WhatsApp Launches New AI-Powered Scam Protection Feature
Meta has introduced 'Scam Alert' in WhatsApp, an AI-driven tool that warns users about potential fraud in real-time. The feature runs locally on devices, ensuring end-to-end encryption remains intact.
Spam and fraud have long moved beyond traditional SMS, finding their way into WhatsApp. Now, Meta is announcing a new weapon in the fight against fraud: a smart AI-based feature that warns users in real-time. Here is how it works and why it does not compromise application encryption.
Protection before financial loss
Meta has launched a new feature in WhatsApp called Scam Alert to help users deal with messages from fraudsters. According to the Meta engineering blog, the company developed this feature to protect users at higher risk, such as children and the elderly.
"If the model identifies a message that is likely from a scammer, a warning will pop up for the user within the conversation, which is not visible to the other party. From there, the user can decide what to do: block the sender, report the message, or continue the conversation," the blog states. If the system makes a mistake and marks a legitimate message as malicious, users can mark it as "safe," and the warning will disappear.
The feature uses AI, but Meta emphasizes that the model runs locally on our devices, not in the company's cloud. Message content never leaves the device, and the feature can be disabled at any time.
According to the Federal Trade Commission, scammers have stolen 425 million dollars in the US alone using Meta's messaging app.
End-to-end encryption maintained
Meta explains that the relatively small model is based on the probabilistic classification of conversation structure and "linguistic signals"—patterns the AI recognizes in the text structure. The model was trained on fraud cases previously reported by users.
To track the feature's effectiveness without reading private messages, Meta collects only statistical data via a secure analytics system. The system adds "mathematical noise" to the data—a random numerical value that prevents the identification of specific users. Meta emphasizes that they have no way to access raw information, meaning your actual messages.
To ensure transparency, every model version is registered in a public log, and Meta will publish the model's weights so security researchers can independently verify that the model is intended only for fraud detection. Users will be able to see within the app exactly what data was shared.
