Study Reveals ChatGPT Exhibits Many Human Decision-Making Biases

A recent study published in the journal Manufacturing & Service Operations Management has found that artificial intelligence systems like ChatGPT display many of the same cognitive biases as humans. Researchers from institutions in Canada and Australia tested OpenAI’s GPT-3.5 and GPT-4 models across 18 well-known human biases, including overconfidence, risk aversion, and the endowment effect. While GPT-4 showed improved performance on clear, mathematical problems, it still exhibited irrational decision-making tendencies in subjective scenarios, often favoring certainty more than humans do. The study revealed that these biases are not merely memorized responses but are embedded in the AI’s reasoning processes, likely stemming from biased training data and human feedback during fine-tuning. Some biases, such as confirmation bias and the hot-hand fallacy, were even amplified in GPT-4. However, ChatGPT avoided others like base-rate neglect and the sunk-cost fallacy. The findings suggest that AI tools are most reliable when used for formulaic, objective tasks but require human oversight for subjective or strategic decisions. Experts emphasize treating AI like an employee who needs ethical guidelines and supervision to prevent the automation of flawed human thinking rather than its improvement.