Source: Cao, Wang & Yi (2025) — “When LLMs Go Abroad: Foreign Bias in AI Financial Predictions”
This is one of those risks that sounds abstract until you imagine explaining it after the fact. Then it suddenly becomes very concrete, very expensive, and very difficult to hide behind a slide deck.
It’s well-established that human investors exhibit home bias — they tend to hold too much of their portfolio in domestic stocks, partly because they feel they understand local companies better. This paper documents a curious reversal of that phenomenon in AI financial models. U.S.-based ChatGPT, when asked to predict the stock performance of Chinese firms, is systematically more optimistic than China-based DeepSeek making the same predictions. And critically, ChatGPT is less accurate. model overestimates Chinese firm performance; the Chinese model is closer to reality. firms, both models produce similar forecasts. The mechanism the researchers identify is information availability. ChatGPT’s training data contains abundant U.S. financial news and commentary but comparatively sparse coverage of Chinese firms, especially those that don’t cross-list on U.S. media coverage of a Chinese firm is limited, ChatGPT’s optimism is strongest. When the researchers inject Chinese-language news into the prompt, the prediction gap narrows substantially. The bias appears to be informational, not architectural. This is a finding worth sitting with for a moment. An AI model trained primarily on one country’s information environment will inherit that environment’s blind spots — not as deliberate policy, but as a structural feature of its data diet.
My takeaway: the danger is rarely the dramatic thing in the headline. It is the quiet gap between knowing a risk exists and assigning someone to do something about it. Very unglamorous. Very important.