ClickGuard: Detecting and Spoiling Clickbait News with Informativeness Measures and Large Language Models

Summary

ClickGuard is a browser extension that flags clickbait articles using a hybrid pipeline that feeds transformer/LLM embeddings, handcrafted linguistic features, and a custom 'clickbaitiness' score into an XGBoost classifier. It reports a 91% F1 score on an unspecified combined dataset. It also generates a one-to-two-sentence 'spoiler' summary after the user opens the article.

Why it matters

Misleading headlines remain a persistent driver of low-quality information consumption. A lightweight, client-side detector can shift the burden from platform curation to individual users. The 'spoiler' approach directly targets the incentive structure that rewards clickbait by presenting the article's essence in two sentences. If the tool works reliably in real browser environments, it offers a practical, privacy-preserving alternative to opaque platform-level content filters.

Highlights

• XGBoost combined with LLM/transformer embeddings and a custom 'clickbaitiness' metric achieves a 91% F1 score on a combined open dataset
• The extension warns users before and after clicking the link; it explains its prediction through underlying feature scores
• Built-in 'clickbait spoilers' compress the entire article into a 1-2 sentence summary to break the bait effect

Critical take

The 91% F1 score is reported on an unspecified 'combined open dataset' without a breakdown of metrics that would determine if the extension actually works in a live browser, such as false positive rates, domain diversity, or real-time latency. The contribution reads like a competent engineering integration of off-the-shelf components like XGBoost and LLM embeddings, rather than a methodological advancement. While the 'spoiler' feature is clever, it is essentially just the short summarization task that any general-purpose LLM can already perform without a specialized pipeline.