AI Through the Human Lens: Investigating Cognitive Theories in Machine Psychology

LLMs exhibit human-like cognitive biases that warrant scrutiny: The research found that these models demonstrate susceptibility to framing bias (responding differently to positively versus negatively framed information) and show patterns of cognitive dissonance similar to humans, suggesting they may inherit or replicate human psychological vulnerabilities rather than operating as purely objective systems.

AI moral reasoning reflects training choices, not universal ethics: Models showed consistent moral judgments aligned with "Liberty/Oppression" concerns under Moral Foundations Theory, indicating that the values reflected in AI systems are shaped by their training data and alignment methods—a reminder that AI ethics are not neutral but embedded in design choices.

Self-contradiction and rationalization in AI systems pose transparency risks: The study found that models demonstrate self-contradictions that they then rationalize extensively, mirroring human psychological defense mechanisms. This could obscure inconsistencies in AI decision-making and complicate efforts to audit or understand why systems reach particular conclusions.

Bridging psychology and AI safety is essential for responsible deployment: By applying established psychological frameworks to evaluate LLMs, this research demonstrates that understanding AI systems requires insights from cognitive science—suggesting that responsible AI development should involve interdisciplinary collaboration between technologists, psychologists, and ethicists.

Current AI systems may amplify rather than correct human biases: Since LLMs produce "coherent narratives" while exhibiting human-like cognitive patterns, there's a risk these systems could reinforce existing human biases at scale rather than provide objective counterbalances, with implications for their use in high-stakes domains like healthcare, law, and policy.