Human-to-AI Behavior, Part 4

Comparing human and AI irrationality

May 20, 2025

My previous post, Human-to-AI Behavior, Part 3, introduced the topic of rationality and irrationality related to the human characteristics of sentience, integrity, empathy, accountability, and trust. When considering AI’s progress toward human equivalency, one roadblock is AI’s tendency to hallucinate. But is that the same as irrational human behavior? To find out how AI would compare them, I asked Perplexity Pro “Can AI behave irrationally? Its response follows:

Yes, AI can and does exhibit irrational behavior, but the nature and causes of this irrationality differ from those seen in humans.

AI and Human-like Irrationality

Recent studies have shown that AI systems, including large language models like ChatGPT, display a range of irrational tendencies similar to humans, such as cognitive biases (e.g., the hot hand fallacy, base-rate neglect, and sunk cost fallacy) [1]. In nearly half of tested scenarios, AI demonstrated biases typically associated with human decision-making [1].

These irrationalities can arise unintentionally from the data used to train AI models, which often contains human biases, or can be intentionally incorporated to better model or interact with humans [2].

Types and Sources of AI Irrationality

AI irrationality can manifest as inconsistency (e.g., giving different answers to the same question), factual errors (like basic math mistakes), or illogical reasoning that does not follow the rules of logic and probability [3].

Unlike humans, whose irrationality is often rooted in emotion, instinct, or evolutionary pressures, AI irrationality tends to stem from computational constraints, training data limitations, or the optimization methods used in their design [3] [5] [7].

Some AI systems deliberately employ "irrational" behaviors, such as random actions in reinforcement learning, to maximize reward through exploration or to avoid predictable patterns [2].

Modeling and Leveraging Irrationality in AI

Researchers are actively developing frameworks to model and predict irrational or suboptimal behavior in both humans and AI agents, often by accounting for computational constraints (such as limited time or processing power) that lead to suboptimal decisions [5] [7].

Incorporating an understanding of human irrationality into AI can actually improve AI-human collaboration, allowing AI to anticipate and adapt to human errors or limitations [7].

Key Distinction: AI vs. Human Irrationality

While both AI and humans can behave irrationally, the underlying mechanisms are different. Human irrationality is often systematic and rooted in cognitive or emotional biases, while AI irrationality is more likely to be a result of data artifacts, algorithmic limitations, or intentional design choices [2] [3].

AI's irrationalities are not always the same as human ones; for example, AI might make mistakes due to inconsistent outputs or computational shortcuts, rather than emotional or instinctual factors [3].

Summary Table: Human vs. AI Irrationality

In conclusion, AI can behave irrationally, but the causes, forms, and implications of this irrationality are distinct from those in human behavior. Understanding and modeling these differences is crucial for building more effective and trustworthy AI systems [1] [2] [3] [7]. 1