On Characterizations of Large Language Models and Creativity Evaluation

Incredible as they may be, Large Language Models (LLMs) have their limitations. While they generate high-quality texts, excel at stylistic reproduction, and tap into an immense pool of information, they can produce wildly inaccurate responses. The hype around LLMs led to them being characterized as "reasoning", "sentient", or "knowing" like humans. We examine these characterizations and discuss what LLMs can't do and what they are surprisingly good at. LLMs are still susceptible to traditional issues with AI, probabilities are not knowledge, and they are not in the world. Nonetheless, LLMs, despite not being human, have great potential to perform various creative tasks. We conclude that LLMs are beyond "mere generation" and perceivable as creative, but we may need to reassess some frameworks for creativity evaluation.