Inteligența Artificială reflectă prejudecățile umane, conform unui studiu.

Recent research has revealed surprising similarities between human cognitive biases and the decision-making processes of artificial intelligence (AI), particularly ChatGPT. A study published in Manufacturing & Service Operations Management evaluates how AI mirrors common human cognitive errors, analyzing its behavior against 18 known biases from psychology.

The research, conducted by academics from Canada and Australia, specifically tested OpenAI’s GPT-3.5 and GPT-4 models. Findings indicate that, while these models demonstrate impressive consistency in their logic, they are not immune to the same irrationalities that humans exhibit. This consistency has both positive and negative implications for users.

Key Findings

1. Cognitive Biases in AI:
   The study explored how AI exhibits human-like biases, such as risk aversion and overconfidence, through hypothetical questions inspired by psychology and real-world scenarios like inventory management. Despite their design, both AI models displayed a tendency to behave like humans, particularly in the realm of subjective decision-making.

2. Performance Comparison:
   GPT-4 generally outperformed GPT-3.5 in scenarios requiring clear mathematical solutions, committing fewer logical errors. However, it also reflected irrational human preferences in subjective situations, showing an extreme preference for certainty over ambiguity. The authors noted, "GPT-4 shows an even greater preference for certainty than humans do."

3. Stable Behavior Across Contexts:
   The study found that the biases of both models remained consistent regardless of whether questions were posed in abstract terms or practical business contexts. This suggests that these biases are intrinsic to the AI’s reasoning processes rather than resulting from mere memorized examples.

4. Amplification of Human Errors:
   A surprising outcome was the observation that GPT-4 sometimes amplified human errors. For instance, in confirmation bias tests, GPT-4 consistently provided biased responses. The authors attribute these human-like biases to the imperfect data used for training AI, which reflects human errors and tendencies.

Implications for Use

The implications of these findings are significant for managers and decision-makers. Yang Chen, the lead author, emphasizes that AI tools are most beneficial when applied to problems with clear solutions. However, when it comes to subjective decisions or those based on personal preferences, the use of AI becomes more complex and requires careful human oversight.

Chen suggests that if one seeks accurate and unbiased decisions, they should use GPT for areas where they would typically trust a computer. For outcomes relying on subjective factors, human oversight—such as adjusting prompts and correcting known biases—becomes essential.

Recommendations

Experts highlight the need for caution when relying on AI for decision-making, particularly in subjective contexts. Human expertise plays a critical role in refining the AI’s outputs and ensuring more rational decisions are made.

Overall, as AI continues to evolve and integrate into decision-making processes, understanding its limitations and the potential for cognitive biases, similar to humans, will be crucial for maximizing its effectiveness while mitigating risks.