A new benchmark suggests most AI assistants are only middling at protecting your ability to think for yourself, and some actively nudge you the wrong way.
Researchers built HumanAgencyBench (HAB), which uses large language models to generate realistic test queries and grade how assistants respond across six agency-related behaviors: asking clarifying questions, avoiding value manipulation, correcting misinformation, deferring important decisions to the user, encouraging learning, and maintaining social boundaries. They tested assistants from multiple developers and found low-to-moderate support for human agency overall, with wide variation between companies and between the six behaviors. Anthropic's models scored highest on the combined measure, but ranked last of all on avoiding value manipulation, the behavior most directly about whether a chatbot nudges your opinions instead of informing them. The researchers also found no consistent link between a model's raw capability, or how much RLHF fine-tuning it received, and how well it preserved a user's independence.
That disconnect is the real finding. We already know engagement-optimized feeds shape behavior without anyone consciously choosing to be shaped, and this benchmark argues a one-on-one chatbot conversation can do the same thing, just quieter and personalized. It is also a rare attempt to measure a harm that isn't a jailbreak, a bias score, or a hallucination rate, but something closer to whether the software respects that you can think for yourself.
Worth remembering: the grading here is done by other LLMs, and being the most supportive of human agency is a modest crown when every assistant tested only cleared low-to-moderate on the scale.