A caseworker clicks “approve” after a system recommends a decision. The audit log records a human click, so a dashboard marks the case as human controlled. But the worker had twelve seconds, could not see the evidence behind the recommendation and would need a manager’s permission to reject it. The click is real. The control it appears to measure may not be.
Meaningful human control is a claim about a human’s ability to shape, understand and answer for what an automated system does. Measuring it requires more than counting approvals or placing a person somewhere in a workflow. This methods guide distinguishes the construct from nearby ideas such as perceived control, trust, compliance and manual override. It offers task designs and validity questions for research and governance, while keeping the difference between a proposed measure and a proven universal scale visible.
Define the construct before choosing a metric
Santoni de Sio and van den Hoven proposed two necessary conditions for meaningful human control over autonomous systems. Under their tracking condition, the system should respond to relevant human moral reasons and facts in its environment. Under tracing, an outcome should be traceable to at least one human within the chain of design and operation. This is a philosophical account, not a questionnaire that directly scores a public office or a clinical tool.
The account is useful because it widens the unit of analysis. A frontline worker may be able to press a stop button while the system’s objectives, thresholds and deployment context were chosen elsewhere. An engineer may design a model but have no authority over how a department uses it. To assess control, a study may need to follow decisions across design, procurement, configuration, operation and appeal, not only the instant when a recommendation appears on a screen.
“Meaningful” must also be tied to a particular risk and action. Control over a draft paragraph is not the same as control over a benefit decision, a medical intervention or an autonomous physical movement. A measure should say who is expected to notice what, within what time, with which information and power to change the outcome. Without that specification, a high score can be generated by a task too easy or too low-stakes to test the claim of interest.
Separate feeling, authority and effective intervention
Van den Bosch and colleagues proposed and tested three components of meaningful control in a simulated pandemic-triage human-AI team: subjective control, normative control and moral control. Their work is valuable because it does not assume that one feeling or button is the whole construct. Participants could report how much control they experienced while researchers also examined whether team decisions aligned with task norms and moral considerations.
Those components are not a finished certification test. The authors themselves note the simplified triage environment and participant-sample limits. They also judged their proposed moral-control measure insufficiently sensitive, so its result should not be read as a settled test of moral agency. Their work asks useful questions: Did the operator understand the team’s behavior? Could they keep decisions within relevant guidelines? Could they act on morally important information? A real clinical or public-sector system would require task-specific validation, including the conditions under which people lack time, evidence or authority.
Feeling in control can diverge from having control. A clear, confident interface may make a recommendation feel easy to approve even when important facts are hidden. Conversely, a worker may feel uncertain while correctly challenging a system. Subjective reports should therefore be read with behavioral traces and task outcomes. Neither should erase the other. A person who feels pressured while successfully overriding deserves attention; an organization cannot dismiss that experience because an override was technically possible.
Authority is another layer. A user may see the correct decision but be unable to change a threshold, pause deployment or appeal a decision upstream. To test authority, a researcher can document which actions are permitted, how long they take and what happens after a disagreement. An override rate alone is ambiguous: a low rate might mean excellent AI advice, deference to a misleading display, lack of authority or a task without meaningful opportunities to disagree.
Design tasks with correct and incorrect advice
A study that only shows correct AI recommendations cannot measure whether people can resist incorrect ones. It can measure agreement and perhaps speed, but a participant who accepts everything will look good. A stronger design includes cases where the AI is right and cases where it is wrong, with a defensible answer established independently. It records when people accept correct advice, reject wrong advice and make errors without advice.
Buçinca, Malaya and Gajos tested interface designs that asked people to think before seeing AI advice or added other forms of cognitive forcing. In their experimental tasks, some designs reduced overreliance on incorrect suggestions, while requiring more effort and sometimes being less preferred. This shows why a friction feature should be evaluated for both benefit and cost. More clicks or a delay do not automatically create better judgment, and the finding does not transfer unchanged to every profession.
Task difficulty matters. If wrong recommendations are obvious, rejecting them will not demonstrate the ability to catch subtle failures. If a task is impossible without specialist knowledge, failure may reflect missing capability rather than overreliance. A useful test includes cases spanning realistic difficulty and records the operator’s expertise. It also asks whether the evidence needed to challenge the AI is available in the interface and whether the operator has enough time to inspect it.
For a high-stakes workflow, the test can include changes in the environment: a case outside the training distribution, missing data, an ambiguous guideline or a system error that arrives when workload is high. The question is not whether a human can catch a planted typo. It is whether the organization has given them a workable opportunity to intervene when a foreseeable failure occurs.
Measure the whole chain, including traceability
Tracing asks where an outcome came from and who could have shaped it. A useful record may include the model version, data available at decision time, recommendation, operator’s initial judgment, final decision, override reason and later correction. These fields should be designed with privacy and legitimate use in mind. A perfect log is not meaningful control if no one can interpret it or use it to correct harm.
Accountability should follow the location of real authority. If a worker is responsible for a final click but procurement, model configuration and performance targets are fixed by others, attributing every adverse outcome to that worker would be misleading. A study can map who had the power to set objectives, change thresholds, suspend the tool and respond to appeals. The tracing condition then becomes a question about actual organizational paths rather than a list of names on an org chart.
NIST’s AI Risk Management Framework identifies roles, responsibilities and oversight arrangements as governance concerns. It asks organizations to map, measure and manage risks in context. The framework is voluntary guidance, not empirical evidence that a particular dashboard creates control. It can help structure a record of responsibilities and tests, but the institution still must show that its controls work in the relevant workflow.
Records must also support review after the fact. A decision maker may override correctly yet leave no explanation; an AI may generate a plausible rationale that hides the decisive feature. Researchers should distinguish an explanation offered by the system from a causal account verified through testing. Traceability is stronger when reviewers can reconstruct which facts, rules and people actually affected the outcome and can identify a route to remedy an error.
Legal requirements are a floor for the test
Article 14 of the EU AI Act addresses human oversight for high-risk AI systems. Its text calls for systems to permit effective oversight by natural persons and, as appropriate and proportionate, support understanding of limitations, monitoring, awareness of overreliance, interpretation, disregard or reversal of outputs, and intervention or safe stopping. This is a legal requirement in its specified scope, not proof that the listed capabilities are present in any individual deployment.
A measurement study can turn those functions into testable scenarios. Can a trained operator identify a known limitation? Can they spot an anomaly before it affects a person? Can they disregard an incorrect recommendation without punitive friction? Can they stop the system safely? The legal text does not supply one universal pass mark, and its applicability and timing must be checked for the particular system and date. The purpose here is to connect a policy requirement to observable practice, not to offer legal advice.
Paper compliance is especially easy to mistake for ability. A manual may say “human oversight” while the interface provides only a final approval screen. Training records may show that staff attended a session but not that they can identify errors under pressure. A meaningful test samples the actual conditions of use, including workload, conflicting incentives, language and access needs. It checks whether the oversight option is reachable before harm, not merely whether it exists somewhere in a settings menu.
Common confounds in control studies
AI accuracy is the first confound. When advice is highly accurate, people may rightly agree with it. Agreement does not prove blind dependence. When advice is poor, frequent overrides may be appropriate rather than evidence of distrust. Report accuracy by case type and compare operator decisions against an independent standard, not simply against the AI.
Task difficulty and expertise are next. A novice may accept a correct recommendation because it helps them; an expert may reject it because of case details the test did not capture. If groups differ in knowledge or case mix, a difference in override rate is not automatically a difference in control. Stratify or design the cases accordingly and report uncertainty rather than assuming one cause.
Interface friction and organizational power can also mimic choice. If rejecting the AI requires a long form or supervisor approval, few overrides may reflect cost. If the operator cannot access the evidence, a formal override permission has limited value. Measure the time, steps and consequences of disagreement. Ask whether the person can pause the workflow, request more information or appeal an upstream rule.
Self-report and logging have their own biases. A participant may say they felt in control because the interface was easy, or say they did not because the decision was morally difficult. A log can record that an operator clicked “reviewed” without showing what they understood. A robust study combines interview, behavioral record, system test and outcome review, while preserving privacy and documenting what each measure cannot reveal.
Build a validity argument, not a single score
Construct validity asks whether evidence supports the interpretation attached to a measure. A useful test begins with a claim such as: “Operators can identify and reverse materially wrong recommendations in this benefit workflow under ordinary time pressure.” It then designs wrong and right cases, observes decisions and reasons, checks authority to act, and examines what happens after an override. The result supports that specific claim to some degree, not all possible forms of human control.
Convergent evidence would come from different methods pointing the same way: task performance, explanations of limitations, usable interruption, traceable responsibility and appeal outcomes. Divergent evidence matters too. A person may have high confidence but poor error detection; a team may catch errors but have no power to stop deployment. A report should preserve those differences rather than averaging them into a reassuring number.
Replication should change cases, staff and conditions. If control vanishes under normal workload, it was not adequately measured in the easy lab condition. If one office succeeds only because a particular expert is present, the organization should not generalize that performance to every site. Measures should be sensitive to improvement and failure, and should reveal unequal burdens on people affected by decisions. A low complaint count may mean satisfaction, or it may mean that contesting an outcome is too hard.
What the evidence does not show
Santoni de Sio and van den Hoven’s tracking and tracing conditions are a philosophical account, not a validated two-item scale. Van den Bosch and colleagues’ triage measures are an empirical proposal tested in a simulation, and their moral-control measure did not achieve the sensitivity they sought; the study does not certify control in public benefits, clinical care or autonomous vehicles. The NIST framework and EU legal text organize responsibilities and requirements, but neither proves that a particular human can intervene effectively.
The cognitive-forcing experiment does not establish that more friction is always better. It reduced overreliance in tested tasks at an effort cost. An override count does not reveal whether the AI was correct, whether the person understood it or whether they had practical freedom to disagree. Even a well-designed laboratory test may fail to capture organizational incentives and the effects of decisions on the people who live with them.
No universal meaningful-control score follows from these sources. The proposed validity argument is an Institute synthesis: specify the decision, the human role, the risk, the available evidence, the intervention window and the route to correction; then test those claims with multiple methods. Its value lies in making “human in the loop” answerable to observable conditions rather than allowing the phrase to stand as proof.
Return to the approval click
The caseworker’s click may be an important part of the record. To know whether it represented control, a reviewer needs to see what evidence the worker had, what alternatives were feasible, how a wrong recommendation could be challenged and who could change the system when a pattern of errors appeared. A count of approvals cannot answer those questions.
Meaningful human control becomes measurable when a claim about it is tied to a particular task and a real capacity to change an outcome. The strongest evaluation does not search for a ceremonial human action. It tests whether people can understand, contest and redirect the system, and whether the organization remains responsible for making that possible.
Research references
- Santoni de Sio, F., & van den Hoven, J. (2018). Meaningful human control over autonomous systems: A philosophical account. Frontiers in Robotics and AI, 5, 15. View source
- Van den Bosch, K., Van Diggelen, J., Verdult, S., Haije, T., & Van der Waa, J. (2025). Measuring meaningful human control in human–AI teaming: Effects of team design in AI-assisted pandemic triage. AI and Ethics, 5, 3329–3353. View source
- Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188. View source
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. View source
- European Union. (2024, consolidated 2026). Regulation (EU) 2024/1689, Article 14: Human oversight. View official text