Singapore puts Gemma on the phone to catch scam conversations

Google DeepMind and Singapore's GovTech are developing a lightweight Gemma model to recognize scam conversations directly on a phone. Local processing is central to the design: DeepMind says private chat data should not leave the device for the cloud.

The model reads the conversation pattern without uploading the chat

The collaboration fine-tunes Gemma for conversational scams. It aims to spot manipulation during a private chat, as an attacker gradually pushes a victim toward a payment or disclosure of sensitive information. The announcement describes a model in development, not a finished protection service available at scale.

During verification, the primary page exposed mainly metadata and a short description. It does not specify the Gemma version, supported apps, deployment date or test results, so no honest claim about accuracy or reach can be drawn from it.

On-device inference reaches places cloud moderation cannot see

For GovTech, the important feature is that protection could follow the course of a conversation without creating a central copy of its content. That offers a route to intervention inside private messengers, where server-side inspection runs into encryption and user trust.

The scale of the problem is concrete. Singapore Police recorded 37,308 scam cases and losses of about S$913.1 million in 2025. In 81.8% of cases, victims transferred the money themselves after being manipulated. At that moment, a warning before payment can matter more than blocking an account afterward.

Privacy does not solve false alarms or adversarial adaptation

Local execution limits content exposure, but it says nothing about accuracy. An oversensitive model will interrupt ordinary conversations, while a cautious one will miss new scripts. Scammers can also change vocabulary, move part of the exchange to a call or coach victims to ignore warnings.

The announcement also leaves open who updates the model, how errors are measured across languages and whether users can review a decision. Without those details, this is a promising security direction rather than a validated product.

Real chats and clear intervention rules will decide the outcome

The decisive signals will be published false-positive and detection rates, support for Singapore's languages and the precise way warnings reach users. It will be equally important to confirm that text remains on the device during model updates and incident reporting.

The project will prove useful only in the field: when people stop a payment after a warning and keep the protection enabled despite occasional mistakes. Technical elegance cannot stop a scammer if the victim's behavior does not change.

Lilith's verdict

Gemma can stand beside a victim before they tap Send. The real test is whether its warning breaks the manipulation instead of becoming wallpaper after the third false alarm.