As scam tactics evolve from impersonation to social engineering and AI-generated lures, WhatsApp is testing a new defence: Scam Alert. The optional feature runs an on-device machine learning model that warns the user about potential scam messages.

Message content never leaves the device

This is the technically significant part of the announcement. No message content leaves the device for classification, and nothing is auto-reported to WhatsApp, Meta or anyone else. The feature complements end-to-end encryption rather than displacing it.

The company ties this to three principles:

  • On-device only: The model and the message data it processes both stay on the device.
  • No automatic reporting: WhatsApp cannot initiate sharing of any user data without the user's action. The only way message content — or even the fact that a scam was detected — reaches the servers is if the user explicitly chooses to report it.
  • User control: The feature can be switched on or off at any time.

How it works

Once the user turns it on, Scam Alert downloads a machine learning model to the device. The model classifies incoming messages from non-contacts according to whether they match known scam patterns.

It is trained on patterns observed in scam conversations that users have reported. The assessment is probabilistic, based on conversational structure and linguistic signals. If the model flags a message as a likely scam attempt, the user sees a warning in the chat — and that warning is not visible to the other person.

Why this is possible now

According to the company, recent advances in on-device machine learning are what make this feasible. The performance, battery and model-size tradeoffs that previously made on-device classification impractical no longer apply, and accurate text classification can run entirely on mobile hardware.

Scam Alert suits that approach particularly well: the model is small enough to run on-device, simple enough to publish for independent review, and effective without server-side components.

Technical overview first, rollout second

The ordering is itself a choice. The company published this overview before making the feature available to all users, and released it only in a limited Beta rollout.

To validate the implementation, it says it will continue working with its bug bounty community to stress-test the system, and welcomes feedback from the broader security research community. Any feature that inspects message content inside an end-to-end encrypted app naturally attracts suspicion; publishing the claim before inviting scrutiny is the most direct answer available.

What it does not solve

The feature's limits are part of its design. The model only looks at messages from people who are not in your contacts; a scam arriving from a familiar number — a message sent from a compromised account, say — falls outside the net.

Because the assessment is probabilistic, both false alarms and misses are possible. That the warning is invisible to the other party matters here: a wrongly flagged message gives the sender no signal at all, leaving the judgment entirely with the recipient.

Training the model only on patterns from user-reported conversations is another constraint. As scam tactics shift, the model has to be updated, and a genuinely new method is by definition absent from the pattern pool the first time it appears.