WhatsApp is testing Scam Alert, a new security feature that uses on-device machine learning to identify conversations that may involve scams. The feature is being introduced as an additional layer of protection against increasingly sophisticated scam attempts, including impersonation and social-engineering messages. Unlike server-based detection systems, Scam Alert performs its classification directly on the user’s device. WhatsApp says no message content leaves the device for classification, allowing the feature to work while maintaining the app’s existing end-to-end encryption protections.

Scam Alert: Key Specifications

• Technology: On-device machine learning


• Detection: Probabilistic classification of potential scam patterns


• Messages analyzed: Incoming messages from non-contacts


• Processing: Performed locally on the user’s device


• Automatic reporting: Not enabled


• User control: Feature can be switched on or off


• Warning: Shown privately to the user


• Current status: Limited beta rollout

How WhatsApp Scam Alert Works

Scam Alert is an optional feature. Once enabled, it downloads a machine learning model to the device. The model analyses incoming messages from people who are not saved as contacts and looks for patterns associated with known scam conversations. WhatsApp says the model uses conversational structure and linguistic signals to make a probabilistic assessment of whether a message is likely to be part of a scam attempt.

It is trained using patterns observed in scam conversations that users have previously reported to WhatsApp. If the system identifies a likely scam, a warning appears inside the chat. The warning is not visible to the other person. The recipient can then choose whether to block the sender, report the conversation, or continue chatting. If a warning is incorrect, users can mark the conversation as trusted. Scam Alert will then stop flagging that particular chat.

On-Device AI Protects Message Privacy

The central difference between Scam Alert and a conventional server-based detection system is where the analysis happens. Meta says both the machine learning model and the message data it processes remain on the user’s device. No message content is automatically reported to WhatsApp, Meta, or another third party. Even the fact that a scam was detected is not automatically sent to the company’s servers.

WhatsApp has also designed the system so that it cannot automatically share user data without an action from the user. The company’s approach is intended to allow scam detection without requiring WhatsApp to access the contents of private conversations. The company says anonymous and aggregated usage information can still be used to understand whether the feature is working. This information is processed through privacy-preserving systems rather than sending the underlying message content to WhatsApp.

What Changes for WhatsApp Users?

The new feature adds a layer of protection that users previously had to handle more manually. Instead of relying only on recognizing suspicious messages themselves, users who enable Scam Alert can receive an in-chat warning when WhatsApp’s model detects patterns associated with potential scams. The system does not automatically block the conversation, however.

The final decision remains with the user, who can block, report, or continue the chat. This is particularly relevant for scams that begin with messages from unknown numbers and attempt to persuade recipients to continue a conversation, share information, or take another action. WhatsApp’s approach also means users do not have to trade the privacy of their message content for the additional warning system. The classification itself remains on the device, while the feature stays optional and user-controlled.

Users Can Help Improve Detection

WhatsApp has added an additional feedback option for incorrectly flagged conversations. If a user marks a chat as trusted, they can optionally share the last five messages received with WhatsApp to help improve the feature’s accuracy. This sharing does not happen automatically. It requires the user’s explicit choice, consistent with WhatsApp’s broader approach of keeping message reporting under user control.

Limited Beta Rollout Underway

Scam Alert is currently being introduced through a limited beta rollout rather than a full release to every WhatsApp user. Meta says it is publishing the technical details early and working with its Bug Bounty community and security researchers to stress-test the system before making it broadly available. The company has also designed safeguards intended to make the feature independently verifiable. WhatsApp says model versions are published on a transparency ledger before deployment, while model weights will be available for independent security researchers to examine.

Conclusion

WhatsApp’s Scam Alert adds an AI-based warning system that can identify potential scam patterns directly on a user’s device without sending message content away for classification. The optional feature gives users another layer of protection while keeping the decision to block, report, or continue a conversation in their hands. For now, Scam Alert remains in limited beta testing as WhatsApp continues evaluating its accuracy, privacy protections, and performance before a wider rollout.l