How it works
- 01
Extract
For a URL, the article title, authors, date and body text are pulled from the page. Pasted text skips this step.
- 02
Score
Four fine-tuned transformer models from the Hugging Face Hub each estimate the probability that the text is fake.
- 03
Average
The four probabilities are averaged. Above 50% reads as fake; confidence is the distance from that line.
- 04
Cross-check
In parallel, the domain is looked up in two lists of known fake-news sources and scanned with VirusTotal.