Do AI explanations help people know when to trust a system?
A South Korean survey found that people who perceived an AI system as clearer also reported more understanding, trust and reliance. It did not test whether they actually understood the system or trusted it appropriately.
The short version
- What happened
- In a convenience sample of 342 South Korean AI users, perceived explainability and transparency were associated with self-reported understanding, while perceived accuracy and understanding were associated with trust.
- Who it affects
- Workers and public-service professionals need performance evidence and practice with error cases, not only explanation panels that increase confidence.
- What remains unproven
- Cross-sectional associations cannot establish that explanations caused understanding, trust or reliance.
Keep in mind
Cross-sectional associations cannot establish that explanations caused understanding, trust or reliance.
Every construct was self-reported; objective comprehension, system accuracy and observed behaviour were not measured.
One perceived-calibration item overlaps with reliance and may inflate the association between those constructs.
The convenience sample was young, frequently used AI and was not representative of South Korea.
The system each participant rated was not recorded, and the English questionnaire had no reported formal pilot or cultural adaptation.
Source analysis · 1 linked source. Primary analysed source date: 14 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The answer: clearer systems felt easier to trust, but calibration was not tested
People who regarded an AI system as explainable and transparent also tended to say they understood it, trusted it and relied on it. That is the central result of a peer-reviewed survey published on 14 October. It is useful evidence about how perceptions cluster, but it does not show that explanations made anyone more accurate, that the system deserved trust or that reliance improved a decision.
In human-factors research, calibrated trust normally means that confidence tracks a system's real capability: people rely on it when it is right and hesitate when it is wrong. Here, calibration was a participant's own judgement that their trust matched performance. The researchers did not independently measure the system, comprehension or the correctness of any action.
For product teams and public-service managers, the lesson is not to add an explanation panel and infer that safe reliance follows. Explanations may support a feeling of understanding. Whether that feeling helps a clinician, teacher, benefits officer or worker spot an error requires behavioural testing against known answers and realistic consequences.[1]
Who took part and what they were asked
Recruitment ran through universities, academic networks and professional communities in South Korea between May and June 2026. The researchers received 463 responses, removed 93 incomplete submissions and 28 duplicates, and analysed 342. The sample was young and experienced with AI: 68.1% were aged 20 to 30, 54.1% were men, and 75.1% reported frequent or very frequent use.
Education and occupation were not collected, so the study cannot show whether students, professionals or high-consequence workers responded differently. Convenience recruitment means the group is not representative of South Korea. Each participant selected one recently used AI system, but the researchers did not record which product. Ratings could therefore concern very different interfaces, tasks and error patterns.
The questionnaire measured eight constructs with three statements each: perceived explainability, transparency, understanding, accuracy, fairness, trust, trust calibration and reliance. Responses used a five-point scale. Every construct was self-reported, so labels such as accuracy and calibration describe what a person believed, not independently verified performance.[1]
How the analysis worked
The team used partial least squares structural equation modelling to estimate its proposed network of associations. It bootstrapped 5,000 samples for significance tests and used ten-fold cross-validation repeated ten times for predictive checks. Statistical tests found strong internal reliability in this sample, but reliability does not establish that a scale captures real understanding or appropriate behaviour.
The questionnaire was in English even though participants were in South Korea. The paper reports no formal pilot, expert review or cultural adaptation. The design took every measure once, from the same person, in the same questionnaire. It therefore cannot establish temporal order or causality and remains vulnerable to recall, enthusiasm and social-desirability effects.
A participant who generally liked AI could rate explanation, accuracy, fairness, trust and reliance highly without any feature having caused another. Structural modelling can show whether the hypothesised pattern fits the responses; it cannot turn cross-sectional opinions into a test of system safety.[1]
What the numbers show
Perceived explainability was associated with self-reported understanding with a standardised coefficient of 0.349, while transparency was associated with understanding at 0.297. Understanding was associated with trust at 0.388. The largest direct predictor of trust was perceived accuracy, at 0.492; perceived fairness also contributed more weakly.
Trust was associated with perceived trust calibration at 0.631, and perceived calibration was associated with self-reported reliance at 0.600. The model accounted for roughly 36% to 54% of variance across main downstream constructs. These are associations within the survey, not percentages of users who became correctly calibrated and not estimates of how much an interface will improve decisions.
Perceived accuracy being more strongly associated with trust than perceived understanding is important. People may trust an AI because it appears right even when they cannot verify it. Fluent, confident presentation can create that impression. A responsible deployment needs visible evidence about actual performance and failure modes, not only an explanation of a result.[1]
Why self-reported calibration is the critical limitation
One calibration item asked whether participants relied on the system appropriately. That wording overlaps with the separate reliance outcome and may inflate the strong relationship between them. More fundamentally, a person cannot confirm that trust is appropriate without an external measure of when the system succeeds and fails.
The study did not present correct and incorrect AI answers, vary explanation quality or observe whether participants accepted bad advice. It did not compare confidence with measured accuracy. Someone who confidently over-trusted a polished but inaccurate system could score highly on perceived accuracy, trust, calibration and reliance simultaneously.
Common-method diagnostics cannot remove that measurement problem. The safest reading is a map of perceptions: explanation, transparency and perceived accuracy sit upstream of reported trust, which sits upstream of reported reliance. The study is not a validation of safe behaviour.[1]
What designers and organisations can do
Designers can test explanations against comprehension. Users should be asked to predict what the system will do in a new case, identify when an input is outside scope, explain a recommendation in their own words and reject intentionally wrong output. Confidence should be compared with correctness rather than collected alone.
Managers should provide task-specific performance information: validated error rates, relevant subgroups, failure examples and an escalation route. A generic accuracy figure can mislead when workplace data differ from a test set. High-consequence deployments need logs and human review that show whether explanations help people catch errors or merely make automated advice feel legitimate.
Training should include disagreement cases. Workers need practice deciding when to consult another source, seek a specialist or stop a workflow. The goal is neither maximum trust nor minimum trust, but discrimination: appropriate use when evidence is strong and deliberate resistance when the model is uncertain, out of scope or contradicted by authoritative information.[1]
Funding, geography and what would change the assessment
The Hoseo University Academic Research Fund 2026 funded the work. The author declared no commercial or financial conflict and stated that generative AI was not used to prepare the article. Cultural expectations can still shape trust, and the young, frequent-user sample plus English questionnaire make generalisation uncertain.
Stronger evidence would randomise explanation and transparency features while holding system accuracy constant, then measure objective comprehension, error detection and actual reliance on correct and incorrect advice. Longitudinal studies could test whether confidence adjusts after repeated success and failure.
Until then, the paper supports a restrained conclusion: explanation and transparency are associated with feeling informed, but feeling informed is not evidence of calibrated trust.[1]