A Scam Detector That Runs on Your Phone: DeepMind and Singapore Fine-Tune Gemma

Google DeepMind and GovTech Singapore, the country’s government technology agency, announced on October 14 that they are building an AI model to flag scams in real time, and it is designed to run on the phone itself. It is based on Google’s lightweight Gemma model, fine-tuned on local scam patterns such as people pretending to be banks, officials or other trusted figures. When the model spots manipulative behaviour, it shows a warning that advises you to pause before sharing sensitive details or sending money.

The design has two stages. A small, cheap classifier first screens incoming messages using very little battery. Only suspicious conversations go on to the fine-tuned Gemma model, which looks at a larger slice of the chat to confirm a real threat or dismiss a false alarm. Local press reports describe it scanning chats in apps such as WhatsApp and Telegram on Android. The motivation is clear: Singapore says it has lost more than S$4 billion to scams since 2019, including S$410.6 million in the first half of 2026.

On privacy, Google says the model can only access on-device conversations with the user’s consent, and that neither Google nor the Singapore Government can read people’s chats through this system. Training used anonymised scam reports and synthetic conversations, not private messages. Worth knowing before you get excited: there is no release date, the rollout will be phased, and the announcement gives no accuracy figures or model size.

What’s behind it: “on-device” means the AI does its thinking on your own phone instead of sending your messages to a company’s servers. For something as personal as your chats, that is a big deal, because the check never needs to leave your hands. Small models like Gemma make this possible: they are compact enough to fit on a phone yet can be taught a narrow job, like spotting the tone of a fake “your account is blocked” message. Fine-tuning means giving an existing model extra practice on specialised examples, here scams in the languages and styles used in Singapore.

What this means for you: if you are not in Singapore, you cannot use this yet, and nothing here says it is coming to other countries. Still, it shows a practical pattern: privacy-friendly AI that protects you quietly in the background. In the meantime, the old rules still work best. Be suspicious of urgency, never move money because a message told you to, and call the bank or person on a number you already know. Treat any future AI scam warning as a second opinion, not a replacement for your own judgment.

Sources

• Google DeepMind: on-device AI scam detection
• The Star: Google and Singapore’s GovTech developing an AI model to flag scams in real time for Android users