A new academic paper argues that existing frameworks for understanding how students and educators interact with generative AI are fundamentally inadequate — and proposes a replacement theory that repositions systems like ChatGPT as active participants in knowledge construction rather than passive tools.

The paper, published on arXiv (arxiv.org/abs/2603.23863), introduces the Human–AI Epistemic Partnership Theory, or HAEPT, developed to explain user experiences that conventional adoption-focused models cannot account for. According to the researchers, standard constructs such as 'usefulness,' 'ease of use,' and 'engagement' were designed to evaluate tools that support tasks — not systems that actively shape what users come to believe.

'Systems such as ChatGPT do not merely support learning tasks but also participate in knowledge construction,' the authors write, arguing that this distinction demands an entirely new theoretical vocabulary.

At the core of HAEPT are three interlocking contracts the researchers say govern every human–AI educational interaction: an epistemic contract, governing who is treated as a credible source of knowledge; an agency contract, governing who drives the intellectual work; and an accountability contract, governing who bears responsibility for outcomes. The theory holds that users do not maintain fixed relationships with AI systems but instead move through 'calibration cycles' — repeated interactions that cause trust and skepticism to coexist and shift over time.

This framing has immediate implications for several debates already roiling education policy. Concerns about academic integrity, teacher reluctance to adopt AI, and student over-reliance on chatbot outputs are recast by HAEPT not as isolated problems but as symptoms of tension within these three contracts. A student who submits AI-generated work, for instance, is not simply cheating — under this model, they have collapsed the accountability contract entirely.

The researchers tested HAEPT's explanatory power against two documented use cases: collaborative learning with AI speakers and AI-facilitated scientific argumentation. Both scenarios, they argue, illustrate how the contract configurations shift depending on task type and context.

The paper arrives as institutions worldwide scramble to set coherent policies on generative AI in classrooms, often doing so without an agreed-upon theoretical basis for understanding what these tools actually do to the learning process. HAEPT represents one of the more ambitious attempts to supply that foundation — though, as a preprint, it has not yet undergone formal peer review.

Whether HAEPT gains traction in the research community will likely depend on whether its three-contract model proves useful to practitioners designing AI-integrated curricula, not just to theorists. The authors suggest it could reframe how educators interpret student behavior and how UX designers build AI systems intended for educational contexts.