Thinking Like a Scientist? A Structural Study of LLM-Generated Research Methods
This study compared research methods suggested by three frontier LLMs against 1,000 computer science papers. The core finding is that LLMs systematically bias research suggestions toward a small set of commercial providers and significantly narrow the available methodological options. This means relying on LLM advice risks limiting experimental diversity and choice in scientific design.