Humans frequently make mistakes in both new and routine tasks, ranging from minor errors to catastrophic ones that can erode trust and potentially have life-or-death consequences. Over time, we have developed security systems to mitigate human errors, such as rotating casino dealers and taking precautions during surgeries. These systems rely on the predictability of human mistakes, which often occur at the boundaries of knowledge or due to factors like fatigue. In contrast, artificial intelligence (AI), specifically large language models (LLMs), are being integrated into society, presenting a different error profile. AI mistakes are unpredictable and can occur randomly, without clustering around specific topics. LLMs might make mistakes that are bizarre, like suggesting unlikely scenarios. Unlike humans, AI systems exhibit confidence in both correct and incorrect outputs, creating trust issues in complex tasks. To address these AI-specific challenges, research is focusing on two areas: engineering LLMs to make more human-like mistakes and developing new systems to address the unique errors of AI.

Approaches like reinforcement learning with human feedback are being used to align AI behavior with human understanding. Existing human-error prevention methods, such as double-checking work, can be applied to AI, but more innovative solutions are needed. Unlike humans, AI can handle repetitive questioning, and asking the same question in different ways can be a strategy to reduce errors. There are also surprising similarities between AI and human error, like the prompt sensitivity issue in LLMs, where slight changes in phrasing yield different responses, similar to human survey biases. AI also demonstrates quirks like repeating familiar terms due to bias. Some intriguing tactics to manipulate AI systems, such as using ASCII art to bypass restrictions, highlight both AI's unique vulnerabilities and potential parallels to human behavior. Ultimately, while humans rarely make random and erratic mistakes, AI systems should be constrained to decision-making tasks that align with their capabilities, considering their distinctive error patterns.