A Harvard Business School study by Charles C.Y. Wang pitted ChatGPT against DeepSeek on stock analysis, giving both models roughly 5,000 publicly traded Chinese companies to evaluate. ChatGPT projected stock prices 12.5% higher and issued "buy" recommendations more often — but its forecast errors were 13% larger than DeepSeek's, according to Harvard Business Review.
The researchers expected each model to show "home bias," favoring its own country's companies. They found the opposite. ChatGPT was more bullish on Chinese companies than DeepSeek was — a pattern the researchers call "foreign bias."
The explanation: Chinese media had been producing negative coverage of these companies, but ChatGPT's training data didn't include much of it. With gaps in its information, the model defaulted to optimism. When the researchers fed ChatGPT Chinese-sourced news and reran the analysis, the excess optimism disappeared. "We show that it's not just missing news. It's missing negative Chinese news that drives this bias," Wang said.
DeepSeek didn't show the same foreign bias when evaluating U.S. companies, because American financial coverage is, as the study put it, "a global commodity" — available in every training dataset.
The same biases showed up in ChatGPT o1, and 40% of institutional investors now use AI for market analysis.
Previously:
