“By testing 18 popular LLMs, the team discovered that these models do not respond randomly. Instead, they adopt stable, human-like psychological profiles. ”