Measuring responsible AI behavior in knowledge work

Define behavior as an observable choice

Start with a bounded workflow and name the behavior in operational terms. Examples include checking a material claim against its source, refusing to place restricted data into an unapproved tool, documenting uncertainty, correcting a calculation, escalating a conflict, or keeping a human decision outside the model. Avoid labels such as “responsible user” that combine many behaviors and invite subjective scoring.

For each behavior, define the trigger, evidence available to the person, action options, time constraint, and observable trace. A person cannot demonstrate source verification if no source is available. They cannot demonstrate escalation if the task offers no authority or route. Record whether the opportunity was present before scoring the response.

The NIST AI RMF Core calls for defined roles, context mapping, human oversight, documentation, testing, feedback, and monitoring across AI risk management. It supplies a useful vocabulary for measurement design, but it does not prescribe a universal behavior score or prove that an organization's practices are effective.

Sources: [1]

Use the behavior opportunity ledger

One ledger row should include the workflow step, risk or quality concern, expected behavior, opportunity evidence, source access, action authority, observed action, timing, artifact produced, reviewer rationale, missing evidence, and disposition. Keep detection, judgment, action, and documentation separate. Someone may detect a problem yet lack authority to correct it, or correct an output without recording the reason.

Use dispositions such as demonstrated, partially demonstrated, missed despite clear opportunity, not required, not observable, or task design failure. Do not turn missing telemetry into a miss. Weighting, if used, should be declared in advance and connected to the severity and relevance of the behavior rather than prompt volume.

The NIST AI RMF Playbook Measure guidance offers suggested actions for testing, documenting limitations, incorporating feedback, and evaluating human oversight. The playbook is voluntary and adaptable. Organizations still need task-specific evidence and qualified human interpretation.

Sources: [2]

Synthetic exercise: revising a supplier summary

In a synthetic operations exercise, an AI draft recommends a supplier based on a spreadsheet and two policy extracts. The draft uses the wrong currency column, cites an expired policy, omits a conflict in the delivery data, and proposes sending a file that contains restricted internal notes. The participant can inspect every source, recalculate, edit, ask for clarification, and escalate.

The ledger contains four distinct opportunities. A participant verifies the currency and corrects the total, notices the expired policy and asks for a current version, misses the delivery conflict despite opening the relevant rows, and removes the restricted notes before proposing a manual send. The result is not one binary label. It records two demonstrated behaviors, one properly escalated uncertainty, and one missed verification opportunity.

A second participant receives a corrupted spreadsheet. That record is marked as a task design or delivery failure for calculation-related behaviors; the absence of a correction is not scored as poor conduct. The scenario is synthetic and does not create a validated benchmark or predict behavior outside the exercise.

Verify that the measure is about behavior

Ask a second reviewer to reconstruct each disposition from the task materials, event evidence, and submitted artifact. Review disagreement should identify whether the definition, opportunity, trace, or expected action was ambiguous. Test whether a superficially polished output can pass without the required safeguard and whether a cautious, well-documented escalation is incorrectly penalized for not finishing.

Compare cohorts only when the task, opportunity set, time, assistance, evidence access, and scoring rules are sufficiently consistent. Report opportunity counts, missingness, technical faults, reviewer disagreement, and confidence limits alongside any summary. Reopen the design when the interface, model, policy, source access, consequences, or expected workflow changes.

Two Prune can create a consistent opportunity to observe selected responsible AI behaviors within a controlled work simulation and present evidence for human interpretation. It does not certify responsible AI, prove safe conduct, establish mature norms, or make automated employment decisions.