Open three tabs and you might see three copies of the same dispatch. The headlines and mastheads change. The reporting behind them doesn’t.

Abdelkader Metales, Miriam Grabenheinrich and Aicha Metales tested what happens when a model judges those copies. In a controlled experiment with claims about invented people, Llama 3.2 3B accepted 6% of false claims when shown one report and 71% when shown three identical copies under different outlet names. For phi4 14B the figures were 21% and 90%. A note about the shared source did little for those two models. Gemini 2.5 Pro responded much more strongly to the note, but in that test it could also reject true claims about invented people.

The authors then simulated a system that uses judges to decide which claims a learner takes in. Showing one report per detected copy chain lowered the share of copied lies accepted by 10 to 24 percentage points across four judges. The largest average drop in acceptance of true claims was 4.2 points. The text comparison missed the generated rewrites. In the simulation, grouping duplicates also left coordinated claims from separate sources untouched.

This matters when an analyst asks an AI system to confirm an incident or a supplier change. The useful question is what each source saw for itself. Keep the wire credit, filing or interview attached to each claim, and let a reviewer inspect why two accounts were grouped. If all three links lead back to the same interview, the analyst still has one source.