Learning with machines: Toward a theory of epistemic co-agency

Synthesis: Samuel (2026) introduces the Epistemic Entanglement Framework, a theory-informed model for understanding how learners engage with generative AI (GenAI) systems. Arguing that existing learning theories (constructivism, sociocultural theory, connectivism) presume human-centered epistemic agency and cannot account for the ways GenAI simulates reasoning, reframes arguments, and co-constructs meaning, the paper proposes epistemic co-agency — a reflexive stance in which learners engage AI outputs dialectically, challenging assumptions, surfacing contradictions, and asserting epistemic sovereignty. The central claim is that the real challenge of AI in education is not technological fluency but cultivating learners who can reason with, through, and against generative systems.

Key Findings

• The paper identifies three epistemic configurations of human-AI interaction: Epistemic Augmentation (EA) — learners treat AI as a content generator, adopting a low-interrogation posture that improves surface productivity but leaves knowledge construction passive; Epistemic Integration (EI) — learners strategically prompt, refine, and coordinate AI outputs into evolving reasoning, managing knowledge production without interrogating the epistemic status of claims; and Epistemic Co-Agency (EC) — learners treat AI outputs as epistemically contestable contributions, engaging them dialectically while maintaining epistemic sovereignty and accountability for warranting knowledge.
• These configurations are not stages of maturity but patterned epistemic relationships that vary by task, design, and learner posture; the task of learning design is to help learners recognize which configuration they are in, understand its limits, and develop agility to shift postures as needed.
• Epistemic entanglement refers to the condition where knowledge construction emerges through reciprocal interaction between human reasoning and AI-generated contributions, such that the boundaries of authorship, authority, and justification become hard to disentangle.
• Epistemic co-agency explicitly rejects "AI partner"/"co-author" metaphors (which imply mutuality and shared intention); it focuses instead on the learner's capacity to maintain epistemic sovereignty while navigating an entangled system of influence and response.
• The framework warns against current AI-in-education tools that prioritize efficiency and risk reinforcing passive epistemic postures and overtrust, calling for design that embeds "epistemic friction" (prompt templates requiring justification, multiple-output displays, confidence visualizations, critique scaffolds like "What's missing?" and "Whose perspective is absent?").

Study Design & Method

This is a conceptual/theoretical paper proposing a framework, not an empirical study. Drawing on distributed cognition, sociomaterialism, posthumanist theory, and the concept of relational agency, the author develops the Epistemic Entanglement Framework and articulates the three configurations (EA, EI, EC) in terms of distinct human postures, cognitive demands, and design implications. The configurations are also mapped onto analytic dimensions from Thürmel's (2015) graduated agency framework (activity level, adaptivity, interaction potential, personification, joint agency). The paper concludes with pedagogical and assessment implications for Scaffolding learners toward co-agency, and acknowledges the author used ChatGPT as a supplementary writing/ideation tool during manuscript development.

What this means for practice

• Instructors. Make the epistemic configuration visible: help learners recognize whether they are in augmentation, integration, or co-agency, and design transitions deliberately rather than assuming uniform critical engagement.
• Instructors. Embed epistemic friction — prompt templates that require learners to justify why an output is appropriate or flawed, multiple-output displays of divergent responses, confidence visualizations or traceable sources, and critique scaffolds such as "What's missing?" and "Whose perspective is absent?"
• Instructors. Assess the reasoning trajectory, not just the artifact: use reflective prompts, process explanations, and evidence of revision, and avoid "AI detection" measures that reduce engagement to surveillance.
• Faculty developers. Embed epistemic ethics, source reliability, and knowledge construction in disciplinary curricula rather than standalone AI literacy modules, and build faculty capacity to teach in AI-mediated environments where disciplinary validation norms differ.
• Designers. Treat GenAI as an object of inquiry rather than an assistant, so learners can see its limitations and the values encoded in its design.

Limitations

As a conceptual paper, the framework has not been empirically tested; the author explicitly calls for future research to test and refine the model. The framework's configurations are theoretical constructs without operationalized measures. The paper does not provide concrete curriculum designs, only illustrative examples. The context is higher education, and the author notes disciplinary norms vary in how knowledge is validated, which may affect how the framework applies across fields.

Citation

Samuel, A. (2026). Learning with machines: Toward a theory of epistemic co-agency. Computers and Education: Artificial Intelligence.