Can AI Crush Online Rumors? Inside Alibaba’s Rumor‑Crushing Machine

Recently the author received a message from his mother urging the spread of a sensational claim that a single Chinese herb could kill 60% of cancer cells within 48 hours, prompting a personal reflection on the pervasive nature of online rumors.

“Urgent! Scientists discover: a single herb can kill 60% of cancer cells in 48 hours!”

The author interviewed Li Quanzhi, a core member of the DAMO Academy NLP team, who holds a Ph.D. in natural language understanding and previously worked as an intelligence officer at Reuters.

Li explained the three‑step AI‑driven rumor detection process:

Step 1 – Source Credibility

The system extracts the original source, analyzes the publisher’s profile (domain, influence, past content, registration time, activity patterns, etc.) and assigns a trust score.

Step 2 – Content Verification

It parses the article’s claims, extracts key entities (e.g., the herb’s ingredients), and matches them against a knowledge graph of verified scientific facts; contradictory or unsupported statements receive penalties.

The model also detects sensational language typical of clickbait, such as exaggerated titles or emotionally charged phrases.

Step 3 – Propagation Path Analysis

The AI builds a network of user interactions (shares, comments, likes), evaluates each participant’s credibility (e.g., verified medical professor vs. casual user), and weighs their influence on the rumor’s spread.

All three layers feed into a neural‑network classifier that outputs a final credibility score for the news item.

Li highlighted that the system learns autonomously, continuously improving as the knowledge base expands, and can perform multi‑task learning to simultaneously assess truthfulness, trace propagation, and profile users.

In the recent SemEval global semantic evaluation, DAMO’s “AI Rumor Crusher” achieved a record‑breaking 81% accuracy in binary fake‑news detection, surpassing previous benchmarks. The same team previously set records in SQuAD (82.44% accuracy) and won multiple categories in the WMT machine‑translation competition.

These advances aim to rebuild public trust by automatically debunking false information and delivering verified corrections to those previously exposed to the rumor.

AI, in this context, is portrayed as a tool for love and truth.