Overview
- WhatsApp began a limited beta of Scam Alert on August 12, deploying an optional on-device machine‑learning model that flags likely scam messages and shows private warnings inside chats.
- All classification happens on the user’s phone so no message content is sent to WhatsApp or Meta for analysis or automatic reporting.
- When the model flags a message from a non‑contact, users see a warning and can block, report, ignore, or mark the chat as trusted, and can opt in to share the last five messages to help improve the model.
- Meta requires every model release to be logged on a third‑party append‑only ledger, verifies Cloudflare‑signed manifests and uses Trusted Execution Environments plus differential privacy to collect only two aggregate telemetry counts so researchers can audit provenance without exposing chats.
- The feature is being tested with researchers and an expanded bug bounty, is presented as an encryption‑preserving alternative to network‑level spam rules that regulators in India have proposed, and Meta has given no date for a wider public rollout.