You Forecast What You Read: Impact of the National Origin of News on an AI Forecaster

We ask whether the national media ecosystem an AI forecaster draws its evidence from

changes the probability it reports and its accuracy. Retrieval-augmented forecasters read the

press before they answer, and two countries in dispute tell different stories. We give a frontier

forecasting agent (Claude Opus 4.8) the same 20 resolved US–China questions with US-only,

Chinese-only or mixed sources, against a random-split control. We find the forecaster robust to

the origin of its news: one country’s press leaves its probabilities and accuracy where the other’s

would, and bound a source effect the test could have missed at 5 probability points. The one

asymmetry lies in the agent’s awareness of where its news comes from, which differs between

the two countries, consistent with a model’s own origin setting what it expects to read.