WhatsApp Scam Alert uses on-device ML to flag likely scam chats in beta

Meta is rolling out Scam Alert as an optional, limited-beta feature on WhatsApp that uses on-device machine learning to flag likely scam messages. When the classifier judges a message to be a probable scam, the user sees a warning inside the chat; the warning is rendered locally and is not visible to the other participant. From the warning, the user can block, report, keep chatting, or mark the chat as trusted, in which case the warning is removed and Scam Alert will not flag that thread again. The feature follows an earlier scam-detection rollout this year for device-linking requests on WhatsApp.

Because classification runs on the user's device, there is a local compute and battery cost on every WhatsApp-capable phone, with the model size and inference frequency determining the real overhead. Meta states that no message content leaves the device for classification and nothing is auto-reported, but the trust flow introduces an explicit data-sharing choice: marking a chat as trusted can surface an opt-in to send the last five received messages in that thread to WhatsApp to improve the model. That five-message upload is a distinct, user-initiated telemetry channel separate from the on-device classifier, and it is the only path by which message content reaches Meta under this feature. The feature is opt-in and can be disabled at any time.

Scam Alert is a limited beta, so coverage is partial and the classifier's precision and recall on real WhatsApp traffic are not publicly documented. A user who marks a chat as trusted disables further warnings on that thread, which means a single dismiss can blind the feature to an evolving social-engineering attempt in the same conversation. The opt-in five-message share is voluntary, so the improvement loop depends on user cooperation rather than continuous learning. The scale of the problem is large: WhatsApp has more than three billion users, and the FTC reported $425 million lost to WhatsApp scams alone and more than $2.1 billion across social-media scams in 2025.