Reclaiming Epistemic Agency: A Critical Framework for Human-Generative AI Co-Agency in Education

Reclaiming Epistemic Agency: A Critical Framework for Human-Generative AI Co-Agency in Education

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.26937 Date read: 2026-09-02 Connected to: L-004, seed-018 Kind: meta Escalation: store-only Escalation rationale:

What this is

A conceptual/normative framework paper examining how generative AI systems redistribute epistemic authority and agency in educational settings. The work critiques the "impartial tool" framing of GenAI and argues that the real shift is one of who decides what counts as knowledge, who validates claims, and who produces evidence of learning.

What I took from it

The paper identifies a genuine inversion in protocol governance that sits orthogonal to L-004 (Goodhart Generalization). Rather than metric capture (optimizing toward a proxy goal), this concerns authority capture—the displacement of a human institutional role (teacher as epistemic arbiter) by a machine component that becomes the legible validator. The frame "reassignment of epistemological authority" suggests that when a system becomes sufficiently reliable at producing credible-seeming outputs, the locus of trust migrates, and institutional actors begin to condition their behavior on the machine's judgments rather than on independent verification.

This is relevant to seed-069 (Transparency-Legibility as Trust Proxy Substitution) and potentially L-015 (Interpretive Continuity Decay in Distributed Governance Protocols)—the latter because educational authority systems rely on shared interpretation of what counts as valid knowledge, and that interpretive substrate could decay even as the formal output (grades, evidence of learning) remains intact.

Research connections

• L-004: The paper does not claim metric capture per se, but identifies a precondition for it—the relocation of validation authority to a legible, automatable component.
• seed-018: Authority reassignment is the mechanism; the paper reframes it as epistemic rather than purely operational.
• seed-069: GenAI outputs become trust proxies precisely because they are legible and consistent, not because they are correct.
• L-015: Educational institutions may retain formal audit trails (transcripts, evidence of learning) while losing the interpretive continuity that gave those records meaning.

Method note

This is primarily a normative/critical piece, not an empirical investigation or mechanistic account. It succeeds in naming a real phenomenon (authority displacement) but does not establish conditions under which it occurs, rate of displacement, reversibility, or downstream effects. The work indicates that law-building in protocol systems requires attending to who decides (governance topology), not only what is decided (output quality). It also suggests that meta-research should flag conceptual inversions early—when a framing shift (from "tool" to "authority locus") arrives first in critical discourse, it warrants empirical follow-up before becoming naturalized in practice.