As language models move from executing fixed instructions to making open-ended decisions, a key question is whether they choose goals the way people do. A new arXiv paper addresses this directly, comparing LLM goal selection with human goal selection in a self-directed learning task.
The authors argue that LLMs are increasingly used in agentic workflows, social research, and chat settings where they effectively replace humans in deciding what to pursue. Their results suggest that this substitution is not seamless: model choices diverge from human choices in the task studied.
The paper does not claim models are better or worse, only that the behaviour differs. That distinction matters for anyone relying on LLMs to set their own objectives, since differences in goal selection could cascade into downstream behaviour.