WhatsApp Scam Alert Uses On-Device AI to Detect Potential Scams

WhatsApp is testing WhatsApp Scam Alert, an optional security feature that uses artificial intelligence to identify conversations that may involve scams. The feature is currently in limited Beta and is designed to analyse messages directly on a user’s device rather than sending message content to WhatsApp’s servers for classification.

The approach aims to address a difficult challenge for encrypted messaging platforms: improving scam detection without routinely accessing users’ private conversations.

How WhatsApp Scam Alert Works

When enabled, WhatsApp Scam Alert uses a machine-learning model installed on the user’s device. The model examines messages from people who are not saved as contacts and looks for conversational and linguistic patterns associated with potential scams.

If the system determines that a conversation may be suspicious, WhatsApp can display a warning to the recipient.

Users remain in control of what happens next. They can block the sender, report the conversation, continue chatting or mark the conversation as trusted.

The system does not automatically report a suspected scam to WhatsApp. Reporting requires an action from the user.

Message Content Stays on the Device

One of the most important elements of WhatsApp Scam Alert is its on-device processing.

WhatsApp says message content used for classification remains on the user’s device. This means the company does not need to receive the underlying private messages simply to determine whether a conversation appears suspicious.

However, it would be inaccurate to say that no information leaves the device. WhatsApp has developed a separate privacy-preserving analytics system to measure the feature’s overall performance.

The analytics system works with aggregate information rather than sending the underlying conversations to WhatsApp.

Users Can Voluntarily Share Messages

WhatsApp also provides an optional way for users to help improve Scam Alert.

After marking a conversation as trusted, users can choose to share the last five received messages with WhatsApp. This requires an explicit action from the user and is separate from the normal on-device classification process.

This distinction is important because Scam Alert does not automatically send suspicious conversations to WhatsApp for review.

Privacy-Preserving Analytics

To understand how the system performs, WhatsApp uses several privacy technologies.

These include Oblivious HTTP (OHTTP), anonymous credentials, confidential computing, secure aggregation and differential privacy.

The process is designed to allow WhatsApp to measure aggregate trends, such as warning activity and user responses, without exposing individual users’ private message content.

WhatsApp also considers different security threats, including external attackers, compromised infrastructure and malicious insiders.

Transparency Around AI Models

WhatsApp is also creating mechanisms to make the AI models used by Scam Alert more verifiable.

Model versions and associated assets are recorded in a public, append-only transparency ledger. Cryptographic hashes can then be used to verify that the model files delivered to devices correspond with the versions WhatsApp has approved.

The company says model information will also be made available for independent security research, while its bug-bounty programme provides another way for researchers to identify potential vulnerabilities.

WhatsApp Prevents Targeted Model Assignment

Another important part of the system is how AI models are delivered.

WhatsApp says its infrastructure is designed so that it cannot simply select a particular Scam Alert model for an individual user.

Instead, model assignment takes place on the client side. This is particularly relevant when WhatsApp tests different experimental versions of its AI system.

The design aims to prevent the server from secretly directing a particular model variant toward a specific individual while making model deployment more transparent and auditable.

Users Can Check Scam Alert Activity

WhatsApp is also providing users with visibility into how Scam Alert operates.

The feature’s activity information can show details such as which messages were analysed, the resulting classification and the model version involved.

This provides users with greater transparency about the activity taking place on their device.

Security Testing and Beta Limitations

WhatsApp Scam Alert remains in limited Beta, so its effectiveness should not be treated as a guarantee that every scam will be detected.

Machine-learning systems can produce false positives and false negatives, while scammers can also change their tactics to avoid detection.

WhatsApp is therefore continuing security testing and working with researchers through its bug-bounty programme as it evaluates the system.

Why On-Device AI Matters for Scam Detection

Scams increasingly rely on social engineering, impersonation and carefully constructed conversations rather than obvious spam messages.

Advances in mobile machine learning have made it more practical to run sophisticated classification models directly on smartphones.

For WhatsApp, this creates a potential way to strengthen scam protection while maintaining its approach to private messaging.

The technology could also become increasingly important as messaging platforms face pressure to detect malicious activity without weakening encryption or routinely examining private communications on central servers.

WhatsApp Scam Alert: Key Takeaways

• WhatsApp Scam Alert is an optional feature currently in limited Beta.
• Scam detection uses an AI model running on the user’s device.
• Messages from non-contacts can be analysed for potential scam patterns.
• Message content used for classification remains on the device.
• Users decide whether to block, report, continue or trust a conversation.
• Users can voluntarily share the last five received messages.
• Privacy-preserving analytics use technologies such as OHTTP and differential privacy.
• WhatsApp is developing a public transparency system for its AI models.
• Model assignment is designed to prevent targeted delivery to individual users.
• Security researchers can test the system through WhatsApp’s security programmes.

Conclusion

WhatsApp Scam Alert represents an attempt to combine AI-powered scam detection with the privacy requirements of encrypted messaging.

Rather than sending private message content to a server for routine classification, WhatsApp is using an on-device AI model to identify potentially suspicious conversations. Around that model, the company has built additional safeguards involving privacy-preserving analytics, confidential computing, model transparency and protections against targeted model assignment.

The feature is still in limited Beta, meaning its real-world effectiveness remains to be evaluated. Nevertheless, WhatsApp Scam Alert demonstrates how on-device AI could provide an additional layer of protection against increasingly sophisticated scams without requiring routine server-side access to users’ private conversations.