WhatsApp has become one of the world’s most common channels for fraud. Impersonation messages, fake job offers, investment pitches, romance scams, and sudden requests for money or one-time passwords now arrive every day from numbers people do not recognise. Meta’s response is a new optional tool called Scam Alert. It is designed to warn users when a message from an unknown contact looks like a scam, without sending private chat content to WhatsApp or Meta for analysis.
The feature is still rolling out in limited Android beta, but Meta has already published a detailed technical explanation of how it works. The core idea is simple: run a small machine learning model on the user’s phone, flag suspicious patterns locally, and leave the final decision with the person who received the message.
Why WhatsApp Needed a New Layer of Protection
End-to-end encryption is one of WhatsApp’s strongest selling points. It also creates a problem for scam detection. If WhatsApp cannot read messages, it cannot scan them on a central server the way an email provider might scan spam. That limitation has helped scammers. A stranger can open a chat, build trust, and then ask for money or personal details before the victim realises what is happening.
WhatsApp already shows some context for unknown numbers, such as whether the number is saved as a contact, whether you share groups, and the country where the number is registered. It also warns users about suspicious device-linking attempts. Scam Alert goes further. Instead of only showing metadata about the sender, it looks at the message itself for linguistic and conversational signs of fraud.
The feature is optional and off by default. That matters. WhatsApp is not forcing an AI filter onto every chat. Users who want the extra warning can turn it on. Users who do not want it can leave it off.
What Scam Alert Actually Checks
Scam Alert only examines incoming messages from people who are not saved in the recipient’s contacts. Chats with saved contacts are left alone. Once the feature is enabled, WhatsApp downloads a compact machine learning model to the device. That model then runs locally whenever a new message arrives from an unknown number.
Meta says the model was trained on patterns found in scam conversations that users had already reported. It does not claim to “understand” every message the way a large language model might. It performs probabilistic classification based on conversational structure and linguistic signals. Typical red flags include urgent requests for money, pressure to share PINs or passwords, investment or lottery language, impersonation, and the trust-building style used in romance or pig-butchering scams.
If the model scores a message as a likely scam, WhatsApp displays a warning inside the chat. Reports from the beta describe the banner as “This may be a scam.” The warning is visible only to the recipient. The sender does not see it and is not told that Scam Alert is active on the other side. That is intentional. If scammers knew the message had been flagged, they could immediately change tactics.
What Happens After the Warning Appears
Scam Alert does not block anyone automatically. WhatsApp leaves three practical choices with the user.
The first is to block and report the contact. That stops further messages and sends a report to WhatsApp in the same way existing user reports already work. The second is to continue the conversation if the warning looks wrong. The third is to mark the chat as trusted. Trusting a chat removes the warning and tells Scam Alert not to flag that contact again.
If a user marks a chat as trusted, WhatsApp may offer an extra option: share the last five received messages so the model can learn from a possible false positive. That sharing is voluntary. Nothing is uploaded unless the user agrees.
This design keeps the feature from becoming a silent censor. A delivery person, a new colleague, a relative using a different number, or a small business could all trigger a warning. The user can override it.
How WhatsApp Claims Privacy Is Preserved
The most important technical claim is that classification happens on the device. Message text is not sent to WhatsApp, Meta, or a third-party server so the model can decide whether it looks like a scam. Meta’s engineering team described three design principles: on-device processing, no automatic reporting, and user control.
On-device processing means the model and the message data stay on the phone. No automatic reporting means WhatsApp cannot start sharing message content, or even the fact that a scam was detected, unless the user taps report. User control means the feature can be switched off at any time.
Meta also published extra safeguards aimed at security researchers. Model versions are recorded on a public transparency ledger before they are deployed. The company says it cannot send a special model to one particular user. Independent researchers are being invited to inspect the system through an expanded bug bounty programme. Meta also says model weights will be published so outsiders can check that the model is built for scam detection rather than broader content surveillance.
WhatsApp still wants some feedback on whether the feature works. For that, it collects only anonymous aggregate counts: how often warnings appear, and what users do next. Those counts are processed with differential privacy inside Trusted Execution Environments. Meta says the aggregates are designed so they cannot identify an individual user. Users can also review an on-device activity log showing when the feature was enabled, which messages were flagged, and which model version was used. Those logs stay on the phone.
How to Turn It On
In the current Android beta, the switch sits under Settings and then Account. After it is enabled, WhatsApp downloads the model and begins checking messages from unknown senders. The same screen can be used to turn the feature off.
The rollout is gradual. Installing the latest WhatsApp beta does not guarantee that Scam Alert will appear. WABetaInfo and other trackers have seen it on recent Android beta builds, including versions in the 2.26.34 range, but only for some accounts. iOS has not received the same public beta access yet. WhatsApp has not announced a date for a full public launch.
What the Feature Cannot Do
Scam Alert is a warning system, not a guarantee. A flagged message may still be legitimate. An unflagged message may still be a scam. Sophisticated fraudsters can write in a more natural style, start with harmless conversation, or use a number that later gets saved as a contact. The model also cannot judge identity. It does not know whether the person on the other side is who they claim to be.
It will not replace basic caution. Users should still treat unexpected requests for money, OTP codes, PINs, bank details, or remote-access apps as high risk. WhatsApp will never ask for a PIN or payment details through a random chat. Official support accounts carry a verified badge. Messages that claim to be from WhatsApp Support without that badge should be blocked.
There is also a practical limit: the feature only watches unknown senders. Once a number is saved, Scam Alert steps aside. That makes sense for everyday chats, but it also means a scammer who persuades someone to save the number can reduce the chance of future warnings.
Why the On-Device Approach Matters
For years, privacy advocates have argued that encrypted messengers cannot add meaningful scam detection without weakening encryption. Scam Alert is Meta’s attempt to answer that criticism. By keeping inference on the phone, WhatsApp can point to the same end-to-end encryption promise it has used since 2016.
The approach is closer to WhatsApp’s on-device voice message transcription than to a cloud spam filter. The model has to be small enough to run on ordinary phones, simple enough to review, and accurate enough to catch common scam language without flooding users with false alarms. That is why Meta is starting with a limited beta and asking users who dismiss a warning to optionally share a few messages. The company needs real-world feedback on false positives before a global launch.
The Bigger Picture
WhatsApp has more than three billion users. That scale makes it a prime target. Regulators and consumer agencies have repeatedly flagged messaging apps as a major source of social-engineering losses, including fake investment schemes and account-takeover tricks. Device-linking scams, in which a victim is persuaded to enter a pairing code, have been a particular problem. Scam Alert sits alongside those earlier warnings rather than replacing them.
If the beta goes well, the feature should eventually reach the main Android app and then iOS. Until then, the safest habit remains unchanged. Pause before replying to an unknown number. Do not send money or codes under pressure. Use two-step verification. Keep the app updated. And if Scam Alert appears on your phone, treat the banner as a prompt to think, not as proof that the other person is a criminal or that the chat is safe.
Scam Alert will not end WhatsApp fraud. It can, however, put a visible stop sign in front of some of the most common opening moves. For a platform that cannot read private messages, that is the point: give users more information, keep the conversation encrypted, and let them decide what happens next.
About The Author
