Meta adds AI screening to detect WhatsApp scams
Meta is rolling out an optional Scam Alert feature on WhatsApp using on-device machine learning to flag suspicious messages and protect user privacy.
Stock photo for illustration only, not from the actual event
- Meta tests Scam Alert on WhatsApp using on-device AI screening.
- Warnings appear only to the user and remain hidden from the sender.
- Users can block, report, or continue the conversation after a warning.
- No message content leaves the device during the classification process.
Meta has begun rolling out an optional Scam Alert feature as part of a limited beta on the messaging platform WhatsApp. The tool utilizes on-device machine learning to identify and flag suspicious messages that may indicate fraudulent activity. Earlier this year, Meta also introduced scam detection features specifically for device linking requests on the platform.
The newly introduced system relies on an artificial intelligence model running directly on the user's smartphone to determine if an incoming chat resembles a scam attempt. When the model identifies a likely scam, a warning message is displayed within the chat interface, which stays completely invisible to the other person involved in the conversation.
Stock photo for illustration only, not from the actual event
Upon receiving a warning, users retain full control and can choose how to proceed with the following actions:
- Block the contact
- Report the chat
- Continue the conversation normally
If a user believes a warning was triggered incorrectly, they can mark the chat as trusted. Doing so removes the warning immediately, and the Scam Alert feature will not flag that specific chat again. Furthermore, users who mark a chat as trusted can choose to opt in by sharing the last five received messages with WhatsApp to help enhance the overall accuracy of the detection feature.
Deploying AI models directly on user devices for scam detection represents a crucial shift toward privacy-first cybersecurity. By keeping data local, platforms can address growing user concerns over data surveillance. However, the true test for such on-device systems lies in balancing high detection accuracy with minimal false positives, especially as scammers continually evolve their tactics across different languages and regions. Allowing users to voluntarily share limited message data upon marking a chat as trusted provides a valuable feedback loop for ongoing machine learning improvements.
With a user base exceeding 3 billion people, WhatsApp remains a primary target for widespread scam activities, including wire transfers and investment schemes. According to the Federal Trade Commission (FTC), victims reported losing $425 million to scams executed solely via WhatsApp out of more than $2.1 billion total lost to social media scams throughout 2025.
Meta emphasizes that no message content leaves the user device for classification purposes when the scam detection feature is active, and nothing is auto-reported to WhatsApp, Meta, or any external parties. Additionally, users retain the flexibility to turn off the Scam Alert feature at any time.
Source: The Verge
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment