LLM Problems Observed in Humans (embd.cc)

A recent evaluation of conversational dynamics between humans and AI has revealed alarming parallels in performance, suggesting that as language models (LLMs) evolve, they inadvertently highlight shortcomings in human communication. The author notes that behaviors once attributed to LLMs—like incoherent rambling and failure to grasp context—are increasingly evident in human conversations. This observation raises concerns about the potential erosion of meaningful interpersonal exchanges, as individuals may lean more on AI for coherent perspectives, thereby diminishing their ability to engage in deep, nuanced discussions with peers. The significance of this trend lies in its implications for social interaction and cognitive engagement. While LLMs can quickly learn and adapt during interactions, humans often struggle to maintain focus or apply learning from previous dialogues. The author highlights cognitive phenomena such as “instruction drift” and “mode collapse,” which show that both humans and AI may fail to sustain the original intent of a conversation. This increasing reliance on AI for problem-solving and conversation could lead not only to a degradation of social skills but also challenges in critical thinking and personal connection, potentially pushing society toward an enhanced or even replaced intelligence paradigm.