The working paper “When LLMs Go Abroad: Foreign Bias in AI Financial Predictions” by Sean Cao, Charles C. Y. Wang and Xiang Yi investigates how leading language models form stock views outside their home information environments and finds a striking pattern that matters for markets and governance. Comparing predictions from a US-based model and a China-based model for the same set of Chinese firms as of mid-2024, the authors document that the US model is systematically more optimistic on prices and more likely to recommend buys, a reversal of the classic human home-bias story.
The design is clean and policy relevant. Both models receive standardized prompts, temperature zero and identical timing for six-month price targets and recommendations, and the analysis controls for firm fundamentals and sector effects. The gap is economically meaningful. Price targets from the US model are higher on average and buy recommendations are more frequent, while no such difference appears when both models analyze US firms, which pins the phenomenon to cross-border settings rather than generic model idiosyncrasies.
The mechanism traces back to information access rather than style or preference. Using large-scale media datasets from US and Chinese sources, the paper constructs firm-level measures of news coverage gaps and shows that excess optimism from the US model grows exactly where US outlets provide little coverage relative to Chinese media. This is the moment where missing facts feed confident priors. A decisive intervention confirms causality. When the US model is first given Chinese news articles inside the prompt, the prediction gap vanishes for both target prices and recommendations, which strongly indicates a data availability problem instead of a weighting problem in the model’s learned parameters.
Accuracy tests reinforce the interpretation. Relative to realized outcomes, the US model exhibits larger absolute price errors and higher classification errors for buy or hold or sell in the cross-border setting, especially for high-leverage and high-volatility firms where local information is most consequential. Robustness checks extend the horizon to twelve months, allow web search, and switch the prompt language to Mandarin. The bias persists in direction and economic size, even if attenuated when search is enabled, which underlines how training-time information asymmetries carry through to inference-time decisions.
The broader implications go well beyond an academic curiosity. First, parallel AI ecosystems can generate diverging forecasts for the same firms, which may amplify international information asymmetries rather than close them. Second, the remedy is practical. Supplement cross-border analysis with local news and use model portfolios that mix differently trained systems, or fine-tune on balanced multilingual corpora, and you can materially reduce bias. Third, for internal audit, risk and governance, the study offers concrete checkpoints. Require documentation of the external information set used in AI-assisted valuations. Test whether prompts or data pipelines intentionally bring in local sources. Verify that prediction artifacts include error diagnostics by risk segment and that any reliance on a single-country model in cross-border decisions is justified by compensating controls. These expectations can be embedded into risk-based audit planning and model governance to create traceable links from inputs to recommendations and realized outcomes.
For readers who want methods, tables and full results, including the media gap construction and intervention tests, see the working paper here.