An automated system recommends a route around a hazard. The screen displays a confidence number and a paragraph explaining the recommendation. A user says the system is transparent, then cannot say which hazard was detected, what assumption the route depends on, or when the recommendation should be ignored. The interface contains information, and the user feels informed, but the critical understanding may still be absent. Measuring automation transparency requires separating what a system discloses from what a person can find, understand, verify, and use. A design can succeed on one of these dimensions while failing on another.
Define transparency for the task
Automation transparency is often used for visibility into an automated system’s state, actions, reasons, limits, or likely future behavior. It is not a single property that an interface either has or lacks. Chen and colleagues’ situation-awareness-based agent transparency model organizes disclosures about current actions and plans, reasoning, and projected outcomes. That framework helps a researcher ask what information a human teammate needs for the task. It does not imply that every interface should show every internal calculation. The measurement target is the usefulness of selected disclosures for particular decisions and users.
System transparency and perceived transparency are different constructs. A system can display a detailed log that nobody can interpret. Another may offer a short, timely explanation that makes its current action predictable. A self-report scale asks how transparent the system seems; a knowledge probe asks what the user can accurately state; a behavioral task asks whether that understanding improves decisions. These measures need not agree. Their disagreement can show whether an interface creates a feeling of clarity without supporting a working model, or supports useful understanding without being liked.
Angelopoulos and colleagues developed the Transparency Of RObots Scale to assess perceived transparency of robot behavior. Their validation identified factors labeled illegibility, explainability, and predictability across studied languages and scenarios. This is a valuable instrument for its stated domain, not a universal test of every automated decision system. A questionnaire about a robot’s visible behavior may not transfer unchanged to a statistical risk score or invisible background process. Researchers should check item meaning, factor structure, and relation to behavioral understanding after adapting it to a new technology.
Transparency also differs from trust. A person may trust a system because it has performed well before, because an institution endorses it, or because an interface looks polished. They may understand a system’s limitations and rationally withhold trust. Dzindolet and colleagues showed that trust matters for reliance decisions in a specific automated-aid experiment. Their result does not make trust a measurement of transparency. A transparency study should report trust separately and ask whether understanding improves the calibration of reliance to the system’s actual reliability, rather than merely raising trust scores.
Measure disclosure, access, and comprehension separately
An interface audit can record what information is available: current mode, input data, action plan, explanation, uncertainty, history, and means to contest or override. That is a property of the design as implemented. It does not establish that users notice the information or can interpret it. To test access, give participants a realistic task and observe whether they can locate the relevant information when needed. To test comprehension, ask them to explain in their own words what the automation is doing and what would change the recommendation. The answer key should be built from the actual system behavior, not from the designer’s hope about what the display communicates.
Comprehension questions should be specific enough to distinguish recognition from a usable model. “Was an explanation shown?” tests memory for a screen element. “Why did the system choose route A, and under what condition would route B become safer?” tests a deeper relation between evidence and action. But the question must not require technical knowledge irrelevant to the user’s role. An operator may need to know that a sensor is uncertain, not the exact optimization algorithm. Task analysis and user involvement help define which facts are necessary for appropriate use.
A projection task can test whether users anticipate the next automated action. After a system changes mode or encounters new evidence, ask what it will do and why. Prediction is especially relevant where a human may need to intervene. Correct prediction under one scripted scenario is not proof of a general mental model, however. The participant might infer from superficial cues or memorize a pattern. Vary conditions and include cases where the same superficial cue leads to different actions. Report performance by scenario so a high average does not conceal a failure on the rare transition that matters most.
Another approach measures error detection. Present recommendations that are sometimes correct and sometimes wrong, and ask whether users accept, reject, or seek more information. The key outcome is not maximum rejection or maximum acceptance. It is calibrated use: following the aid when it is sound and recognizing limits when it is not. This requires a well-characterized set of cases and a defensible ground truth. If errors are obvious, the task may overstate human oversight; if cases require inaccessible expertise, it may unfairly make the explanation look ineffective. Case difficulty should be measured or balanced.
Transparency manipulations have side effects
Selkowitz and colleagues varied agent-transparency displays in a virtual robot monitoring task and found differences in situation awareness, cognitive processing, and trust across interface levels, without detected differences in workload or perceived usability in that study. The result supports the value of testing disclosure levels experimentally. It does not prove that more information always helps. The levels bundled particular kinds of content, and the task and population constrain transfer. A later nautical-navigation study by van de Merwe and colleagues explicitly examined transparency alongside scenario complexity, illustrating that effects may depend on the surrounding task.
More explanation can impose work. Text that appears at the wrong moment competes with hazard monitoring, and a detailed rationale can bury the one fact needed for action. A transparent interface must be evaluated for timing and attention cost as well as content. The same information might help in a post-event review but distract during a fast handoff. Researchers can compare concise and detailed variants while holding layout and visual salience as stable as feasible. If the richer version also uses a larger panel, any outcome may reflect display area rather than explanation quality alone.
Kizilcec’s field experiment on an algorithmic peer-assessment interface found that transparency affected trust differently depending on whether outcomes violated expectations and how much information was provided. It is a reminder that disclosure does not act independently of the user’s prior beliefs or stakes. A result about grading should not be generalized wholesale to robots or clinical aids, but the measurement lesson transfers: record whether the system agreed with expectations, whether the user was affected by the outcome, and what information the explanation actually added. A before-and-after trust score without these conditions can be hard to interpret.
Bussone and colleagues studied explanations in clinical decision support and reported a tradeoff: fuller explanations raised trust but also raised over-reliance in their study, while less detailed ones led to self-reliance problems. The implication is not that explanations should be withheld. It is that success must be evaluated against appropriate reliance and decision quality, not against trust alone. A system can be liked, believed, and dangerous when wrong. A transparency measure should make the user’s opportunity to verify the system part of the test, especially when consequences are high.
Confounds in surveys and experiments
Perceived transparency can be inflated by fluent language, visual polish, or an explanation that sounds coherent but is not faithful to the system’s actual process. A participant may rate a confident narrative as understandable because it fits expectations. One way to test this is to include questions whose correct answers depend on the disclosed evidence and compare them with the explanation’s content. Researchers should also verify technical fidelity: if the explanation describes a post hoc story unrelated to the model’s calculation, user comprehension of that story is not comprehension of the automation. The transparency claim must specify what it is transparent about.
Prior experience changes both understanding and ratings. Experts may need fewer words but be more skeptical about missing information; novices may feel informed by a simple diagram while remaining unable to detect error. A system tested only with students may not work for trained operators, and an operator study may not represent public users affected by automated decisions. Stratify or measure expertise, and avoid treating group differences as deficits in users. If the design requires rare technical expertise to understand, that is a property of the design and its intended audience.
Outcome feedback is another confound. A participant told that the aid was accurate on every practice trial may trust a later explanation more regardless of its content. Conversely, one visible failure can dominate perception. Experiments should balance reliability and explanation conditions where possible, report the history of observed system behavior, and analyze whether effects differ after correct and incorrect recommendations. Blindness to condition can be difficult when interfaces visibly differ, but objective scoring of decisions and prespecified outcomes can reduce interpretive flexibility.
Timing matters for memory as well as behavior. A posttask question about what the system showed can be influenced by the final outcome, and participants may infer reasons after seeing whether the decision succeeded. Probes delivered during work preserve temporal relevance but may interrupt the task. Logs of information seeking can help show when explanations were opened, yet a click is not evidence of comprehension. Combining logs with time-sensitive questions and decisions is stronger than using any one indicator. Each addition should be justified against the disruption it creates.
Transparency, situation awareness, and control
Automation transparency can support situation awareness without being identical to it. Knowing the automation’s current mode is one part of understanding the larger environment. An operator may understand the automated plan but miss a separate external hazard. Conversely, they may understand the hazard while being surprised by the automation. A study can use situation-awareness probes that distinguish world state from automation state. If all questions concern the system’s explanation, it has measured understanding of the explanation rather than full operational awareness.
Control is a separate property too. A user may know why an automated decision occurred but lack the ability to correct it. Another may have an override button without understanding when to use it. Transparency can be assessed alongside directability, contestability, and auditability, but one should not stand in for the others. In a planning or clinical system, an explanation may need to support review by a team after the event as well as immediate use by an individual. Different audiences and timescales call for different measures.
Teams complicate interpretation. The person who reads a diagnostic explanation may not be the one who acts on it. Information may be communicated inaccurately at a handoff. A team measure could test whether the relevant reason and uncertainty reach the decision maker at the moment needed, rather than simply survey each member’s impression of transparency. Communication logs and role-specific probes can help. The question is whether the designed information flow supports coordinated action, not whether every individual can recite every internal detail of the system.
Report the measure that was actually used
A transparent methods report names the automation, its actual reliability, the information shown in each condition, the user role, and the decisions participants made. It distinguishes design-level disclosure, perceived transparency, tested comprehension, prediction of future action, trust, and reliance. It describes how correct answers were defined, which scenarios contained errors, and whether outcome feedback was given. For a self-report scale, include item wording, adaptation, and factor evidence. For behavioral tests, report false acceptance and false rejection, not only overall accuracy. These details let another team see whether a result can transfer to its own setting.
Interpret group averages alongside distributions. An explanation that helps most users but confuses a minority in rare high-risk cases may be unacceptable in a safety-critical workflow. A subgroup comparison requires adequate sample and careful interpretation, but hiding variability is not neutral. Qualitative accounts of confusion can guide redesign: users may need an uncertainty cue, a clearer mode indicator, or a way to inspect the evidence. The goal of measurement is to improve the fit between what the system reveals and what people must decide.
What the evidence does not show
The cited studies do not show that more disclosure always yields better decisions, that trust is equivalent to understanding, or that one perceived-transparency scale applies to every automated system. A successful explanation in a controlled experiment may fail when timing, stakes, expertise, or system reliability changes. Showing the internal workings of an algorithm does not by itself make its output fair, correct, or contestable. The evidence supports targeted tests of what users can comprehend and do with disclosures, bounded by task and population. It does not support a single context-free transparency score.
A useful transparency study ends with a sharper claim than “users understood the AI.” It can say that under specified conditions users identified the current mode, predicted the next action, recognized a defined failure, or calibrated reliance more accurately. It can also say that a polished explanation increased confidence without improving error detection. Both results matter. By keeping disclosure, perception, comprehension, trust, and action distinct, measurement can help designers make automation answerable to the people who depend on it.
Research references
- Chen, J. Y. C., Lakhmani, S. G., Stowers, K., Selkowitz, A. R., Wright, J. L., & Barnes, M. (2018). Situation awareness-based agent transparency and human-autonomy teaming effectiveness. Theoretical Issues in Ergonomics Science, 19(3), 259–282. View source
- Selkowitz, A. R., Lakhmani, S. G., & Chen, J. Y. C. (2017). Using agent transparency to support situation awareness of the Autonomous Squad Member. Cognitive Systems Research, 46, 13–25. View source
- Angelopoulos, G., Lacroix, D., Wullenkord, R., Rossi, A., Rossi, S., & Eyssel, F. (2025). Measuring transparency in intelligent robots. Scientific Reports, 15, 43809. View source
- Kizilcec, R. F. (2016). How much information? Effects of transparency on trust in an algorithmic interface. Proceedings of CHI 2016, 2390–2395. View source
- Bussone, A., Stumpf, S., & O’Sullivan, D. (2015). The role of explanations on trust and reliance in clinical decision support systems. 2015 International Conference on Healthcare Informatics. View source
- Dzindolet, M. T., Peterson, S. A., Pomranky, R. A., Pierce, L. G., & Beck, H. P. (2003). The role of trust in automation reliance. International Journal of Human-Computer Studies, 58(6), 697–718. View source
- van de Merwe, K., Mallam, S., Nazir, S., & Engelhardtsen, Ø. (2024). The influence of agent transparency and complexity on situation awareness, mental workload, and task performance. Journal of Cognitive Engineering and Decision Making, 18(2), 156–184. View source