ChatGPT and Perception Biases in Investments: An Experimental Study

Applications of artificial intelligence (AI) in finance have been met with concerns about algorithmic bias, following issues observed in domains such as medical treatment and lending. We ask whether AI models accurately capture investment preferences across demographics. We elicit investment preferences from over 1,200 survey participants and compare the data directly to investment ratings generated by OpenAI's ChatGPT (GPT4). We find that ChatGPT predicts investment preferences with high accuracy across demographics. Specifically, ChatGPT correctly predicts that women rate stocks lower than men, older individuals prefer holding cash, and higher incomes are associated with higher ratings for stocks and bonds. Moreover, free-form responses from ChatGPT focus on the same aspects as human free-form responses. Most common themes in both responses are "risk" and "return," and "knowledge" and "experience" play an important role for stock market participation. One difference is that ChatGPT responses are almost always transitive, whereas human responses are more prone to violating transitivity, especially when expressing indifference. Overall, the use of AI in finance offers a promising direction for augmenting human surveys in preference elicitation, with important applications for areas such as robo-advsing.