You Forecast What You Read: Impact of the National Origin of News on an AI Forecaster
As AI systems start to be used for geopolitical analysis and decision-making, I have been interested in how the source of their input shapes their output on such topics. Their input is news written for human readers, who are known to be sensitive to cognitive bias; on top of that, different media outlets do not frame the same facts the same way and do not cover the same topics.
I built a dataset to find out: 20 resolved forecasting questions on US–China relations, each with two equal-size article packs, one from US outlets and government sources, one from mainland-Chinese outlets and state organs. I then ran a ReAct-style frontier AI forecaster on each pack.
What I found:
- Swapping the US pack for the Chinese pack moves the forecast about as much as swapping two random halves of the same articles. Accuracy is the same on every pack, and doubling the evidence does not improve it.
- One asymmetry: in a few runs the agent (Claude Opus 4.8) is surprised that all its sources are Chinese; it never remarks that all its sources are American.
This suggests that frontier models of 2026 already understand the stakes of well-known geopolitical situations from their training alone, and are fairly robust to source manipulation at inference time. It also suggests that Chinese and US models might differ in judgement through their training alone, independently of the news they are given at inference.