AI also makes irrational mistakes and exhibits biases just like humans.

Despite operating on algorithms and massive datasets, artificial intelligence (AI) systems like ChatGPT are still susceptible to common human cognitive biases, namely overconfidence and bias.

AI performs best when tasked with solving problems that follow clear procedures and fixed formulas. However, when dealing with situations requiring subjective judgment or involving personal preferences, AI can make mistakes similar to humans.

A new study published in the journal Manufacturing & Service Operations Management, conducted by scientists at five universities in Canada and Australia, evaluated the behavior of ChatGPT (specifically, OpenAI's GPT-3.5 and GPT-4 models) on 18 prominent cognitive biases commonly found in human psychology, such as confirmation bias, the possessive effect, the sunk cost fallacy, and the certainty bias.

The results showed that in nearly half of the scenarios tested, ChatGPT behaved exactly like a human when faced with irrational or emotional choices, despite the system's reputation for consistent logical reasoning.

Is AI truly as "impartial" as expected?

Dr. Yang Chen, associate professor of operations management at the Ivey Business School (Canada) and lead author of the study, stated: "AI performs best when tasked with solving problems that follow clear processes and fixed formulas. But when dealing with situations requiring subjective judgment or involving personal preferences, AI can make mistakes just like humans."

The research team incorporated familiar psychological scenarios into ChatGPT, adding practical contexts such as inventory management or supplier negotiations.

It is noteworthy that AI models still exhibit cognitive bias even when the context of the question is shifted from abstract to practical business context.

GPT-4 is smarter, but it's not perfect.

GPT-4 is an upgraded version of GPT-3.5 that excels at solving logic or probability problems. However, in subjective simulations, such as choosing a risky option to increase profit, GPT-4 exhibits a tendency to behave more emotionally than humans.

In particular, the study noted that GPT-4 consistently provides biased responses in confirmation bias testing scenarios and is more prone to the "hot-hand fallacy," a tendency to believe that random events will repeat in a sequence, than GPT-3.5.

Conversely, AI has the potential to avoid some of the biases that humans often make, such as overlooking base-rate neglect or the sunk cost fallacy.

The reason ChatGPT exhibits human-like bias stems from its training data, which is riddled with our own flawed behaviors and thinking patterns. - Image: AI

The origin of AI bias: From human data itself.

According to the research team, the reason ChatGPT exhibits human-like bias stems from the training data itself, which is full of our own flawed behaviors and thinking patterns. The fact that the AI ​​is refined based on human feedback further reinforces this tendency, as the model is "rewarded" for responses that seem reasonable, rather than entirely accurate.

"For accurate and unbiased results, use AI for tasks you already trust a computer to handle well," Dr. Chen recommends. "However, for strategic or emotional problems, human oversight and intervention are still necessary, even if it's simply by rewriting the question."

Co-author Meena Andiappan, associate professor of human resources and management at McMaster University (Canada), argues: "AI should be viewed as an employee with significant decision-making power, meaning it needs to be supervised and adhere to ethical principles. Otherwise, we are inadvertently automating flawed thinking instead of improving it."

Read more Back to the Home page

Back to the topic

MINH HAI

Source: https://tuoitre.vn/ai-cung-mac-sai-lam-phi-ly-va-thien-vi-nhu-con-nguoi-20250505103652783.htm