WhatsApp is testing Scam Alert, an optional feature that runs a machine learning model on your phone to warn you when a message from someone outside your contacts looks like a scam. Meta published a technical overview of how it works on 12 August, and the feature is currently in a limited Beta rather than a general rollout.
How WhatsApp Scam Alert Works
Scam Alert is off by default. Turn it on and WhatsApp downloads a machine learning model to the device, where it classifies incoming messages from non-contacts against known scam patterns. Meta says the model was trained on patterns observed in scam conversations that users themselves reported, and that it does probabilistic classification based on conversational structure and linguistic signals rather than anything resembling a keyword blocklist.
When the model thinks a message is a likely scam, a warning appears inside the chat. The person who sent the message does not see it. From there you can block, report, or simply carry on. If the warning looks wrong, you can mark the chat as trusted and it will not be flagged again.
The design question everyone asks about a feature like this is whether it breaks end-to-end encryption. On Meta’s account it does not, because nothing is scanned anywhere but on your handset. From the company’s own post:
No message content leaves the device for classification or is auto-reported to WhatsApp, Meta, or anyone else.
Meta lists three stated principles behind that: on-device only processing, no automatic reporting, and user control over turning the feature on or off at any time. The company also says it publishes model weights so outside security researchers can check the model was built for scam detection and nothing else, and that every model version — including experimental ones — goes onto a public transparency ledger before deployment, so neither Meta nor WhatsApp can quietly ship a special model to one specific user.
Where Your Messages Can Still Leave the Phone
The on-device claim holds for classification. It is worth reading the rest of Meta’s post carefully anyway, because there is one path where message content does travel, and it is attached to the gesture most people will use most often.
If you mark a chat as trusted — that is, if you tell WhatsApp the warning was wrong — you can also opt in to share the last five messages you received with WhatsApp to help improve the model’s accuracy. It is opt-in and it sits behind an explicit choice, which is the right way to build it. But it means the “this is not a scam” button is also the point where you are asked to hand over five messages from a private conversation, and a tap made in mild annoyance is not the same as a considered decision about sharing chat content. Read that prompt rather than dismissing it.
The other thing to keep in perspective: the model itself was built from message patterns in conversations that other users reported. That is a reasonable way to train a scam classifier, and it is also a reminder that “on-device” describes where the inference runs, not where the training data came from.
Checking What Got Flagged
Meta says users can audit the feature themselves through in-app logs. The transparency log listing which messages the on-device model flagged sits at Account > Request Info > Scam Alert Activity in WhatsApp’s settings. Anonymous, aggregated telemetry about whether the feature is working — warning counts and user action counts, not content — is processed inside confidential virtual machines and sent to WhatsApp only as differentially private aggregates.
Availability
Scam Alert is in limited Beta right now, and Meta framed the technical overview as an early look published deliberately ahead of a wider release, alongside continued work with its bug bounty community. There is no date for general availability and no market-by-market rollout schedule. If you do not see the toggle yet, that is expected.
For a platform where scam and impersonation messages are a daily fact of life for hundreds of millions of users — across India, Brazil, Indonesia and elsewhere — a warning that runs locally and leaves the decision to the recipient is a better shape of solution than server-side scanning would have been. The test is whether the flags are accurate enough that people do not train themselves to swipe them away.
