Rethinking Epistemic Agency
in Knowledge Building

in the Age of AI

Bodong Chen, Katherine Guangji Yuan, Chew Lee Teo, Monica Ong,
Lydia Cao, Xueqi Feng

KB Connects, April 2026

Agenda

Introduction & framing
Tracing Epistemic Agency (Chen)
Vibe-coding Projects for Secondary Students (Yuan, Teo & Ong)
Teachers as Inventors Through KB and Vibe Coding (Cao, Feng, Teo)
Open Discussion

Continuing Dialogue

KB Connects Feb 2026: The Centrality of Epistemic Agency in LLM–Student Knowledge-Building Partnerships (Earl Woodruff & Jim Hewitt)

KBSI 2026 Symposium on Epistemic Agency:

Orchestrating the “Dance of Agency” in Science Education (Tao)
Vibe-coding Projects for Secondary Students (Yuan, Teo & Ong)
Promoting Epistemic Agency via AI (Cacciamani & Khanlari)
Manifesting New Knowledge Building Attitudes (Weng & Horng)
Shared Epistemic Agency with Generative AI (Oshima & Naganuma)
Promote Epistemic Agency via Meta-discourse (Yu, Sun, & Chan)

Guiding Questions of the KBSI Symposium

What do we mean by epistemic agency in Knowledge Building? Which tenets of epistemic agency should we examine and nurture?
How is epistemic agency distributed among students, teachers, intelligent technologies, and other agentic actors in KB?
How should we design purposefully to advance key tenets of epistemic agency?

With AI, each question gets a new actor and a new design surface.

Your ideas about epistemic agency

What words come to mind?

If you were asked to define epistemic agency in one or two phrases, which words feel essential?

Drop them on the board.

All are improvable ideas.

2:00

https://www.tldraw.com/f/-xioWzB3sH2VKi8xWtpV9?d=v0.0.1468.936.page

Tracing Epistemic Agency in Knowledge Building

Bodong Chen

Epistemic Agency in KB

Scardamalia & Bereiter (1991)

“Teacher C … aims to give children a higher level of agency in the knowledge-building process.”

“Giving children a higher level of agency … would require providing a new structure that supports children in growing into a question-asking role.”

Two threads:

agency as a capacity the child grows into
agency as something a structure must support

Epistemic agency as capacity

Simply Put:

Every time we act in pursuit of an epistemic goal, we express our epistemic agency.

But it’s not choosing to believe whatever one wants.

The capacity to actively engage in and control one’s own belief formation and reflection processes, rather than passively accepting information.

— Ahlstrom-Vij (2013)

Epistemic agency has many faces

Aspect	What the agent does
Reasons-responsiveness	Form beliefs in light of reasons
Active endorsement	Actively reason your way to a belief
Apt judgment	Achieve truth through competence, not luck
Epistemic dependence	Rationally rely on expertise beyond one’s own
Group agency	Collective understanding, intention, responsibility

The old voluntarist picture — agency as believing at will — was dismantled by Bernard Williams and then by Hieronymi (2008) and Boyle (2009), who showed that belief is constitutively answerable to reasons: you cannot simply decide to believe, because belief already owes itself to evidence and argument. Rows 1–2 draw on their work: Hieronymi on reasons-responsiveness, Boyle on how actively reasoning to a belief lets one own it as one’s own. Row 3 follows Sosa (2015): an apt judgment is true because it manifests the agent’s competence — not because of luck. Rows 4–5 shift to social epistemology: Hardwig (1985) on the rationality of deferring to experts; Fricker (2007) on contesting testimony as an agentive (and politicized) act; Tollefsen (2015) and Lackey (2021) on genuine group-level agency and knowledge. KB tends to foreground only the managerial face — who controls inquiry — while leaving these richer dimensions largely implicit.

Two kinds of agency

Evaluative control

Agency in forming the belief itself

settle what is true
own a judgment
revise in light of reasons
stand answerable to reasons

Managerial control

Agency over the work around forming the belief

choose what to read
decide a strategy
organize collaboration
evaluate progress

This is Hieronymi’s distinction. Managerial control is the indirect agency we exercise over attitudes — taking a course, looking up evidence, deciding to keep thinking. Evaluative control is the constitutive agency we exercise in believing — settling for ourselves whether p, in light of reasons. The crucial point is the second: we are responsible for what we believe even though we cannot believe at will, because belief is constitutively answerable to reasons. Hieronymi also shows that managerial control necessarily involves evaluative control at two points — when one forms the intention to manage, and when the managed attitude is actually formed (pp. 153–154). So the two are not cleanly separable. Applied to KB: even the managerial activities KB foregrounds — deciding what to pursue, judging what counts as progress — must involve evaluative control. KB doesn’t escape it; it just doesn’t theorize it. That gap is what we need to design for.

Epistemic agency as capacity in KB

Scardamalia (2002):

Epistemic agency — students taking on “high-level decision-making about long-range goals, inquiry strategies, social arrangements, and criteria of success.”

High-level agency emphasizes the managerial:

who decides
what to pursue
how to organize
what counts as success

Evaluative control – Worth more attention?

KB’s key commitment (Bereiter, 2002): shift students from belief mode — ideas as claims to accept or reject — into design mode — ideas as artifacts to be improved. That move is sound and much needed!

But what’s under-developed: students’ evaluative control over beliefs (Hieronymi, 2008) and the intellectual virtues that support it, e.g., curiosity, humility, attentiveness.

AI heightens the need for evaluative control. LLMs produce confident prose but through opaque processes. Evaluative agency and associated intellectual virtues are even more important.

Bereiter’s distinction in Education and Mind: belief mode treats ideas as claims to be accepted or rejected; design mode treats them as conceptual artifacts to be improved. KB’s signature contribution has been pushing students into design mode — and that has been the right move against schools that default to surface-level belief work. What has been under-developed alongside it is the work Hieronymi calls evaluative control: the constitutive agency we exercise in settling, on our own grounds, what we believe. KB does evaluative work — judging revisions, weighing rise-above options — but the truth-aimed, individual-belief-settling kind has not been at the center of the practice. Responsibilist virtues from Zagzebski and Baehr — curiosity, humility, carefulness, intellectual courage — are the dispositions that hold evaluative control as a habit, and they too get less explicit attention in the KB literature than the design-mode practices. AI changes the stakes: large language models produce stylistically confident output without exposing reasoning, and without evaluative agency as a stable disposition, students have little to push back with. The gap is now urgent.

Agency in environments

Agency is simultaneously enabled and constrained by the environment.

values and norms are built into tools
they differ in epistemic affordances — representation, interaction, traceability, socialization…
scaffolds may either cultivate human agency, or undermine it (e.g., hostile scaffolding)

Three implications of taking the agent–environment coupling seriously. First, tools embed values: an artifact’s design has already committed to which moves are easy and which are hard before any user picks it up. Second, environments differ in their epistemic affordances. What gets represented, how interaction is structured, how credibility surfaces, whether reasoning is traceable — these are not minor variations. An argument-map view and a chat thread are not the same epistemic environment; they make different actions possible and foreclose others. Third, scaffolds either cultivate the capacities they support or undermine them. KF scaffolds repeated for years cultivate; engagement feeds and default-answer AI substitute. The question is whether the environment we put students in supports the agency we are trying to cultivate, or works against it.

AI can do either when embedded in an environment

Cassinadri (2024) — contextual virtue epistemology

A calculator in early arithmetic and a calculator in advanced astrophysics are not the same artifact.

The educator’s design judgment is what makes one cultivate epistemic agency and the other substitute for it.

Whether AI cultivates or substitutes for epistemic agency depends on:

who produces the idea — student or AI?
whether productive frictions (struggle, forgetting) are preserved
whether humans get to traverse complex idea landscapes
whether AI output contributes to community knowledge through discourse

Scenario 1 — AI as oracle

Class: “AI, which of our notes is most promising?”

AI: Top 3 promising notes:
1. Note #42 (0.91) 2. Note #17 (0.88) 3. Note #29 (0.84)

Students accept the ranking.

The model ranks. Students select.

judgment largely shifts to AI model
AI output enters discourse as conclusions, disguised as suggestions
the capacity (judging promisingness) atrophies

AI output as conclusions – which narrow imagination.

Scenario 2 — AI outputs new material

“AI, convert this transcript into an argument map
to help us see what we know and on what basis.”

Argument map showing claims, evidence, and counter-claims as nodes and links

The model surfaces. Students judge.

judgment stays with the community
AI output enters discourse as new material
struggle, disagreement, and friction are preserved

AI output enters as material — which offers new representations for ideas.

Beyond the classroom

Students step into a wider world of epistemic environments.

Libraries · museums
slow, curated, citation-rich

Skate parks
agency-first, failure-normed

Social media feeds
engagement-optimized

AI chatbots
helpful, but frictionless

Consider broader spaces where epistemic agency can be excercised and where intellectual virtues can be nurtured.

Students step from Knowledge Forum into a much wider world of epistemic environments. On one side: libraries, museums, reading groups, considered conversation around a dinner table — physical environments that have been preserving intellectual virtues for centuries, often more reliably than the digital spaces we have built. On the other side: recommendation feeds and default-answer AI — environments designed for engagement and frictionless delivery, which scaffold hostilely in Timms and Spurrett’s sense and substitute for the very capacities we are trying to build. The point is not digital-bad, physical-good. It is that the function of the environment matters, and KB’s design vocabulary — interposing community discourse, structuring evaluative control, building toward dispositions across years — should travel beyond the LMS into the digital and physical environments where students actually form beliefs.

Work ahead

Be specific about the agency we cultivate. Consider multiple faces of epistemic agency in KB practice.
Treat each tool feature as a design decision. Every scaffold sits somewhere between benign and “hostile”.
Empower humans to build their epistemic environment. Agency is contextual. Entrust teachers and students with the agency in shapring their environments – within and beyond their classroom.

Vibe-coding Projects for Secondary Students

Yuan, Teo & Ong

Teachers as Inventors Through KB and Vibe Coding

Cao, Feng, Teo

Open Discussion

Questions to consider:

Collective reflection — Looking at the words you put on the board: which face of agency were you reaching for?
Different paths — In your KB practice, do you see instances of AI cultivating or substituting epistemic agency? What leads to this?
Beyond the LMS — Which epistemic environments are shaping your students this week? Is epistemic agency scaffolded?
Design moves — If you integrate AI in KB practice, which areas of epistemic agency will you protect? How?

Key References

Ahlstrom-Vij, K. (2013). Epistemic Paternalism. Palgrave.
Baehr, J. (2011). The Inquiring Mind. Oxford UP.
Bastani, H., et al. (2025). Generative AI without guardrails can harm learning. PNAS.
Battaly, H. (2018). Can closed-mindedness be an intellectual virtue? RIP Suppl., 84.
Bereiter, C. (2002). Education and Mind in the Knowledge Age. Lawrence Erlbaum.
Boyle, M. (2009). Active belief. Canadian Journal of Philosophy, 39 (Suppl. 35).
Cassinadri, G. (2024). ChatGPT and the technology-education tension. Philosophy & Technology, 37(14).
Chan, C., Burtis, J., & Bereiter, C. (1997). Knowledge building as a mediator of conflict in conceptual change. Cognition and Instruction, 15(1).
Fricker, M. (2007). Epistemic Injustice. Oxford UP.
Hardwig, J. (1985). Epistemic dependence. Journal of Philosophy, 82.
Heersmink, R. (2018). A virtue epistemology of the internet. Social Epistemology, 32(1).
Hieronymi, P. (2006). Controlling attitudes. Pacific Philosophical Quarterly, 87.
Hieronymi, P. (2008). Responsibility for believing. Synthese, 161.
Lackey, J. (2021). The Epistemology of Groups. Oxford UP.
Nguyen, C. T. (2020). Echo chambers and epistemic bubbles. Episteme, 17(2).
Pritchard, D. (2012). Anti-luck virtue epistemology. Journal of Philosophy, 109(3).
Scardamalia, M. (2002). Collective cognitive responsibility. In B. Smith (Ed.), Liberal Education in a Knowledge Society.
Scardamalia, M., & Bereiter, C. (1991). Higher levels of agency for children in knowledge building. Journal of the Learning Sciences, 1(1).
Sosa, E. (2015). Judgment and Agency. Oxford UP.
Sterelny, K. (2010). Minds: extended or scaffolded? Phen. Cogn. Sci., 9(4).
Timms, R., & Spurrett, D. (2023). Hostile scaffolding. Philosophical Papers, 52(1).
Tollefsen, D. (2015). Groups as Agents. Polity.
Zagzebski, L. (1996). Virtues of the Mind. Cambridge UP.