A person asks an AI system a difficult question.
The answer is competent, organized and fast.
The next question is not whether the answer is useful.
It is whether the person still owns enough of the inquiry to know what to do with it.
Who decided what counted as the problem? Who chose the evidence? Who noticed what was missing? Who set the threshold for confidence? Who remains responsible when the answer travels into the world?
These are questions of epistemic agency.
Definition
In this Knowledge Library, epistemic agency means:
the capacity to take an active, responsible role in forming questions, seeking and evaluating evidence, constructing interpretations, revising beliefs and shaping the conditions under which knowledge is produced.
Agency here does not mean solitary independence.
Human knowing is deeply social. We depend on teachers, instruments, institutions, colleagues, databases and experts. Epistemic agency concerns how a person or group participates inside that dependence.
The opposite of agency is not collaboration.
It is becoming only the endpoint of somebody else’s epistemic process.
The term already has a research history
Epistemic agency is not an Institute invention.
In learning sciences and science education, researchers have used the term to describe learners’ active role in directing inquiry, constructing knowledge and shaping the norms and structures through which a community learns.
Jianwei Zhang’s 2022 study, for example, examines how Grade 5 students enacted epistemic agency by co-constructing inquiry directions and flexible group structures during sustained collaborative knowledge building.
That work makes an important point: agency can be collective.
A classroom does not have to choose between teacher control and isolated student independence. Students can take responsibility for how inquiry unfolds together.
Epistemic agency is broader than “having an opinion”
Agency does not mean that every belief deserves equal authority because it is personally chosen.
A person who refuses evidence can be highly self-directed while exercising poor epistemic agency.
The capacity therefore includes answerability.
Questions are answerable to evidence. Interpretations are answerable to method and context. Confidence is answerable to uncertainty. Participation in a knowledge community is answerable to other people who can challenge, correct and extend what one knows.
Epistemic agency is freedom with epistemic obligations attached.
It is also broader than critical thinking
Critical thinking concerns the quality of analysis, evaluation and inference.
Epistemic agency asks who is doing the epistemic work and who has meaningful control over it.
A person may critically evaluate a set of options while another system determines the problem frame, retrieves the evidence and excludes alternatives.
Conversely, someone may choose the question and direct an inquiry while using weak reasoning.
Agency and critical thinking support one another, but they are not substitutes.
From individual skill to the architecture of inquiry
Epistemic agency is often discussed as if it lived entirely inside a person.
But environments distribute agency.
A classroom assignment that specifies the question, sources, method, acceptable answer form and evaluation criteria leaves less room for learner-directed inquiry than one that allows students to formulate and refine problems.
A workplace dashboard can similarly pre-select what counts as relevant performance. A search engine ranks candidate sources. A recommendation system determines what enters view. A generative model proposes the categories through which a problem is described.
The person may still be “deciding,” but the decision space has already been shaped.
This is why epistemic agency belongs in institutional and technological design, not only individual education.
Generative AI makes the construct newly visible
Wu and colleagues’ 2025 conceptual article on human epistemic agency in partnership with generative AI argues that AI-rich learning should preserve an active human role rather than reducing learners to consumers of generated knowledge.
The paper is conceptual, not an intervention proving one framework’s effectiveness.
Its value is diagnostic. It identifies a real design problem: generative systems can participate in knowledge production so fluently that users may stop noticing which epistemic operations have moved outside themselves.
This does not mean AI must be excluded.
The more useful question is how human agency is configured within the partnership.
Does the learner set goals, ask follow-up questions, seek counterevidence and revise conclusions? Or does the model define the problem, generate the answer, explain it and verify itself?
The latter workflow can be efficient while leaving the user epistemically thin.
Agency can be supported by friction
Some forms of friction are waste.
Others preserve cognition that matters.
An AI interface that asks the user to state their current hypothesis before revealing a recommendation may create useful epistemic friction. A system that shows competing interpretations rather than one polished answer can preserve comparison. A source view that exposes provenance can keep verification within the human loop.
These design moves should not be assumed effective without testing.
They illustrate the principle that efficiency and agency are different objectives.
A system optimized only to minimize effort can remove exactly the operations through which people learn to judge.
Confidence is part of agency because uncertainty guides action
Epistemic agency includes knowing when to defer.
The mature epistemic agent is not the person who insists on personal verification of everything.
Modern knowledge makes that impossible.
Agency includes choosing when expert reliance is rational, when a claim requires primary-source inspection, when one lacks the competence to judge, and when uncertainty should delay action.
This is where metacognition enters.
Fischer and Huff’s 2025 argument that metacognitive accuracy matters for science communication is relevant because people act not only on knowledge but on confidence in that knowledge.
A person who cannot locate their uncertainty may appear agentic while being poorly calibrated.
Shared agency needs accountability, not blurred responsibility
Human-AI systems are often described as partnerships.
That language can obscure responsibility.
An AI model does not carry professional, legal or moral accountability in the same way a teacher, physician, engineer, journalist or public official does.
If decision processes become distributed, institutions need clarity about who remains responsible for epistemic checks.
“Human in the loop” is not enough if the human sees only the final output and lacks time, evidence or authority to challenge it.
Meaningful epistemic agency requires more than ceremonial approval.
Measurement is difficult because agency is relational
There is no single gold-standard scale that captures the Institute’s full definition.
Researchers may infer epistemic agency from observable participation in question generation, evidence seeking, explanation, knowledge-building discourse or control over inquiry decisions.
Self-report can capture perceived agency but not necessarily enacted agency.
Task designs can capture local behaviours while missing institutional constraints.
Because agency depends partly on available choices, identical behaviour can mean different things in different environments.
A student who follows a prescribed source list may be passive in one task and appropriately selective in another where the list was co-designed.
Measurement must therefore specify what choices were available, not only what choices were made.
European digital competence contains a related autonomy thread
DigComp 3.0 repeatedly links digital competence to autonomy, critical evaluation and the ability to identify needs and select tools appropriately.
It does not define “epistemic agency” in the Institute’s sense.
The relevance is structural: European digital competence is not framed as mere button knowledge. It includes exercising choice, evaluating information, staying informed about digital developments and building capacity to operate autonomously in digital environments.
This provides a legitimate European bridge while preserving conceptual boundaries.
The institutional problem: who is allowed to know differently?
Epistemic agency can be constrained by authority structures even without technology.
If employees learn that questioning assumptions harms promotion, they may stop surfacing counterevidence. If students are rewarded only for reproducing the teacher’s framing, inquiry becomes performance. If researchers face incentives that punish null results, the knowledge system itself narrows what gets pursued.
Agency therefore includes the social permission to revise shared understanding.
Institutions that want epistemic agency need correction channels, dissent protection and spaces where uncertainty can be expressed without loss of status.
Agency is visible in question formation
The most obvious sign of epistemic agency is not disagreement.
It is the ability to formulate a question that has not already been handed to you.
Question formation determines what evidence will be sought, which alternatives are imaginable and what counts as a satisfactory answer. A learner who can only respond to teacher-generated questions may display knowledge without exercising much epistemic agency. A professional who can only select among options produced by a dashboard may retain decision authority while losing authority over the problem definition itself.
This distinction matters in AI-assisted work because language models are unusually good at supplying frames. They can propose categories, interpretations and next questions before the user has articulated their own. That can be helpful, especially for novices. It can also narrow the field silently.
One practical safeguard is to preserve a short period in which the human states the problem, uncertainty and desired evidence before asking the system to structure it. This is not a universal rule. It is a way of preserving evidence that the inquiry still has a human origin point.
Epistemic agency is relational
Agency is sometimes described as if it were independence from other people.
Knowledge rarely works that way.
Researchers depend on instruments, journals and colleagues. Patients depend on clinicians. Citizens depend on journalists and public institutions. Students depend on teachers. Professionals depend on standards, databases and specialist expertise.
The relevant question is not whether dependence exists. It is whether the person or group can participate intelligently in that dependence.
Relational epistemic agency includes knowing what to ask an expert, what evidence to request, when a consensus deserves weight, how to notice conflicts of interest and how to act when competent authorities disagree.
This is why agency can increase through well-designed scaffolding. Support does not necessarily reduce agency. Support can make a person capable of taking part in a practice that would otherwise remain inaccessible.
The danger begins when scaffolding becomes opaque substitution and the learner cannot tell which parts of the inquiry they are no longer performing.
From classrooms to organizations
The education literature gives useful conceptual examples because classrooms make the distribution of epistemic authority visible.
Who gets to pose a problem? Who can challenge an explanation? Who decides when the class knows enough? Who records the shared model? Who may revise it?
The same questions apply inside organizations.
A company can centralize epistemic authority in a small analytics function while telling everyone else to be “data driven.” A public institution can invite consultation after the relevant categories and options have already been fixed. A product team can use user research while allowing only metrics already present in the dashboard to count as evidence.
Epistemic agency is therefore partly an organizational-design question.
Distributed expertise requires channels through which local knowledge can alter collective conclusions. Otherwise participation becomes decorative.
Agency without epistemic isolation
The Institute rejects a romantic version of agency in which a mature person must verify everything personally.
That would be impossible and often irrational.
Human knowledge works through justified trust. The challenge is to make trust discriminating and revisable.
A person with epistemic agency can say, “I am relying on this specialist because of these reasons, within these limits, and I know what evidence would make me reconsider.”
That is different from uncritical deference.
It is also different from reflexive suspicion.
This middle position is crucial for the AI age. External intelligence will increasingly participate in research, drafting, search, translation, coding and decision support. The human task is not to preserve autonomy by refusing assistance. It is to preserve enough epistemic structure that assistance remains supervised, contestable and attributable.
Indicators without pretending to have a scale
Because epistemic agency is broad, the Institute does not propose a single score.
Researchers can instead observe domain-specific indicators: who initiates questions, who selects evidence, whether alternatives are generated, whether sources are interrogated, whether uncertainty is represented, whether revisions occur and whether participants can explain why a conclusion changed.
Each indicator has limitations. Speaking more often does not equal greater agency. Challenging authority can be performative. A quiet participant may have substantial internal control over their judgment.
The value of indicators is diagnostic, not reductive.
They make the distribution of knowledge-making visible enough to ask whether a system is developing human capacity or merely extracting compliant outputs.
What the evidence does not show
The literature does not support treating epistemic agency as one universally standardized psychological trait.
The Zhang classroom study shows how agency can be enacted in one sustained inquiry context. It does not establish a general population effect.
The 2025 human-GenAI framework is conceptual and should not be presented as validated intervention evidence.
The CHI critical-thinking survey does not prove that generative AI causes loss of agency.
European competence frameworks are normative and educational reference systems, not causal demonstrations.
And greater personal control does not automatically produce better beliefs.
Agency without evidence discipline can become stubbornness.
What this changes for Inner Technology
The Institute uses epistemic agency to name a design question that often disappears behind performance metrics:
Which parts of knowing must remain actively exercised by the human if the goal is not only a correct output, but the continued development of judgment?
That question can be asked at several levels.
For a learner: Did I formulate the question and test the answer?
For a professional: Can I inspect the evidence behind the recommendation?
For a team: Can disagreement change the shared model?
For an institution: Are people rewarded for detecting error?
For a society: Do citizens have the competence and access needed to interrogate increasingly automated information systems?
These are Institute synthesis questions built on established adjacent literatures.
Before the prompt
The person returns to the AI interface.
This time, before typing, they write down what they think the problem is.
They note what evidence would change their mind.
They decide which parts they are willing to delegate and which they must still be able to explain themselves.
The answer may arrive just as quickly as before.
But the person meeting it is no longer only a consumer.
They are still inside the act of knowing.
Research references
- Zhang J. Epistemic agency for costructuring expansive knowledge-building practices. Science Education. 2022;106. https://doi.org/10.1002/sce.21717
- Wu JY, Lee YH, Chai CS, Tsai CC, et al. Strengthening Human Epistemic Agency in the Symbiotic Learning Partnership With Generative Artificial Intelligence. Educational Researcher. 2025. https://doi.org/10.3102/0013189X251333628
- Lee HP, Sarkar A, Tankelevitch L, et al. The Impact of Generative AI on Critical Thinking. CHI 2025. https://doi.org/10.1145/3706598.3713778
- Fischer H, Huff M. Metacognition as a target of science communication. Nature Reviews Psychology. 2025;4:554–555. https://doi.org/10.1038/s44159-025-00478-2
- Cosgrove J, Cachia R. DigComp 3.0: European Digital Competence Framework, fifth edition. European Commission Joint Research Centre. 2025. https://doi.org/10.2760/0001149