Meta's anti-scam push moves detection to the moment before a scam happens

Warnings placed at the decision point, not just the account

Meta's existing pitch has been that its systems find and remove malicious accounts. The new tools acknowledge a gap in that model: scammers create accounts that don't behave maliciously right away, so a takedown-only approach misses them until damage is done.

The three consumer-facing features all target the moment a user is about to engage. WhatsApp will now alert you when behavioral signals suggest a device-linking request is suspicious — the specific attack where a scammer coaxes you into sharing a linking code or scanning a QR code to attach their device to your account. The alert shows where the request originates and prompts you to pause.

On Facebook, warnings on friend requests key off signals like few mutual friends or a profile country that differs from expectations. On Messenger, expanding to more countries this month, a chat with a new contact that matches scam patterns — such as suspicious job offers — triggers a prompt asking whether you'd like to share recent messages for an AI scam review.

That Messenger flow is notable because it makes the AI review opt-in on a per-conversation basis rather than a silent background scan, which matters for a private-messaging surface.

Multi-signal AI aimed at impersonation and lookalike domains

Meta frames its AI investment around scam types that evade traditional detection because they rely on framing and context rather than obvious rule violations. The stated approach analyzes text, images, and surrounding context together.

For celebrity and brand impersonation, that means examining fake fan sentiment, misleading bios, and claimed associations with public figures — the kind of contextual reasoning that single-signal classifiers handle poorly. For deceptive links, Meta says it proactively detects pages built to mimic legitimate ones, protecting thousands of brands against domain impersonation.

The claim is that combining signals catches a broader range of patterns at higher precision. Meta doesn't publish precision or recall figures for these systems, so the specific gains remain asserted rather than measured in the announcement.

Verification tied to a revenue percentage rather than a rule

The most concrete commitment is financial, not technical. Meta says it will expand advertiser verification so that verified advertisers drive 90% of its ad revenue by the end of 2026, up from 70% today.

Framing the goal as a share of revenue rather than a share of advertisers is a deliberate choice. It concentrates verification on the highest-risk, highest-spend categories while leaving the remaining 10% for low-risk businesses — Meta's example is a local ice cream shop. The measure captures where the money and the abuse risk concentrate, not the raw count of accounts.

The 159 million scam ads removed globally in 2025, with 92% taken down before any report, sits behind this shift. Verification is being positioned as a preventive layer on the ad system rather than a reactive cleanup of ads already running.

The enforcement numbers point to industrialized scam networks

Meta's own framing is that scams are being industrialized. The figures it cites are meant to show scale on both sides: 10.9 million accounts on Facebook and Instagram tied to criminal scam centers taken down last year, and over 150,000 accounts disabled in a joint operation with global law enforcement against Southeast Asian scam syndicates.

That operation targeted specific playbooks — impersonating law enforcement to stage fake 'digital arrests' over video calls, and pushing fraudulent crypto investments. In India, Meta banned more than 12.1 million pieces of ad content in 2025, with over 93% removed proactively, alongside the third edition of its Scam se Bacho awareness campaign with I4C and SEBI.

What the combination signals for platform-scale fraud defense

The specific implication of this announcement is that Meta is treating scam defense as a stack rather than a single control: pre-action user warnings, opt-in AI review of private chats, multi-signal impersonation detection, revenue-weighted advertiser verification, and law-enforcement takedowns.

Each layer covers a failure mode of the others. Account removal misses dormant scammers, so warnings intercept at the point of contact. AI classifiers miss contextual fraud, so verification gates ad spend. The revenue target — 90% by end of 2026 — is the one line here that can be checked later, which makes it the clearest test of whether the strategy delivers.