Tech & Science
WhatsApp is trialing a new optional AI feature that runs locally on users’ phones to detect suspicious messaging patterns and warn users—without sending message content to Meta’s servers.

WhatsApp is testing a new fraud-detection capability designed to identify potentially deceptive messages before users engage with them. The feature operates entirely on the user’s device using on-device artificial intelligence, avoiding transmission of message content to Meta’s servers and preserving end-to-end encryption and chat privacy.
The “fraud warning” feature is optional and activates a machine learning model that runs directly on the user’s smartphone—not in the cloud. According to a recent report, the system scans incoming messages from contacts not saved in the user’s address book, analyzing linguistic patterns and conversational style for indicators commonly associated with scam attempts.
Meta confirms that message content never leaves the user’s device during classification. No automatic report is sent to WhatsApp or Meta when a suspicious message is flagged—only if the user chooses to manually report it. The detection process does not compromise the confidentiality of encrypted chats.
When a conversation is flagged as potentially fraudulent, a warning appears exclusively within the chat interface—visible only to the user, not to the sender. From there, the user may choose to block the sender, report the conversation to WhatsApp, or dismiss the alert and continue the exchange.
If a warning proves inaccurate, the user can mark the conversation as “trusted,” which prevents future alerts for that contact. In such cases, the user also has the option to share the most recent five messages with WhatsApp to help refine the fraud-detection system’s precision.
This initiative forms part of Meta’s broader effort to integrate artificial intelligence into WhatsApp’s anti-fraud infrastructure while maintaining strict adherence to end-to-end encryption standards and user privacy protections.



