No serious person knows alone.
The water is safe because somebody tested it. The aircraft flies because teams maintained standards the passenger cannot personally inspect. A medical guideline compresses years of distributed work. A search engine returns sources the user did not index. An AI system can summarize literature no individual could hold in working memory.
Dependence is not an exception to knowledge.
It is the ordinary condition of modern knowledge.
Epistemic autonomy asks what responsible judgment can mean inside that dependence.
Definition
For this Knowledge Library:
epistemic autonomy is the capacity to form and revise beliefs through one’s own accountable judgment while making appropriate use of testimony, expertise, institutions and tools.
The crucial word is accountable.
Autonomy does not require personally reproducing every experiment or deriving every conclusion from first principles. It requires retaining enough evaluative control to decide when reliance is warranted, what confidence is justified and what could trigger revision.
This makes epistemic autonomy different from self-sufficiency.
The self-sufficient knower is a myth
Philosophical work on testimony has long challenged the picture of a rational person who should believe only what they can independently verify.
Elizabeth Fricker’s work on testimony and epistemic autonomy examines how discriminating trust in others can coexist with responsibility for belief. Jesús Vega Encabo similarly distinguishes autonomy from self-sufficiency, arguing that epistemic dependence need not eliminate a hearer’s own contribution to knowledge.
These arguments matter because modern expertise makes total independence irrational.
Refusing all testimony would not produce a more autonomous citizen. It would produce someone unable to benefit from the division of epistemic labour.
Dependence can support autonomy when the terms of reliance are visible
Autonomy survives reliance when a person can ask questions such as:
Why is this source relevant?
What makes the expert competent in this domain?
What is the evidence lineage?
Are independent sources converging?
What uncertainty remains?
What would make me seek a second opinion?
These questions do not eliminate dependence. They make dependence discriminating.
A person can rationally defer while still owning the conditions under which they defer.
Expert autonomy has system-level value
Finnur Dellsén’s work on the epistemic value of expert autonomy argues that independence among experts can improve the informational value of joint testimony when their judgments are not merely copies of one another.
The point is subtle.
Consensus among ten genuinely independent expert assessments carries different information from ten statements generated from the same upstream source.
This matters well beyond philosophy.
In science, duplicated dependence can create an illusion of convergence. In journalism, multiple outlets may repeat the same wire report. In AI systems, several generated answers may trace back to overlapping training data or the same search index.
Autonomy therefore has an ecological dimension: diversity of genuinely independent inquiry pathways matters.
Epistemic autonomy and epistemic agency are related but distinct
Epistemic agency concerns active participation in forming questions, seeking evidence and shaping inquiry.
Epistemic autonomy concerns whether a person’s final epistemic position remains responsibly theirs rather than merely inherited from authority or automation.
One can have agency without autonomy. A person may actively search and compare sources but ultimately defer because social pressure makes dissent impossible.
One can also retain autonomy with limited agency. A patient may not design a clinical study but can still make a responsible decision about which expert advice to trust and what uncertainties matter to them.
The distinction is useful because “agency” and “autonomy” are often collapsed into a general language of empowerment.
AI makes source dependence less visible
Generative AI complicates autonomy because it can hide the social and documentary structure of knowledge behind one fluent interface.
When a model produces a synthesized answer, users may not know which sources influenced it, which disagreements were flattened or where uncertainty entered.
The system can therefore create a feeling of direct access to knowledge while increasing actual epistemic dependence.
This is not unique to AI. Encyclopedias, search engines and expert reports also compress complexity.
The difference is conversational fluency and adaptive confidence. The system can answer follow-up questions in a way that resembles an accountable interlocutor even when its provenance is incomplete.
Epistemic autonomy in this setting requires new habits of provenance checking and confidence management.
Autonomy does not mean distrusting experts
A recurring misunderstanding is that autonomy requires resisting authority.
That can turn autonomy into anti-expertise.
Relevant expertise is often the reason rational deference is warranted.
The autonomous move is not to treat all claims as equally open to amateur adjudication. It is to distinguish deference from submission.
Deference says: this person or institution has relevant competence and evidence I cannot fully reproduce, so I will give their judgment substantial weight.
Submission says: because they are authoritative, no further conditions of trust matter.
The first can be rational. The second abandons epistemic responsibility.
Institutions can either preserve or absorb human judgment
Consider two decision systems.
In the first, an employee receives an algorithmic recommendation together with the input variables, uncertainty, provenance and a documented override process.
In the second, the employee receives a score and is told that deviations require managerial approval.
Both have a human technically “in the loop.”
Only one gives that human meaningful epistemic autonomy.
Autonomy therefore depends on institutional permissions as well as individual skill.
A person cannot exercise judgment they are not allowed to use.
European digital competence contains a practical autonomy principle
DigComp 3.0 describes digital competence as combining knowledge, skills and attitudes for confident, critical and responsible engagement with digital technologies.
Its proficiency model explicitly includes increasing autonomy in handling tasks, and its competence framework includes evaluating digital information, asserting rights, exercising choice and building capacity to operate autonomously in digital environments.
This is not the same philosophical construct as epistemic autonomy.
It provides a concrete European policy context in which autonomy is treated as part of mature digital competence rather than as technological abstinence.
The AI Act’s literacy obligation is relevant but not sufficient
As of September 2026, Article 4 of the EU AI Act requires providers and deployers to take measures supporting AI literacy for staff and others operating AI systems on their behalf. The Commission states that the obligation has applied since 2 February 2025 and that national supervision and enforcement began in August 2026; amendments entering into force in July 2026 retained the obligation without prescribing one specific “sufficient” individual level.
This matters for autonomy because institutions cannot responsibly delegate consequential work to AI while leaving users unable to understand basic capabilities and limitations.
But compliance training alone does not guarantee epistemic autonomy.
A person may know what a model is and still lack access, time or authority to question its recommendation.
Measurement requires relational evidence
There is no single validated instrument that measures the Institute’s whole definition of epistemic autonomy.
Philosophical analysis clarifies norms of testimony and judgment. Organizational research may examine discretion, override rights or reliance. Educational work may study how learners justify source choice and revise beliefs.
Self-report can capture perceived independence but may confuse confidence with autonomy.
Behavioural tasks can show willingness to challenge advice but are highly context-specific.
A useful assessment therefore needs to describe the relationship of dependence being studied: who or what is relied on, what information the person can inspect, what alternatives are available and what happens if they disagree.
Autonomy can fail in opposite directions
One failure mode is over-deference.
The person treats an expert, institution or AI system as epistemically final.
Another is under-deference.
The person rejects warranted expertise because accepting dependence feels like surrender.
Both can be autonomy failures.
The first abandons judgment. The second mistakes self-sufficiency for responsibility.
Mature epistemic autonomy has to tolerate the discomfort of saying both “I cannot know this alone” and “I am still responsible for how I rely on those who know more.”
Autonomy is compatible with justified dependence
The strongest conceptual mistake is to treat epistemic autonomy as self-sufficiency.
No contemporary person can independently reproduce the evidence behind most of what they reasonably know. We rely on aviation engineers, epidemiologists, historians, electricians, statisticians, archivists and countless infrastructures we cannot personally audit end to end.
The philosophical literature on testimony makes this unavoidable. The question is not whether people depend on others. It is what kind of dependence is epistemically responsible.
Dellsén’s work on expert autonomy is useful because it shows that some degree of independence among experts can improve the reliability of joint testimony. If experts merely copy one another, apparent agreement contains less independent evidential value than genuinely convergent judgment.
That insight transfers cautiously to AI-mediated knowledge work. Ten generated answers from systems sharing the same training lineage or retrieval source are not equivalent to ten independent evidential confirmations.
Autonomy has a social infrastructure
A person cannot exercise meaningful epistemic autonomy if relevant information is inaccessible, explanations are withheld or disagreement carries disproportionate punishment.
This makes autonomy partly institutional.
Transparent procedures, rights to explanation, access to sources, plural expertise and correction mechanisms can widen the space in which individuals make informed judgments. Conversely, an organization can formally leave a decision to an employee while controlling every informational input that makes the decision intelligible.
Autonomy therefore has at least two layers:
personal capacity, including judgment, metacognition and willingness to revise; and environmental affordance, including access, contestability and informational plurality.
Neither substitutes for the other.
A highly capable person can be constrained by an opaque system. A transparent system does not guarantee that every participant has the capacity to use it well.
The danger of manufactured independence
Digital environments sometimes produce the feeling of independence without the substance.
A user can choose among hundreds of sources that all repeat the same upstream report. Search results can look plural while ranking systems shape which options are visible. AI systems can produce multiple “perspectives” that are generated from overlapping patterns rather than independent investigations.
Epistemic autonomy therefore requires attention to source independence, not just source count.
This is one reason provenance becomes central. Where did the information originate? Which sources depend on which others? Has one claim been copied across an ecosystem until repetition resembles consensus?
Autonomy is strengthened when the person can distinguish diversity of presentation from diversity of evidence.
Autonomy and authority
Experts matter precisely because knowledge is unevenly distributed.
The mature response to expertise is neither submission nor dismissal.
It is calibrated reliance.
Questions of competence, track record, domain relevance, conflicts of interest, consensus and evidence quality all matter. So does the listener’s own competence to evaluate those dimensions.
In some domains, the most autonomous judgment is to defer to a strong expert consensus while understanding why. In others, expert disagreement may justify suspension of judgment.
This is why epistemic autonomy should not be measured by frequency of disagreement with authority.
Independent thought can lead to agreement.
The key issue is whether the agreement remains connected to reasons rather than mere obedience.
AI and the preservation of exit
One practical dimension of autonomy is the ability to leave or override an epistemic pathway.
Can a user inspect sources outside the AI interface? Can they recover the original documents? Can a recommendation be challenged? Can a decision proceed without the model? Can the person know when automation has influenced what they were shown?
These are design questions about epistemic exit.
A system may be highly convenient yet autonomy-reducing if it becomes the only practical route to information. Conversely, an assistant can support autonomy by making evidence more accessible, highlighting uncertainty and helping users compare alternatives.
The Institute therefore treats AI neither as inherently autonomy-eroding nor automatically empowering.
The effect depends on design, skill, context and the distribution of decision rights.
What a strong autonomy claim would require
A future empirical programme on epistemic autonomy would need more than self-report.
It would need tasks that capture independent question formation, source comparison, willingness to override assistance, appropriate reliance on expertise and sensitivity to source dependence. It would need longitudinal evidence showing whether those capacities transfer across contexts.
Until such measures mature, broad claims about “increasing epistemic autonomy” should remain cautious.
The concept is valuable because it organizes a real problem.
Its measurement remains less settled than the confidence with which the phrase is sometimes used.
Autonomy can include suspension
A subtle form of autonomy is the ability not to decide yet.
Digital systems often reward rapid response, while organizations treat decisiveness as competence. But when evidence is weak or authorities genuinely disagree, suspension of judgment can be the most autonomous position available.
That stance is active rather than passive. It records what is unknown, what evidence is missing and what threshold would justify movement.
The capacity to remain undecided without surrendering inquiry is especially important when AI systems can always generate one more plausible answer.
What the evidence does not show
The literature does not provide one settled definition of epistemic autonomy across philosophy, education and technology research.
Dellsén’s argument concerns the epistemic value of expert autonomy under specific theoretical conditions; it is not a field experiment on organizations.
Testimony theory does not provide a simple checklist for whom citizens should trust.
DigComp 3.0 is a competence framework, not causal evidence that autonomy training improves decision quality.
The AI Act establishes legal and organizational obligations around literacy, not proof that those obligations preserve human judgment.
And independent judgment is not automatically accurate judgment.
What this changes for Inner Technology
The Institute treats epistemic autonomy as a design criterion for human capacity in systems of dependence.
The practical question becomes:
Can the person still understand why they are relying, what they are relying on and what conditions would justify changing course?
That question can guide education, AI interface design, professional governance and institutional review.
It shifts autonomy away from the fantasy of the isolated thinker and toward something more demanding: accountable participation in a distributed knowledge world.
The borrowed map
A person unfolds a map they did not draw.
They trust the surveyors. They could not have measured the entire terrain themselves.
But they still need to know where they are going, which route fits the conditions and when the map may be out of date.
That is the shape of epistemic autonomy.
Not refusing the map.
Remaining capable of navigating with it.
Research references
- Fricker E. Testimony and Epistemic Autonomy. In: Lackey J, Sosa E, eds. The Epistemology of Testimony. Oxford University Press. 2006:225–250. https://doi.org/10.1093/acprof:oso/9780199276011.003.0011
- Vega Encabo J. Epistemic merit, autonomy, and testimony. THEORIA. 2008;23(1):45–56. https://doi.org/10.1387/theoria.5
- Dellsén F. The epistemic value of expert autonomy. Philosophy and Phenomenological Research. 2020;100(2):344–361. https://doi.org/10.1111/phpr.12550
- 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
- European Commission. AI Literacy Questions & Answers; AI talent, skills and literacy. Current pages reviewed 23 September 2026. https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers