Meta is testing a new way to catch suspicious WhatsApp messages before users get pulled into a scam, and the interesting part is where the detection actually happens: on the phone itself.
The company has offered an early look at WhatsApp Scam Alert, an optional feature powered by an on-device machine learning model. Instead of sending private conversations to Meta’s servers for analysis, the system is designed to examine certain incoming messages locally on a user’s device.
That distinction matters. WhatsApp has spent years making end-to-end encryption central to its privacy pitch, so introducing automated scam detection could easily raise questions about who — or what — is reading people’s messages.
Meta seems very aware of that problem.
WhatsApp Scam Alert Runs Directly on Your Phone
Once enabled, Scam Alert downloads a machine learning model onto the user’s device. The model checks incoming messages from people who aren’t in the user’s contacts and looks for patterns associated with known scams.
Meta says the model was trained using patterns found in scam conversations previously submitted through user reports. Rather than simply searching for individual suspicious words, it can consider conversational structure and linguistic signals when estimating whether a message looks fraudulent.
The classification itself stays on the phone. WhatsApp says message content isn’t automatically sent to Meta, WhatsApp or another third party just because the system analyzes it.
A Suspicious Message Won’t Automatically Be Reported
Getting flagged doesn’t mean a sender is immediately reported.
If Scam Alert believes a message could be part of a scam, WhatsApp displays a warning inside the conversation. The person who sent the message doesn’t see that warning.
The recipient then decides what happens next. They can block the sender, report the conversation or simply continue chatting. If the alert appears to be wrong, the user can mark the chat as trusted, preventing Scam Alert from continuing to flag that conversation.
That leaves the final decision with the user instead of turning an automated prediction into an automatic enforcement action.
Meta Is Trying to Add AI Without Weakening WhatsApp Encryption
This is probably the bigger story behind the feature.
AI-powered safety tools usually work better when they can analyze large amounts of information on centralized servers. WhatsApp has a different problem. Its end-to-end encrypted architecture is built around limiting access to private message content in the first place.
Scam Alert attempts to work around that tension by moving the machine learning model to the user’s device rather than moving the conversation to Meta.
Meta says the feature is optional and can be switched on or off by the user. It also says no message content leaves the device as part of the normal classification process.
That’s a pretty deliberate design choice, particularly at a moment when AI features inside private messaging apps are receiving much more scrutiny.
WhatsApp Will Still Collect Limited Performance Signals
Scam Alert isn’t completely disconnected from WhatsApp’s systems.
Meta wants to know whether its warnings are actually useful. The system therefore measures limited information such as how frequently scam warnings appear and what users do after seeing them.
According to Meta’s technical explanation, those measurements are reduced to aggregated counts and processed through privacy-preserving infrastructure. The company says it does not receive individual message content or conversation-level information through this analytics process.
Users can separately choose to share message samples to improve the feature. For example, someone marking a conversation as trusted can opt to send the last five received messages to WhatsApp. That sharing requires the user’s action rather than happening automatically.
Meta Wants Researchers to Be Able to Check the System Too
There’s another unusual piece to Meta’s approach: outside verification.
The company says versions of the scam-detection model will be recorded through a public transparency system before being distributed. Meta also plans to make model information available so security researchers can examine whether the technology is actually limited to detecting scams.
Users will eventually have access to transparency logs showing information such as which messages were analyzed, whether the model flagged them and which model version was involved. Meta is also expanding its Bug Bounty work around the system.
That doesn’t automatically settle every privacy concern, obviously. It does make the system considerably more inspectable than a black-box server-side classifier.
WhatsApp Scam Alert Is Still in Beta
Don’t expect every WhatsApp account to suddenly have the feature.
Meta says Scam Alert is beginning with a limited beta rollout and will be tested further before any broader production release. Security researchers and members of Meta’s Bug Bounty community are being invited to stress-test the system during this stage.
Social Media Today reported on August 12, 2026, that Meta plans to continue refining the feature before expanding availability.
No firm worldwide release date has been announced yet.
Scam Detection Is Becoming a Bigger WhatsApp Priority
The new system isn’t appearing out of nowhere.
WhatsApp has already been adding protections aimed at suspicious account-linking attempts, questionable group invitations and other common scam tactics. Meta has also been expanding scam-detection systems across Facebook and Messenger.
Scammers keep changing their scripts. Some impersonate businesses or family members. Others push fake jobs, investment schemes or elaborate social-engineering conversations that may begin with an innocent-looking message.
A warning arriving before that conversation gains momentum could be useful.
The harder problem is doing it without turning private messaging into another stream of data sent back to a technology company’s servers. Meta’s answer, at least with Scam Alert, is to put the detection model on the device and give users the final call.
Whether the system catches enough real scams without constantly flagging harmless conversations will become clearer once the beta expands.