A participant hears the same unfamiliar claim from two people. One has been accurate before; the other has repeatedly guessed. The participant chooses the first person’s answer. That choice is evidence of selective reliance in this task, but does it mean the participant generally trusts knowledge communicated by others? It might instead reflect memory for prior accuracy, a preference for a familiar speaker, or a sensible response to this particular claim. The measurement question begins where the simple choice ends.
Measuring epistemic trust requires a precise account of what is trusted, from whom, under which conditions, and with what consequence for learning. Epistemic trust concerns receptivity to communicated knowledge as potentially reliable, relevant and useful beyond the immediate exchange. It is not identical to liking a person, agreeing with a statement, judging a website credible, or being vulnerable to misinformation. This Methods & Measures entry compares self-report and behavioral approaches, explains what each can support, and shows how a researcher can test a claim without treating a questionnaire score or a single choice as the whole construct.
Specify the trust claim before choosing a measure
The word trust can hide several distinct questions. A person may regard a particular speaker as knowledgeable about medicine but not about finance. They may believe a statement is accurate while finding it irrelevant to their own situation. They may listen politely yet never use the information later. They may accept a claim in a low-stakes experiment but seek independent confirmation before acting in daily life. Calling every one of these responses epistemic trust makes the construct too broad to measure well.
Campbell and colleagues define epistemic trust in relation to communicated knowledge that a recipient regards as significant, personally relevant and generalizable. Their Epistemic Trust, Mistrust and Credulity Questionnaire, or ETMCQ, treats trust, mistrust and credulity as related but distinct reported stances. The distinction matters: reasonable openness is not the opposite of every doubt, and uncritical acceptance is not simply a high amount of trust. A study that scores these positions on one line from suspicious to trusting would impose a structure that the questionnaire developers did not claim.
A second tradition studies selective trust in testimony by asking whether people use what one informant says rather than what another says. In Koenig, Clément and Harris’s experiment, young children were shown speakers with different histories of correct labeling before being asked about unfamiliar objects. Their choice allowed an inference about reliance on prior accuracy under the task conditions. It did not measure a durable adult trait, a therapeutic relationship, or a person’s ability to detect every false claim. These traditions overlap around learning from others but operate at different levels of description.
An evaluation should therefore state its target in a sentence: “We are measuring an adult’s self-reported stance toward socially transmitted knowledge,” or “We are measuring whether children use an informant’s prior accuracy when deciding whom to ask.” Such a sentence determines whether a questionnaire, an informant-choice task, a longitudinal observation of learning, or some combination is appropriate. It also makes missing evidence visible. A claim about generalization to new settings requires a later or different setting, not merely an answer given immediately after a vignette.
What self-report questionnaires can and cannot capture
The ETMCQ is a concrete way to study reported epistemic stance in adults. Campbell and colleagues developed and examined it in two studies, with samples of 500 and 705. They reported a three-factor pattern for trust, mistrust and credulity, alongside reliability analyses and associations with other reported experiences. This is evidence that responses can be organized meaningfully in those samples. It is not direct observation of whether a participant actually learns from a particular person when stakes, relationships or evidence change.
Questionnaire answers are interpretations of one’s typical conduct. Someone may endorse caution because they see themselves as thoughtful, even if they accept persuasive claims in practice. Another may endorse trust because they value social openness while carefully checking unfamiliar expertise. A statement such as “I am open to information from others” can reflect personality, social desirability, recent experience, or the imagined kind of other person. Researchers should specify the reference period, response options, language version and population before comparing scores across groups or times.
Internal consistency asks whether items assigned to a scale tend to move together. It does not show that the scale predicts real-world updating, captures a stable underlying cause, or works identically in another language. Test-retest stability can be useful for a proposed trait-like stance, but a perfectly stable score would be a poor test of a theory that also expects change after a relationship becomes safe or unsafe. Conversely, change in a score after an intervention could reflect demand effects, mood, changed interpretation of the questions or genuine change. A credible design measures more than one of these possibilities.
Cross-language validation illustrates why item wording matters. In a French-language validation, Greiner and colleagues reported a 12-item version and a three-factor solution in 302 online participants. That result supports use of their adapted instrument in a defined population; it does not establish item-for-item equivalence with every other ETMCQ version. A researcher comparing national means must first test whether people in the groups understand and use the items similarly. Translating “trust” accurately as a word is not enough if the social situations imagined by respondents differ.
Self-report can still answer valuable questions. It can describe how people characterize their own openness, wariness and credulity; identify correlations worth testing; and track reported changes over time. It is especially useful when combined with a specified theory and independent outcomes. The error is not using a questionnaire. The error is reporting a questionnaire association as proof of a hidden learning mechanism, a diagnosis, or a treatment effect without the designs needed for those conclusions.
Behavioral tasks: selective reliance is not unrestricted trust
A behavioral task can make the informational decision observable. The classic selective-testimony design gives participants a chance to learn that one informant has been accurate and another inaccurate, then presents a new question for which participants lack an answer. Researchers can count who is asked, whose claim is endorsed, or whether the participant changes an earlier answer. Koenig, Clément and Harris showed that some three- and four-year-olds monitored earlier accuracy and later used the more reliable informant. The result supports selective trust in that experimental arrangement, particularly among children who reliably identified the informants.
Koenig and Harris later showed that task details and developmental stage affect performance. In their experiments, four-year-olds more consistently predicted, sought and endorsed information from an accurate speaker than three-year-olds in particular conditions. This is a warning against treating a single age contrast as a pure measure of trust. To choose accurately, a child must understand the instructions, remember who said what, identify which prior labels were correct, and tolerate the social demand of rejecting an adult’s answer. A weaker choice pattern can arise from any of those steps.
Task design should separate knowledge from benevolence. A speaker who is wrong because they lack access to a fact differs from one who deliberately misleads. A friendly expert can still be mistaken outside their field. A speaker who was accurate about familiar object names may not be reliable about a novel scientific claim. If an experiment makes the “good” informant also warmer, more fluent, more familiar or more authoritative, preference cannot cleanly be attributed to prior accuracy. Counterbalancing speaker identity, voice, position and order helps; it does not remove every contextual cue.
The response itself also matters. Asking whom to approach measures an information-seeking choice. Endorsing an answer measures immediate reliance. Recalling it a week later measures retention. Applying it to a new situation tests transfer. These are related outcomes, but none is interchangeable with the others. A person may know whom to ask and still be unwilling to rely on the answer; they may repeat an answer without understanding it; they may understand it while refusing to act because the costs of error are high. Researchers should align the outcome with the proposed mechanism.
One useful design offers a speaker whose accuracy history is known, a neutral speaker, and a reason to verify the new claim. If the participant checks evidence when it is available, that may indicate calibrated vigilance rather than mistrust. Sperber and colleagues’ account of epistemic vigilance emphasizes scrutiny of both the source and the communicated content. In this framework, critical evaluation is part of successful learning from testimony, not evidence that trust is absent. A task that rewards blind acceptance may therefore be a poor measure of adaptive epistemic trust.
Construct validity requires convergence and separation
A measure gains credibility when it relates to outcomes that the theory predicts while remaining distinguishable from adjacent constructs. For epistemic trust, this means examining whether questionnaire scores relate to observed use of trustworthy information, openness to correction, or transfer of useful knowledge across contexts. It also means asking whether the same scores merely reproduce general interpersonal trust, social desirability, attachment insecurity, anxiety, or agreement with the experimenter. Correlation with a neighboring measure is a clue to investigate, not a proof of identity.
Campbell and colleagues found associations between mistrust or credulity and several developmental and mental-health self-report variables. They also reported that trust itself was not associated with lower symptom levels in the same straightforward way. Those patterns argue against a simplistic “more trust always means better mental health” message. Their cross-sectional analyses can describe covariation and test statistical models. They cannot establish that early adversity caused a later epistemic stance, which then caused symptoms. That causal chain would require stronger temporal and alternative-explanation controls.
The French validation likewise reported associations involving attachment, adversity and loneliness. It used an online sample and mediation models. A mediation coefficient in cross-sectional questionnaire data does not demonstrate a sequence unfolding in real life. People may report loneliness, mistrust and symptoms together because of a shared influence, because one affects another in a different direction, or because item wording overlaps. Researchers should treat these analyses as hypotheses for longitudinal and experimental work.
The convergence question can be tested more directly. One might first measure reported stance, then observe how participants update beliefs when a speaker provides accurate, inaccurate and ambiguous information across several domains. A later test could ask whether learned content is retained or generalized. Multiple speakers, independent evidence checks and different stakes would help identify when reliance is flexible. The study would still need to separate ability from preference: an adult who fails to use a speaker’s information may already know the answer, doubt its relevance, or lack motivation to finish the task. A well-designed protocol records these possibilities rather than forcing a single interpretation.
Separation from source credibility is particularly important. Source credibility is often an appraisal of a particular communicator or outlet for a particular claim. Epistemic trust is an orientation toward taking knowledge from others as potentially worth learning and using, though it is enacted in relationships and situations. A person can have a generally open stance and judge one outlet unreliable. Conversely, they can distrust social learning in general yet accept a specific expert’s explanation after careful checking. This entry examines the measurement of that orientation and its situated expression; the related Source Credibility entry concerns properties and judgments of the source itself.
Susceptibility to misinformation is another neighboring outcome. Believing a false claim may reflect low knowledge, repeated exposure, a trusted relationship, identity motives or a persuasive format. It cannot automatically be scored as high epistemic trust. Credulity in the ETMCQ is a reported stance of insufficient vigilance, but a scale score is not a direct estimate of how many false claims a person will accept. Likewise, rejecting a true claim is not automatically epistemic mistrust. Accurate classification of the content and the participant’s opportunity to verify it are necessary before a task can speak to calibration.
Design a defensible measurement protocol
Begin with the intended inference. If the question is whether a training program changes adults’ willingness to learn from a mentor, identify the mentor relationship, the kind of knowledge, the duration of follow-up and the observable learning outcome. If the question is whether children track past accuracy, construct a task with controlled informant histories and age-appropriate comprehension checks. If the question is whether a therapeutic alliance supports learning beyond the session, do not substitute a general trust questionnaire for observations of what was learned and carried into daily decisions.
Choose indicators at more than one level when the claim spans levels. A self-report scale can represent perceived stance; an informant-choice task can represent immediate reliance; a delayed transfer task can represent one form of generalization. These indicators should not be averaged into a single “epistemic trust score” unless the study has established that the combination has a coherent meaning. Report each outcome and the proposed link between them. A disagreement can be informative: a participant may feel trusting but discriminate sharply among speakers, or report mistrust while learning well from an identified expert.
Document the informational environment. Was the message relevant to the participant’s goal? Did the source have genuine knowledge? Could the participant inspect the evidence? How costly was a wrong answer? Were there incentives to agree with the speaker or the researcher? A condition in which all speakers are equally correct cannot test selective reliance on accuracy. A condition in which one speaker has an obvious status advantage cannot isolate trust from authority. Pretesting whether claims are known, comprehensible and personally relevant reduces ambiguity.
Record task confounds prospectively. For developmental studies, assess memory for speaker identity and prior performance. For adults, test whether respondents understood scale items as referring to friends, professionals or strangers. For translated questionnaires, examine factor structure and measurement invariance before interpreting group differences. For longitudinal studies, account for attrition and changing relationships. For digital environments, record interface cues that can confer apparent authority. None of these controls makes inference perfect; each closes a specific route by which a different process could mimic the target.
Interpret the evidence in layers. An item score supports a claim about answers to those items. A replicated factor structure supports an account of how responses cluster in tested samples. A choice between informants supports a claim about behavior in that task. Repeated, context-sensitive use of accurate testimony supports a stronger claim about calibrated social learning. Clinical or social outcomes require their own evidence. This layered language makes a result more useful because readers can see exactly where observation ends and interpretation begins.
What the evidence does not show
No cited study establishes a universal unit, cutoff or biomarker for epistemic trust. The ETMCQ and its adaptations do not diagnose a person as safe, gullible or incapable of learning. Selective-testimony experiments do not prove a durable trait from a single choice. Neither a correlation with symptoms nor a statistical mediation model demonstrates that changing trust will improve treatment. Developmental findings from controlled speaker tasks cannot simply be projected onto adult workplaces, political information environments or psychotherapy.
The evidence does not show that trust should always increase. Appropriate vigilance can protect learning when a source lacks knowledge or has reason to mislead. Nor does it show that every hesitation is maladaptive mistrust. A participant may have good reasons to seek corroboration, and those reasons can change with the stakes. Measurement that labels all checking as a deficit would confuse careful judgment with closedness.
Existing research also does not settle whether self-reported stance, laboratory selection of an informant and long-term relational learning share a single underlying cause. Their relationships are empirical questions. The original and translated questionnaires provide important psychometric work, but differences in item sets, language, population and setting limit direct score comparison. This uncertainty is a reason to report the actual instrument, sample and task rather than retreat to a vague claim that “epistemic trust was measured.”
Reading a result with care
When a paper reports an epistemic-trust finding, a reader can ask five concrete questions. What definition did the authors use? What behavior or answer was observed? What did the participant know about the source and the claim? Which alternative explanations were tested? How far beyond that task did the authors generalize? These questions do not demand a perfect measure. They keep a useful construct connected to observable evidence.
The strongest research program will allow trust to be selective, revisable and situated. It will measure when a person is open to knowledge, when they question it, and whether those responses fit the evidence available to them. That approach respects the value of learning from others without mistaking automatic agreement for wisdom. It also makes epistemic trust a sharper research question: not simply whether someone says they trust, but how they decide when another person’s knowledge deserves to change what they know.
Related Learning Hub entries
Research references
- Campbell, C., Tanzer, M., Saunders, R., et al. (2021). Development and validation of a self-report measure of epistemic trust. PLOS ONE, 16(4), e0250264. View source
- Greiner, C., Besch, V., Bouchard-Boivin, M., et al. (2025). Epistemic Trust, Mistrust and Credulity Questionnaire (ETMCQ) validation in French language: Exploring links to loneliness. PLOS ONE, 20(3), e0303918. View source
- Koenig, M. A., Clément, F., & Harris, P. L. (2004). Trust in testimony: Children’s use of true and false statements. Psychological Science, 15(10), 694–698. View source
- Koenig, M. A., & Harris, P. L. (2005). Preschoolers mistrust ignorant and inaccurate speakers. Child Development, 76(6), 1261–1277. View source
- Sperber, D., Clément, F., Heintz, C., et al. (2010). Epistemic vigilance. Mind & Language, 25(4), 359–393. View source
- Harris, P. L., Koenig, M. A., Corriveau, K. H., & Jaswal, V. K. (2018). Cognitive foundations of learning from testimony. Annual Review of Psychology, 69, 251–273. View source