WhatsApp is testing a Scam Alert feature that uses on-device machine learning to detect potentially risky messages from people outside a user’s contacts.
InfoQ reports that Meta’s architecture keeps message content on the device while using privacy-preserving techniques to measure performance and protect model delivery. The system includes confidential computing, Oblivious HTTP, differential privacy, and model transparency.
The design reflects the trade-off facing encrypted messaging services. Scam detection is useful only if it can catch suspicious behavior early, but scanning private conversations centrally would undermine the privacy expectations that make those services valuable.
A limited beta means the feature is still being tested, not universally available. The important detail is the direction: WhatsApp is trying to move more safety detection onto the device while still collecting enough privacy-preserving telemetry to understand whether the model works.