguish scam communications from legitimate messages that may use similar language or concern similar topics.

Google and Singapore collaboration

The development follows the National AI Partnership between Google and Singapore’s Ministry of Digital Development and Information (MDDI). The partnership was announced in May 2026, and the collaboration on scam detection was established במסגרת this framework.

The project’s practical significance will depend on how it is deployed. A pilot or public release date, a list of compatible Android devices, and a list of supported communication services are still unavailable. Google says it will provide further details later.

In future announcements, it will be important whether the organizations involved publish testing methodology, accuracy data, and mechanisms for limiting false alerts. Independent verification of how data processing and user privacy are handled in practice will also be important.

Verified and updated: October 14, 2026 06:23