Researchers investigated whether the ways large language models (LLMs) process information are similar to how humans do. They analyzed responses from both humans and six LLMs on two types of assessments: quantitative reasoning and chemistry. The results show that while humans' underlying cognitive constructs can be understood by experts, the same cannot be said for LLMs. In fact, the mechanisms behind LLMs' performance are often 'statistically opaque', meaning they don't follow patterns or logic that humans can easily understand. This finding challenges the assumption that AI and human cognition share similar principles. --- Why it matters: This study's findings have implications for how we evaluate and improve large language models. If LLMs operate on distinct mechanisms, it may be more difficult to develop effective training methods or fine-tune their performance. Researchers in AI need to understand these differences to create more reliable and transparent AI systems. Source: https://arxiv.org/abs/2608.17810