Google and Singapore teach on-device AI to catch chat scams

Google DeepMind and GovTech Singapore are working together on a fine-tuned model that detects scams. It is built on lightweight Gemma models that run directly on the user's device. The target is conversational fraud in private chats, where the scammer builds the deception through dialogue.

The key design choice is that nothing leaves the phone: analysis happens locally and no data goes to the cloud. That removes the main objection to scanning private messages, namely who gets to read them. Conversational scams are also hard to catch with keyword filters because the fraudster moves step by step, so the model has to understand context.

What happens if an employee gets a message from the 'director' asking to pay an invoice urgently? Check whether you have a rule to confirm such payments through a second channel, and show your team what chat-based fraud looks like. Suspicious operations in the bank and 1C can be caught automatically too: 1Cifer compares periods and flags anomalies before a person would notice them.