Main analysis / arXiv

A browser extension that scores the bait before you bite

A new paper pairs transformer embeddings with a homemade "baitness" score to warn readers before and after they open an article.

BriefSprout desk · 24 July 2026 · 1 min read · source-backed

⚡ AI Snapshot

  • 91% F1 on open combined dataset
  • Custom baitness score plus transformer embeddings
  • Explains each verdict with visible metrics

What changed

Researchers introduced ClickGuard, a browser extension that detects clickbait and explains why it flagged something. The system blends transformer-based embeddings with hand-built linguistic features and a custom "baitness" score, feeding an XGBoost model that hits a 91% F1-score on an open combined dataset. It warns users both before and after they open an article, and after opening shows a percentage likelihood that the piece is clickbait, with the metrics behind the call.

Why it matters

Most clickbait tools give you a verdict and nothing else, which is easy to ignore and hard to trust. The interesting move here is the explanation layer — showing the metrics behind a score turns "trust me" into "here's why," which is the only version that changes behavior. Whether a 91% F1 on a curated dataset survives contact with the live, adversarial web is the real question, and one benchmarks can't answer.

Why you should care

Researchers
A hybrid baseline with an explainability layer
Developers
Feature-engineering recipe for content filtering
The source trailEcosystem sources

Related briefs