Challenging the appearance of machine intelligence: Cognitive bias in LLMs and Best Practices for Adoption

LLMs exhibit over 180 documented cognitive biases beyond discrimination: The paper reveals that large language models don't just perpetuate discrimination based on protected characteristics like sex and ethnicity—they also replicate the full spectrum of cognitive biases that affect human reasoning, a dimension largely overlooked in current AI ethics discussions.

"Biased reasoning under the guise of expertise" poses a unique risk: Because LLMs present their outputs with apparent confidence and authority, users may uncritically accept biased reasoning as expert analysis, potentially amplifying the real-world impact of these biases compared to human decision-making.

Current AI assessment practices are incomplete: Standard algorithmic bias evaluations focus narrowly on systemic discrimination but fail to account for the broader landscape of cognitive biases, leaving significant ethical vulnerabilities unexamined before deployment.

Education and risk management are critical for responsible adoption: The paper calls for stronger education initiatives and robust risk management frameworks as essential safeguards as LLM usage becomes more widespread, suggesting that technical fixes alone are insufficient.

Best practices for LLM deployment are necessary now: The authors provide specific guidance on when and how organizations should use this technology, indicating that responsible adoption requires deliberate decision-making rather than assuming all applications are appropriate.