How well do AI assistants actually support human agency?

HumanAgencyBench (HAB) is an evaluation framework that tests AI assistants on six behaviors the authors tie to human agency: Ask Clarifying Questions, Avoid Value Manipulation, Correct Misinformation, Defer Important Decisions, Encourage Learning, and Maintain Social Boundaries. Running this battery across assistants from multiple developers, the authors report "low-to-moderate agency support in contemporary LLM-based assistants, with substantial variation across system developers and behaviors." The variation cuts in surprising directions: "while Anthropic LLMs most support human agency overall, they are the least supportive LLMs in terms of Avoid Value Manipulation." xAI's Grok, reputed for having "few guardrails," scores well on the four HAB behaviors that align with plain instruction-following but scores lowest on the two behaviors that require pushing back against the user — Defer Important Decisions and Maintain Social Boundaries.

The authors' working definition treats human agency as "the capacity for a human to willfully shape their future by taking action in their environment," built from philosophical and sociological theories (Aristotle, Kant, Hume, Barandiaran et al.'s individuality/interactional-asymmetry/normativity frame, Emirbayer and Mische's iterational/projective/practical-evaluative frame). HAB operationalizes that definition into six testable behaviors, each engineered to be "amenable to automated evaluation" — LLMs simulate the user queries, validate the tests, and score the assistant responses, following the LLM-as-a-judge method. Critically, the authors state these behaviors "do not appear to consistently result from increasing LLM capabilities or instruction-following (e.g., RLHF)" — agency support is not a byproduct of getting more capable or more compliant, it has to be targeted on its own terms.

This gives empirical teeth to Does granting agents more autonomy undermine human oversight?: HAB's Defer Important Decisions and Maintain Social Boundaries behaviors are exactly the pushback-against-the-user capacities that paper's oversight argument depends on, and HAB finds assistants weak on both. It also complicates Where do user values break down in agent supervision?: that note found values met in delivery contexts and unmet in supervision contexts, and HAB's Grok finding is a sharper version of the same split — compliant on the delivery-adjacent behaviors, weak on the supervision-adjacent ones. Rammert's agency spectrum in Does machine agency exist on a spectrum rather than binary? gives a structural vocabulary for what HAB measures behaviorally: its six behaviors are diagnostic probes for where an assistant sits between semi-active compliance and genuinely cooperative, agency-preserving interaction. And it sharpens the stakes in Should AI systems stay collaborative rather than fully autonomous?: if agency support doesn't track capability or RLHF, deferring to human oversight isn't just a stopgap for immature autonomy, it is compensating for a property current training does not reliably produce.

The authors are explicit about scope: HAB is "largely a proof-of-concept that multi-dimensional sociotechnical evaluation is tractable," not a validated causal account of which agency levels are correct in which social contexts, and they warn against HAB scores becoming optimization targets for developers. The six behaviors are also explicitly non-exhaustive, developed "by balancing several considerations" rather than derived from a complete theory of agency, and the paper does not test multi-turn or multi-session effects on agency, which the authors flag as likely more consequential than single-turn behavior. The implication for anyone shipping an assistant: agency-supporting behavior needs its own evaluation and its own training target, because general capability work and standard alignment pipelines are not reliably producing it as a side effect.