Researchers test duplicate filtering to reduce false claims accepted by AI judges

Researchers report in an arXiv preprint that filtering copied news reports reduced false claims accepted by AI judges in a simulated continual-learning stream. Showing judges one report per detected copy chain cut acceptance of copied lies by 0.10-0.24 across four judges, at a mean cost of no more than 4.2 percentage points of accepted truths. The study examined 200,000 English CC-News articles and linked 18.9% to a near-copy from another outlet. Its initial detector missed all tested LLM rewrites. A separate test using named entities reduced acceptance of reworded-copy lies from 0.67 to 0.20, or 0.40 when rewrites invented details. Colluding sources remain an open problem.