WhatsApp Tests AI-Powered Fraud Detection Feature on Device

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.

How the AI model works locally

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.

Privacy-first detection and user control

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.

User-facing warnings and response options

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.

Feedback loop for improving accuracy

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.

Balancing security and encryption

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.