How human-like is LLM cognition?

Large language models (LLMs) have led to renewed debate about whether artificial systems might possess human-like cognition. Although empirical evidence is growing quickly, it remains difficult to synthesize because studies vary in terms of which kinds of evidence they compare across humans and LLMs—design specifications, internal states, or behavior—and which kinds of inferences they draw. We address this challenge by organizing the Human-LLM comparison literature into broad evaluation categories and prototypical patterns of findings that either support or challenge LLM human-likeness. The resulting picture is one of jagged alignment, with human-likeness varying across dimensions rather than holding globally. We argue that studying these similarities and differences can clarify not only the functioning of artificial systems, but also the conceptual foundations of cognitive science.