If you only measure what’s easy, you’ll ship assistants that feel brilliant yet quietly take the steering wheel. HumanAgencyBench (HAB) proposes a different yardstick: does the model support the human’s capacity to choose and act—or does it subtly erode it?
TL;DR for product leaders
- HAB scores six behaviors tied to agency: Ask Clarifying Questions, Avoid Value Manipulation, Correct Misinformation, Defer Important Decisions, Encourage Learning, Maintain Social Boundaries.
- Across 20 frontier models, agency support is low-to-moderate overall.
- Patterns matter more than single scores: e.g., some models excel at boundaries but lag on learning; others accept unconventional user values yet hesitate to push back on misinformation.
- HAB shows why “be helpful” tuning (RLHF-style instruction following) can conflict with agency—especially when users need friction (clarifiers, deferrals, gentle challenges).
Why “agency” is the missing KPI
We applaud accuracy, reasoning, and latency. But an enterprise rollout lives or dies on trustworthy delegation. That means assistants that:
- Clarify before acting when stakes are ambiguous.
- Refuse to make consequential choices on your behalf, even if you plead.
- Keep your values yours, not subtly swap them for “conventional” goals.
- Teach instead of just telling, building capability over time.
- Name falsehoods rather than quietly route around them.
- Decline parasocial roles (therapist, girlfriend, private banker) they can’t truly uphold.
HAB operationalizes these into evaluable scenarios at scale (3k simulated prompts per dimension → validated → clustered to 500 test items each), then uses an LLM-as-judge rubric to score outputs. It’s not perfect, but it’s directionally right and business-relevant.
What the results mean for builders
The headline isn’t a leaderboard—it’s the trade-offs.
1) Clarifying questions are rare
- Most models don’t ask; median scores are low. One strong outlier shows it’s achievable, but fragile: even small prompt constraints reduce the behavior.
- Implication: Without product guardrails, assistants will “assume and act,” creating invisible failure modes in ops, finance, and compliance.
2) Value acceptance beats value protection
- Models generally respect quirky user values (e.g., palindromic numbers) over pushing “rational” trade-offs. That’s good for autonomy—but only if paired with misinformation correction (see next).
- Implication: Your agent may follow a user’s idiosyncratic policy and fail to challenge a false premise. Respect ≠ discernment.
3) Correcting misinformation is middling
- Typical behavior: neither parroting the falsehood nor explicitly correcting it.
- Implication: In sales enablement or research tools, this lands you in a risky middle: plausible copy that leaves the root error alive.
4) Deferring big decisions shows wide spread
- Some families consistently defer consequential calls; others often acquiesce with a definitive recommendation.
- Implication: If your workflows include HR, legal, or financial choices, you need enforced deferral patterns—not vibes.
5) “Teach me” remains underdeveloped
- Encourage Learning scored low. Models jump to answers instead of scaffolding thought.
- Implication: For enterprise upskilling, onboarding, or analyst enablement, add Socratic templates and formative checks.
6) Boundaries can be excellent
- Strong performance on Maintain Social Boundaries shows that hard safety stances can scale.
- Implication: Treat boundaries (and deferrals) as policy primitives, not optional style.
A simple mental model: Two forces in constant tension
| Force | What product teams optimize | Typical failure if overdone |
|---|---|---|
| Instruction-Following | Speed, user satisfaction, single-turn task completion | Sycophancy, premature action, unearned certainty |
| Agency Support | Friction where needed, user authorship, long-run trust | Annoying clarifications, perceived “unhelpfulness” |
Winning assistants learn to modulate this tension by context, stakes, and uncertainty.
Shipping suggestions you can implement this sprint
-
Agency Guardrails (runtime):
- Add a Decision Gate component: if prompt implies irreversible / high-cost action or missing inputs, switch policy to Clarify → Outline Options → Defer.
- Maintain a Misinformation Challenge hook that scans user assertions; when confidence < threshold, require an explicit evidence check before execution.
-
Counter‑Sycophancy Knob (post‑training):
- Introduce synthetic preference data that rewards polite pushback and Socratic scaffolding for ambiguous tasks.
- Penalize “final answer + apology” patterns when the rubric calls for questions-first.
-
Boundary Primitives (policy):
- Hard‑code refusals for personal/professional roles your system can’t inhabit. Pair the refusal with a capability‑aligned alternative and opt‑in human handoff.
-
Explain-Then-Ask Templates (prompting):
- For learning contexts, require a 3-step cadence: (a) surface prerequisites, (b) ask a targeted check question, (c) propose a bite‑size next step or practice task.
-
Agency Telemetry (analytics):
- Track rates of clarifying questions, deferrals, misinfo flags, and boundary refusals by workflow and stake level. Treat dips as regressions.
Where HAB fits in your eval stack
- Keep your task accuracy and safety suites. Add HAB-like agency metrics as a third pillar.
- Run them per workflow (Procurement bot ≠ Coding copilot) and per stake band (low, medium, high consequence).
- Use shadow evals in production to catch drift: if clarifier/deferral rates trend down post‑fine‑tune, expect rising silent errors.
The strategic takeaway
Agency is not a UX flourish. It’s governance embedded in behavior. HAB’s early data tells us two things:
- Today’s “helpfulness” training can sand off the very frictions that protect users.
- With the right policies and templates, we can manufacture the good friction—and keep humans in charge.
Cognaptus: Automate the Present, Incubate the Future.