Paper deep dive

Generative AI User Experience: Developing Human--AI Epistemic Partnership

Xiaoming Zhai

Year: 2026Venue: arXiv preprintArea: cs.CYType: PreprintEmbeddings: 89

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 3/26/2026, 2:12:48 AM

Summary

The paper introduces the Human–AI Epistemic Partnership Theory (HAEPT) to address the limitations of traditional technology adoption models (like TAM/UTAUT) in explaining the user experience of Generative AI in education. HAEPT posits that GenAI interaction is a dynamic negotiation of three interlocking contracts—epistemic, agency, and accountability—rather than simple tool usage, accounting for phenomena like redistributed cognition, negotiated authority, and trust calibration.

Entities (5)

ChatGPT · software · 100%Generative AI · technology · 100%Human–AI Epistemic Partnership Theory · theory · 100%Xiaoming Zhai · researcher · 100%Technology Acceptance Model · theory · 90%

Relation Signals (3)

Xiaoming Zhai → authored → Human–AI Epistemic Partnership Theory

confidence 100% · this paper develops the Human–AI Epistemic Partnership Theory (HAEPT)

Human–AI Epistemic Partnership Theory → explains → Generative AI

confidence 95% · HAEPT, explaining the GenAI user experience as a form of epistemic partnership

HAEPT → comprises → Epistemic Contract

confidence 90% · features a dynamic negotiation of three interlocking contracts: epistemic, agency, and accountability.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Generative AI (GenAI) has rapidly entered education, yet its user experience is often explained through adoption-oriented constructs such as usefulness, ease of use, and engagement. We argue that these constructs are no longer sufficient because systems such as ChatGPT do not merely support learning tasks but also participate in knowledge construction. Existing theories cannot explain why GenAI frequently produces experiences characterized by negotiated authority, redistributed cognition, and accountability tension. To address this gap, this paper develops the Human--AI Epistemic Partnership Theory (HAEPT), explaining the GenAI user experience as a form of epistemic partnership that features a dynamic negotiation of three interlocking contracts: epistemic, agency, and accountability. We argue that findings on trust, over-reliance, academic integrity, teacher caution, and relational interaction about GenAI can be reinterpreted as tensions within these contracts rather than as isolated issues. Instead of holding a single, stable view of GenAI, users adjust how they relate to it over time through calibration cycles. These repeated interactions account for why trust and skepticism often coexist and for how partnership modes describe recurrent configurations of human--AI collaboration across tasks. To demonstrate the usefulness of HAEPT, we applied it to analyze the UX of collaborative learning with AI speakers and AI-facilitated scientific argumentation, illustrating different contract configurations.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

89,108 characters extracted from source content.

Expand or collapse full text

Generative AI User Experience: Developing Human–AI Epistemic Partnership Xiaoming Zhai xiaoming.zhai@uga.edu AI4STEM Education Center, University of Georgia Athens, GA, USA Abstract Generative AI (GenAI) has rapidly entered education, yet its user experience is often explained through adoption-oriented constructs such as usefulness, ease of use, and engagement. We argue that these constructs are no longer sufficient because systems such as ChatGPT do not merely support learning tasks but also participate in knowledge construction. Existing theories cannot explain why GenAI frequently produces experiences characterized by negotiated authority, redistributed cognition, and accountability tension. To address this gap, this paper develops the Human–AI Epistemic Part- nership Theory (HAEPT), explaining the GenAI user experience as a form of epistemic partnership that features a dynamic negotiation of three interlocking contracts: epistemic, agency, and accountability. We argue that findings on trust, over-reliance, academic integrity, teacher caution, and relational interaction about GenAI can be reinterpreted as tensions within these contracts rather than as iso- lated issues. Instead of holding a single, stable view of GenAI, users adjust how they relate to it over time through calibration cycles. These repeated interactions account for why trust and skepticism often coexist and for how partnership modes describe recurrent configurations of human–AI collaboration across tasks. To demon- strate the usefulness of HAEPT, we applied it to analyze the UX of collaborative learning with AI speakers and AI-facilitated scientific argumentation, illustrating different contract configurations. Keywords generative AI (GenAI); ChatGPT; user experience; epistemic agency; epistemic trust; Human–AI Epistemic Partnership Theory (HAEPT); academic integrity ACM Reference Format: Xiaoming Zhai. 2026. Generative AI User Experience: Developing Human– AI Epistemic Partnership. In Proceedings of ACM. ACM, New York, NY, USA, 15 pages. https://doi.org/10.1145/n.n 1 Introduction Educational technologies have historically been theorized as in- struments that deliver content, structure practice, represent phe- nomena, or coordinate interaction (Januszewski & Molenda, 2013). Even when learners become emotionally engaged with games or socially connected through platforms, the technology has typically remained epistemically subordinate: it is not presumed to participate ACM, 2026. ACM ISBN 978-x-x-x-x/Y/M https://doi.org/10.1145/n.n in knowledge construction as a quasi-interlocutor. Generative arti- ficial intelligence (GenAI) like ChatGPT changes the phenomenol- ogy and the stakes of educational interaction by producing fluent language that resembles explanation, critique, and reasoning on demand (Samala et al., 2024). In classrooms, such outputs function as candidate knowledge claims and candidate artifacts, entering directly into the epistemic stream of learning activity—often prior to teacher mediation or peer review (Memarian & Doleck, 2023). Two structural conditions accelerate the educational significance of this shift. First, much GenAI is widely accessible and used beyond institutional control, which reduces the “gatekeeping” function that schools historically exerted over instructional technologies. The OECD’s Digital Education Outlook 2026 underscores that, unlike earlier waves of education technology, GenAI is often freely ac- cessible, intuitive, and adopted outside school oversight, and it warns that outsourcing tasks to GenAI without pedagogical guid- ance can yield performance gains without learning gains (OECD, 2026). Second, the policy and regulatory environment remains in flux; iterative releases of publicly available GenAI are outpacing na- tional regulatory adaptation, and the absence of regulations leaves data privacy unprotected and institutions underprepared to validate tools pedagogically and ethically (UNESCO, 2023). The emerging research base suggests that users are already con- structing distinct forms of experience around GenAI. Students pri- marily used ChatGPT for brainstorming, summarizing texts, and finding research articles; they viewed it as useful for simplifying complex information but less reliable for providing information and supporting classroom learning, and they expressed concerns about cheating and plagiarism alongside positive emotions such as curiosity and calmness (Ravšelj et al., 2025). This user experience is not reducible to convenience such as immediacy (reassurance and validation), equity (access and perceived degree value), and in- tegrity (policy ambiguity and group-work vulnerability), which are central themes of lived experience (Holland & Ciachir, 2025). More- over, students’ experience of these GenAI is significantly associated with their academic background and personality (Deng et al., 2025). Educators’ experience shows parallel complexity. For example, instructors at a U.S. university reported moderate-to-high famil- iarity with GenAI concepts but limited direct instructional use; importantly, trust and distrust were related yet distinct rather than opposite poles of a single continuum, pointing to the need for cali- bration rather than simple persuasion (Lyu et al., 2025). Meanwhile, K–12 teachers reported substantial uncertainty and skepticism, with many teachers unsure and a quarter reporting that AI tools do more harm than good (Lin & Pew Research Center, 2024). Despite the mixed perceptions, most teachers anticipate major shifts to teaching and assessment. Some believe GenAI will have a major or profound arXiv:2603.23863v1 [cs.CY] 25 Mar 2026 ACM, 2026, Zhai impact and identify anticipated changes that emphasize learning with AI, higher-order thinking, ethical values, process-oriented assessment, and the centrality of face-to-face relational learning (Bower et al., 2024). These findings and patterns point to a gap—how to theorize the epistemic role of GenAI in users’ experience of learning and teaching? Much of the educational technology user experience (UX) literature still privileges constructs that presuppose tool-like medi- ation: usability, engagement, perceived usefulness, and intention to use. Such constructs remain valuable, but they do not explain why GenAI frequently produces experiences characterized by ne- gotiated authority (“Should I believe this?”), redistributed cognition (“Who is doing the thinking?”), and accountability tension (“Whose work is this—and who is responsible?”). To address this gap, we propose the Human–AI epistemic part- nership theory (HAEPT), adapting the educational UX question to become epistemic and normative. HAEPT proposes that GenAI user experience in education is best explained as the dynamic ne- gotiation of an epistemic partnership rather than as the adoption of a tool. HAEPT intends to supply an explanatory core that can organize—and sometimes re-interpret—findings that otherwise ap- pear fragmented across motivation, trust, integrity, and classroom orchestration. 2 Review of generative AI user experience The empirical literature on GenAI, like ChatGPT, in education is ex- panding quickly. Systematic reviews in education synthesize bene- fits ranging from accessibility, engagement, and cognitive/emotional support to concerns about academic integrity, over-reliance, and technostress (Bahroun et al., 2023; Heung & Chiu, 2025). Moreover, research across educational sectors highlights opportunities (e.g., individualized support; teacher workload reduction through au- tomation) while repeatedly identifying integrity and ethical risks as salient obstacles to integration. Within this emerging literature, several UX regularities are already visible. 2.1 Use is widespread, but “use” does not equal “trust” Although research has suggested broad uptake and a stable set of common tasks—brainstorming, summarization, and information finding, a practicum in the classroom is paired with ambivalence about reliability and classroom learning value. These seemingly positive use patterns, which appear to indicate high acceptance, can often mask epistemic hesitation and strategic caution. Bouyzourn and Birch (2025) found that perceived expertise and ethical risk could mostly predict overall trust in GenAI, followed by concerns about ease of use and transparency. Their findings also suggest that trust is highly task-contingent, particularly for complex academic work (e.g., coding). Instructor evidence reinforces this distinction. The technical bar- riers to AI for teachers are much lower than those for conventional technologies, resulting in greater access to AI. However, teachers often show insufficient trust in AI due to various factors, such as a lack of human characteristics and transparency, which grows their anxieties when using AI in classroom settings (Nazaretsky et al., 2021). In addition, Guo et al. (2024b) found that users’ fa- miliarity and knowledge of AI may positively impact UX, thereby increasing their trust without inflating it. Instructors’ limited use for direct instructional tasks in the presence of conceptual familiar- ity suggests professional caution shaped by accountability and role responsibility, not merely by perceived usefulness. 2.2 Experience is relational and affective, not only instrumental A striking development in measurement research is the move from “interaction” to “relationship.” Traditional technologies have been highly instrumental, emphasizing the interactivity that transforms the learning experience. In contrast, GenAI is usually dialogic, with smart intelligence mimicking human behavior (e.g., grading stu- dents’ work). This feature grants AI the agencies that are usually taken by humans, thus learners experience AI interaction with some of the same social-cognitive textures historically associated with human interaction. That is, students experience more relational and affective than instrumental. Thus, a robust learner–GenAI relation- ship is essential and, sometimes, predictive of learning engagement, perceived cognitive, and motivational effects (Jin, 2025). These findings matter theoretically because they indicate that GenAI UX is partly constituted by relational cues—responsiveness, conversational coherence, and social presence-like dynamics rather than by interface usability alone. Users do not simply “operate” GenAI; they engage in a patterned dialogue with an entity that can feel like an interlocutor. 2.3 Integrity and policy ambiguity are not peripheral; they are constitutive of UX Academic integrity concerns are now a persistent theme in both em- pirical studies and reviews of research on GenAI use, revealing how GenAI influences student behavior and academic honesty. Findings have emphasized both benefits and risks; thus, the field is calling for research and policy development that are responsive to this emerging landscape. Exemplar studies yield substantial empirical evidence that implies both practical and theoretical developments. For example, by examining college students’ use of ChatGPT in assessment practices, Kofinas et al. (2025) suggested that academic integrity in the age of GenAI is not an external governance is- sue layered on top of educational systems; instead, it is embedded in the very experience of assessment. They found that markers were largely unable to distinguish AI-assisted submissions from non-assisted ones. Consequently, the presence of GenAI altered how markers approached the assessment process itself. In other words, integrity concerns reshaped the lived experience of marking, not merely its outcomes. This has fundamentally changed both students’ and teachers’ experiences regarding integrity. While integrity is experienced as problematic, it is not merely because cheating is possible; instead, institutional transparency and policy play significant roles (Bretag et al., 2019; Eaton, 2023). Research has found that these institutional infrastructures are often absent, leaving students uncertain about what is legitimate (Mc- Cabe et al., 2012). Moreover, this uncertainty leaves student groups vulnerable when one member’s “inappropriate” use risks collective misconduct. This ambiguity also causes confusion among teacher Generative AI User Experience: Developing Human–AI Epistemic Partnership ACM, 2026, users when evaluating student work produced with emerging tech- nologies (Kasneci et al., 2023). Thus, it is not surprising that many teachers encounter contradictory user experiences with generative AI: they acknowledge that GenAI might harm academic integrity by enabling students to present AI-generated work as their own, while simultaneously emphasizing that educational impact depends on pedagogical guidance rather than unstructured outsourcing (Cotton et al., 2024). 2.4 Users adopt “stances” toward integration that reflect moral and cultural positioning UX is also shaped by how communities narrate what AI should be in education. These collective narratives, embedded in policy discourse, professional practice, and public debate, construct shared imaginaries about AI’s appropriate roles, capabilities, and limita- tions. Such imaginaries orient users’ expectations, shape their trust and skepticism, and implicitly define the criteria by which AI sys- tems are judged as effective or ethical. Consequently, UX emerges not solely from interface features or technical performance, but from the sociocultural meanings that communities attach to AI and the normative visions they advance for its place in teaching and learning. Prior research on technology adoption in education similarly shows that perceptions of digital tools are mediated by in- stitutional norms, cultural values, and professional identities, which together influence whether technologies are framed as empowering resources or disruptive threats to established pedagogical practices (Dwivedi et al., 2023). Research has identified distinct viewpoints among students and professors, including profiles characterized by ethical guardian- ship, balanced integration, and convenience-oriented enthusiasm. These research findings highlighted that perceptions of GenAI (e.g., ChatGPT) are not merely individual preferences but culturally situ- ated stances toward legitimacy, fairness, and pedagogical coherence (Tsiani et al., 2025). Higher education communities show substantial variation in how students and faculty interpret the benefits, risks, and appropriate uses of generative AI, particularly around issues such as academic integrity, intellectual dependency, and the preser- vation of critical thinking (Johri et al., 2024). While many students perceive generative AI as a practical learning aid that can enhance productivity and provide personalized support, educators often adopt a more cautious stance due to ethical responsibilities and concerns about assessment validity and the erosion of disciplinary learning processes (Sah et al., 2025). Such findings align with the broader observation that debates about AI integration in education are frequently debates about what education is for in an AI-rich environment. Discussions surrounding AI often invoke competing visions of educational purpose—whether emphasizing efficiency and skill augmentation, the cultivation of independent reasoning, or the preservation of humanistic forms of inquiry (Selwyn, 2021; Zhai, 2022). In this sense, user experience with AI systems is inseparable from broader normative debates about the goals of schooling, the meaning of academic work, and the kinds of intellectual capacities that education should foster in an increasingly automated society. Taken together, current evidence suggests that GenAI UX is best described as a composite experience structured by (a) relational interaction patterns, (b) epistemic uncertainty and evaluation de- mands, and (c) institutional and moral accountability pressures. This composite is not well captured by one-dimensional models of acceptance. 3 Existing EdTech user experience theories While existing UX theories have largely accounted for conventional educational technologies, they remain limited in explicating the distinctive user experiences emerging from the use of GenAI. First, technology acceptance models in the TAM and UTAUT traditions have been highly effective in explaining adoption behavior, partic- ularly through constructs such as perceived usefulness, perceived ease of use, performance expectancy, and effort expectancy (Davis et al., 1989; Venkatesh et al., 2003). Yet their explanatory center typically remains the user’s perception of whether a system helps achieve goals efficiently. With GenAI, however, “usefulness” is en- tangled with epistemic validity: outputs can feel useful, fluent, and responsive while still being inaccurate or misleading (Guo et al., 2025; Wang et al., 2026). Recent research in higher education shows this tension clearly. For example, students often report positive attitudes toward GenAI because of its personalized, immediate sup- port and perceived usefulness, while simultaneously expressing substantial concern about factual inaccuracy and the inability of these systems to handle complex tasks reliably (Chan & Hu, 2023). Research in computer science and engineering education found similar outcomes. That is, GenAI (e.g., ChatGPT) seems effective in programming tasks but performed less well in more complex do- mains that require deeper analytical reasoning and domain-specific expertise (Waqas et al., 2025). All that being said, perceived output quality is a significant driver of students’ learning motivation and outcomes, underscoring how UX with GenAI is shaped by the ten- sion between fluency, utility, and uncertainty rather than by stable instrumental utility alone (Bai & Wang, 2025). Second, cognitive and motivational theories explain learning interaction with the technology, treating the latter as an instruc- tional medium, which fails to cover GenAI’s increasing role an epistemic actor. For example, cognitive load theory helps explain why some interfaces reduce search costs and extraneous processing demands during learning (Sweller, 1988). Motivation theories such as self-determination theory likewise clarify why learners may experience greater autonomy or competence when they receive immediate, personalized assistance (Ryan & Deci, 2000). However, these frameworks were not designed to explain the social-epistemic negotiation involved when learners interact with a system like GenAI that produces authoritative-seeming answers, shapes be- lief revision, and may encourage reliance even when its outputs warrant scrutiny. Recent work on epistemic agency argues that AI affects not only what people know, but also how they form and re- vise beliefs, thereby raising questions that extend beyond cognitive efficiency or motivational support (Coeckelbergh, 2026). Third, even sociocultural and activity-theoretic approaches, which are better suited to contextualizing tools in practices, were largely developed for artifacts that mediate action rather than generate domain-relevant explanations and arguments at scale that GenAI attends to. Wertsch (1998) suggest articulates how sociocultural theories emphasize the ways that tools shape human action within ACM, 2026, Zhai institutional and historical contexts, while Engeström (2001) em- ploys activity theory to similarly examine how tools mediate col- lective activity systems and their transformations. With GenAI, however, mediation increasingly becomes co-production. This shift is now being recognized in educational scholarship that frames GenAI as reconfiguring epistemic authority and even functioning as a “surrogate knower” or “innovator,” thereby challenging epis- temic agency and disrupting the justificatory practices through which knowledge is traditionally built in classrooms (Jose et al., 2025; Zhai, 2024). Chen (2025) further argues that GenAI should be understood as epistemic infrastructure rather than as a neutral tool, because it reshapes the conditions under which epistemic agency can be exercised and may influence long-term habit formation in knowledge work. Fourth, the human–AI teaming literature offers important in- sights into trust, coordination, transparency, and shared cognition, but education introduces additional stakes, including developmen- tal aims such as building learners’ agency, institutional norms such as assessment and credentialing, and asymmetrical responsibility, insofar as teachers remain accountable for students and the cur- riculum. Practices in the science domain have shown that scientists treat AI as a collaborative agent rather than a tool, highlighting the role of GenAI in the epistemic process (Herdiska & Zhai, 2024; Zhai, 2025). That is, human–AI teaming through adding an AI team- mate can reduce coordination, communication, and trust, and poor mutual understanding frequently undermines team performance (Schmutz et al., 2024). To keep the effectiveness, Endsley (2023) suggests that shared situation awareness, transparent information displays, explainability, and trust calibration are essential. More re- cent frameworks also argue that collaborative human–AI contexts require new approaches to managing trust over time and to un- derstanding how responsibilities are distributed across interaction processes and task phases (McGrath et al., 2025). These insights are highly relevant to GenAI UX but require a theory that explicitly integrates epistemic authority and institutional accountability with learning aims. Finally, measurement innovation itself signals theoretical insuf- ficiency. The emergence of epistemic trust measures suggests that researchers are moving beyond traditional adoption constructs to capture the distinctive experience of learning with GenAI. For ex- ample, the Epistemic Trust in GenAI for Higher Education Scale (ETGAI-HE) identifies six dimensions of epistemic trust, including cognitive evaluation of trustworthiness, interpersonal and contex- tual influences, dependability and safety, system predictability and transparency, performance expectation, and user control and au- tonomy, but the validated six-factor model explains 70.8% of the total variance (Pandey et al., 2025). This progress in measurement is significant because it indicates that trust in GenAI is multidimen- sional and structurally central to the educational experience. At the same time, the need for such new measures suggests that existing UX theories do not yet provide an integrating explanatory account of why these dimensions matter, how they interact, and how they are enacted in authentic educational activity. 4 Human–AI Epistemic Partnership Theory To fill the gap, the Human-AI epistemic partnership theory (HAEPT) was proposed, claiming that user experience with GenAI in educa- tion is fundamentally an epistemic partnership. This partnership is a dynamic, context-sensitive negotiation in which a human and a GenAI system jointly shape knowledge work (e.g., explanation, ideation, argumentation, modeling, evaluation) under institutional norms that allocate authority and responsibility. Rather than con- ceptualizing GenAI use simply as tool interaction, HAEPT positions human–AI engagement as a form of distributed cognition in which knowledge production is mediated through interactions among human reasoning, technological affordances, and sociocultural con- text (Hutchins, 1995; Salomon, 1997). In this sense, the experiential dimension of GenAI use arises not merely from interface design but from how users perceive the system’s epistemic role within knowledge-making processes. HAEPT begins from an empirical observation in addition to a theoretical stance. Empirically, learners and teachers frequently behave as though GenAI can make epistemic contributions (Zhai & Nehm, 2023). That is, they consult it, ask for a rationale, request alternatives, and integrate its proposals into knowledge artifacts. Observational and survey studies across higher education contexts show that students commonly use generative AI systems such as ChatGPT for brainstorming, explanation, and drafting tasks, treating the system as a conversational partner that can provide feedback or generate alternative perspectives (Dwivedi et al., 2023; Nyaaba et al., 2024). Systematically, this interaction resembles a partnership arrangement in which roles and responsibilities must be negotiated. Theoretically, HAEPT aligns with scholarship arguing that con- temporary AI systems occupy a novel position within epistemic ecosystems: they are neither traditional information tools nor au- tonomous experts, but generative systems capable of producing plausible knowledge representations at scale (Bender et al., 2021; Floridi et al., 2018). As a result, educational use of GenAI requires pedagogical approaches that support critical discernment, epistemic vigilance, and reflective engagement rather than passive acceptance of machine-generated outputs (Zhai et al., 2026). From this perspec- tive, UX with GenAI is inseparable from broader questions about epistemic authority, responsibility, and the social organization of knowledge production in digital learning environments. 4.1 Epistemic partnership in education A human-AI epistemic partnership in education can be conceptu- alized as a socio-cognitive configuration in which human learn- ers, educators, and AI jointly construct knowledge under specific epistemic and institutional constraints. This configuration features three analytically distinct yet theoretically grounded conditions: First, the interaction must involve the production, transformation, or evaluation of knowledge claims or artifacts (e.g., explanations, ar- guments, solutions, instructional designs, or feedback), with GenAI functioning as an active contributor to these epistemic processes. In Human–GenAI contexts, this construction is increasingly mediated by AI-generated outputs that externalize and extend human think- ing. These outputs—such as generated explanations or drafts—serve as epistemic artifacts that can be inspected, critiqued, and iteratively Generative AI User Experience: Developing Human–AI Epistemic Partnership ACM, 2026, refined (Bereiter & Scardamalia, 1993). Importantly, GenAI does not merely transmit information but participates in shaping the form and direction of knowledge construction, thereby functioning as a cognitive agent that augments and reorganizes epistemic activity. Second, the interaction must entail the implicit or explicit dis- tribution of epistemic roles between humans and GenAI (e.g., gen- erator, critic, verifier, or authority), with these roles dynamically negotiated rather than fixed. Researchers have argued that cogni- tive processes in Human-AI partnership are not confined to the human learner but are distributed across human and artificial agents (Hutchins, 1995; Salomon, 1997). GenAI systems may assume roles such as content generator or feedback provider, while humans may act as evaluators, curators, or meta-level regulators of the interaction. However, unlike traditional tools, GenAI can simu- late roles associated with epistemic authority, raising questions about trust, over-reliance, and the calibration of human judgment. Thus, the allocation of epistemic roles is not only constitutive of how knowledge-building processes are organized (Chi, 2009; Dil- lenbourg, 1999), but also central to understanding how agency and control are negotiated in AI-mediated learning environments. Third, the interaction must entail normative consequences, par- ticularly regarding authorship, accountability, and the legitimacy of knowledge claims in AI-supported contexts. Educational systems continue to assign value to processes of knowledge production, jus- tification, and individual contribution, embedding learning within broader structures of assessment, credentialing, and professional re- sponsibility (Gee, 2000). In Human–GenAI partnerships, these nor- mative dimensions become more complex because AI's involvement challenges conventional assumptions about authorship and intellec- tual ownership. Epistemic actions—such as submitting AI-assisted work or relying on generated feedback—are evaluated against in- stitutional norms that define what counts as valid knowledge and legitimate participation. Consequently, Human–AI epistemic part- nerships are inherently normative, as they intersect with evolving expectations about transparency, responsibility, and the appropriate use of AI in educational practice (Crippen et al., 2026). Taken together, these three conditions—knowledge production, role distribution, and normative consequence—are not arbitrary but reflect foundational concerns in educational research regarding how cognition is mediated, distributed, and evaluated. In particu- lar, they align with longstanding inquiries into how technologies reconfigure the division of cognitive labor and reshape the pro- cesses through which learners construct and validate knowledge (Salomon, 1997). Here, we deliberately distinguish epistemic part- nership from mere interactivity. Many educational technologies are interactive, but they are not typically experienced as “contributors” that propose expansive content and reasoning on demand. GenAI differs in that it can produce extended explanations, arguments, and creative proposals that simulate expert discourse, thereby blurring the perceived boundary between tool and collaborator (Kasneci et al., 2023). 4.2 The three-contract architecture We conceptualize HAEPT as a negotiation among three interlocking “contracts.” We use the term “contracts” deliberately to emphasize that these relationships are not merely guided by abstract principles or unilateral rules, but by mutually constituted expectations that carry implications for rights, responsibilities, and accountability. The contract metaphor reflects sociotechnical research showing that human–technology relationships often rely on tacit expectations regarding trust, responsibility, and legitimacy (Floridi et al., 2018). It also highlights that these expectations are shaped through inter- action and can be revised, contested, or even broken, features that are not fully captured by terms like “rules” or “guidelines.” These are not formal documents; they are lived, often implicit, normative arrangements that become visible in moments of uncertainty, error, evaluation, or conflict. 4.2.1 The Epistemic Contract. The Epistemic Contract refers to how users decide what to trust and what counts as valid knowledge when interacting with AI. It addresses questions such as: What role is ChatGPT playing in this task—is it a source of information, a helper, a critic, or simply a text generator? What kinds of evidence or explanations are expected? How should uncertainty or possible errors be treated? This contract is particularly important in the context of GenAI because these systems produce fluent and confident responses re- gardless of their accuracy. This can lead users to overestimate the system’s authority or rely on it uncritically, increasing the risk of automation bias (Bender et al., 2021). Prior research has described this phenomenon as treating AI as a “surrogate knower,” in which learners may accept outputs without verifying the evidence or en- gaging in reflection (Kasneci et al., 2023). Empirical studies further highlight this tension: while students often find generative AI tools helpful and engaging, they also express concerns about their relia- bility and accuracy for academic purposes (Deng et al., 2025). In this sense, the Epistemic Contract is where issues of fluency, trust, and learner judgment come together. It shapes whether learners treat AI-generated responses as tentative inputs requiring evaluation or as authoritative answers that can be accepted without further scrutiny. 4.2.2 The Agency Contract. The Agency Contract indicates how thinking and decision-making responsibilities are shared between the user and the AI. It focuses on how cognitive and metacognitive tasks are distributed, including who generates ideas, who evaluates their quality, who monitors progress, and who determines when a task is complete. It raises key questions such as: Who is actu- ally doing the thinking in this interaction? Am I outsourcing core cognitive work or using AI to support and extend my reasoning? Who controls the direction of inquiry, and who is responsible for verifying the accuracy and quality of the output? In this sense, the Agency Contract makes explicit the often implicit division of intellectual labor between human and AI, highlighting how differ- ent configurations of this division can shape both the process and outcomes of learning. This contract helps explain why simply “using” AI does not automatically lead to meaningful learning. The OECD (2023) dis- tinguishes between improved performance (e.g., completing tasks more quickly) and actual learning gains, noting that heavy reliance on generative AI can increase productivity without improving un- derstanding if users disengage cognitively. The Agency Contract describes how this gap can emerge. When AI takes over both idea ACM, 2026, Zhai generation and evaluation, learners may become less actively in- volved in the thinking process. In contrast, when AI is used as a partner that encourages questioning, critique, and reflection, learn- ers are more likely to stay cognitively engaged. Research in science education supports this distinction. Studies of AI-supported dialogic learning environments show that well- designed conversational agents can promote perspective-taking, argumentation, and reflective reasoning, particularly in discussions of complex issues. These findings suggest that the educational value of AI depends not simply on whether it is used, but on how responsibility for thinking is distributed within the interaction (Guo et al., 2024a; Watts et al., 2025). 4.2.3 The Accountability Contract. The Accountability Contract refers to how authorship, responsibility, and ethical expectations are defined when using GenAI, particularly in contexts where the consequences of use matter. It focuses on how credit is assigned, what kinds of disclosures are required, and who is held responsible for the outcomes produced with AI support. Key questions include: Who “owns” the work when AI contributes to its creation? To what extent should AI use be made visible to others, such as instructors or collaborators? Who is accountable if the output is inaccurate, bi- ased, or violates academic or professional standards? More broadly, how do institutional policies, assessment practices, and disciplinary norms determine what counts as legitimate use? By making these issues explicit, the Accountability Contract highlights that AI use is not only a technical or cognitive matter, but also a social and ethical one, shaped by expectations of integrity, transparency, and responsibility. This contract is central to the UX in educational settings be- cause these environments attach high stakes to academic integrity, authorship, and the value of credentials. As a result, even small ambiguities about the use of GenAI can create significant uncer- tainty for both students and instructors. Recent analyses describe generative AI as a major disruption to established norms of orig- inality and authorship, requiring institutions to rethink policies, assessment design, and instructional practices (Gao et al., 2025). Student perspectives further illustrate how this disruption is ex- perienced in practice: learners often raise concerns about fairness (e.g., unequal access or inconsistent rules), transparency (e.g., un- clear expectations for disclosure), and vulnerability in collaborative work (e.g., uneven contributions when AI is used differently across group members). At the same time, instructors tend to approach GenAI more cautiously, not simply due to skepticism about the technology, but because accountability is concentrated in their role. They are responsible for upholding academic standards, ensuring fair assessment, and protecting the credibility of credentials, all of which heighten the perceived risks of AI use. This asymmetry helps explain why educators may impose stricter boundaries or require clearer justification for AI-assisted work, and more broadly, why judgments about the legitimacy of AI use are deeply shaped by accountability considerations. In all, the three contracts showed a significant difference be- tween traditional EdTech and GenAI (see Table 1). Compared with traditional EdTech, GenAI reconfigures all three contracts, mak- ing previously stable expectations more fluid and negotiable. In traditional EdTech, the Epistemic Contract is anchored in curated, institutionally validated knowledge, allowing learners to rely on relatively stable sources of authority with a limited need for ongo- ing verification. The Agency Contract is similarly structured, with technology primarily supporting rather than performing cognitive work, leaving learners responsible for core processes of thinking and evaluation. The Accountability Contract, in turn, remains largely human-centered, with clear norms around authorship, responsibil- ity, and academic integrity. By contrast, GenAI destabilizes these arrangements. Its ability to generate fluent yet uncertain outputs shifts the Epistemic Contract toward continuous calibration of trust and evidence. At the same time, the Agency Contract becomes more fluid, as cognitive responsibilities can be redistributed between human and AI in ways that may either support or undermine mean- ingful learning. Finally, the Accountability Contract becomes more complex and contested, as questions of authorship, disclosure, and responsibility are no longer self-evident but must be actively nego- tiated within evolving institutional and social norms. 4.3 Calibration cycles: how partnership UX evolves over time HAEPT emphasizes that epistemic partnership with GenAI is not fixed but continually evolving. Rather than holding a single, stable view of GenAI, users adjust their relationship to it over time through repeated interactions. This process can be understood as a series of “calibration cycles,” in which users refine their expectations and behaviors based on experience. A typical cycle includes: (a) encountering an AI-generated response, (b) evaluating it—either implicitly or explicitly—by checking, cross-referencing, or accepting it, (c) forming an affective reaction such as confidence, relief, or doubt, and (d) adjusting their approach moving forward, such as trusting the AI more, relying on it less, or changing how they disclose its use. Over time, these cycles shape how users position AI in terms of trust, responsibility, and use. This process is similar to the concept of calibrated trust in human–automation interaction, in which users gradually adjust their reliance on automated systems based on their perceived reliability and task demands (Lee & See, 2004). One example of the evolving process is reflected in users’ seem- ingly contradictory views of GenAI, which are better understood as part of an ongoing calibration process rather than as fixed at- titudes. Empirical studies consistently show that students report high satisfaction with generative AI tools due to their usability and perceived quality of responses, while simultaneously expressing concerns about the accuracy, reliability, and trustworthiness of AI-generated outputs (Lund et al., 2026; Ng et al., 2025). Similarly, instructors’ trust in AI varies by context, task, and perceived risk, underscoring the importance of teaching strategies that encourage the critical evaluation of AI outputs. This variability and evolution are accounted for by users’ epistemic trust, which is influenced by factors such as transparency, predictability, and user control. Within HAEPT, these factors serve as mechanisms shaping how users continuously adjust and renegotiate epistemic authority, the distribution of cognitive work, and accountability in human–AI interactions. Generative AI User Experience: Developing Human–AI Epistemic Partnership ACM, 2026, Table 1: Comparison between traditional Educational Technology and GenAI across three contracts ContractTraditional Educational TechnologyGenAI Epistemic Contract (What counts as knowledge? What to trust?) Knowledge is pre-curated, stable, and institutionally vali- dated (e.g., textbooks, LMS content, simulations). Trust is largely delegated to external authorities such as teachers, publishers, and curriculum designers. Uncertainty is mini- mized and bounded. Knowledge is dynamically generated, probabilistic, and potentially fallible. Trust must be actively evaluated by the user. Outputs are fluent but may be inaccurate, requiring verification and critical evaluation. Risks include automa- tion bias and treating AI as a surrogate knower. Agency Contract (Who does the thinking?) Cognitive roles are clearly structured: learners engage in tasks designed by instructors, while technology supports delivery, practice, or visualization. Thinking remains pri- marily human-driven, with tools scaffolding specific pro- cesses. Cognitive labor is fluid and negotiable. AI can generate ideas, evaluate responses, and guide inquiry, potentially taking over core thinking processes. Learning depends on whether users outsource cognition or use AI as a partner for reflection and critique. Risks arise when agency shifts too heavily to AI. Accountability Contract (Who is responsible?) Authorship and responsibility are relatively clear and human-centered. Outputs are typically produced by stu- dents, with tools playing a transparent, supportive role. Institutional norms for assessment and integrity are well established. Authorship becomes ambiguous and distributed between human and AI. New norms are required for disclosure, credit, and ethical use. Responsibility for errors, bias, or misconduct is contested and context-dependent, raising concerns about fairness, transparency, and academic in- tegrity. 4.4 Partnership modes: recurrent configurations of the three contracts To make the idea of contract negotiation more concrete and ob- servable, we introduce the concept of partnership modes in HAEPT. Partnership modes refer to relatively stable patterns in how the three contracts—epistemic, agency, and accountability—are config- ured during a particular task. This approach is grounded in prior human–AI interaction research, which shows that people adopt different “epistemic relationships” with AI systems depending on context, perceived reliability, and purpose of use. In practice, users do not renegotiate each contract from scratch in every interaction. Instead, they tend to settle into recognizable ways of working with GenAI (e.g., using it as a tool, a collaborator, or an authority), at least within a given context or activity. These recurring patterns emerge because users rely on prior experience, task demands, time con- straints, and institutional expectations to guide their interactions. As a result, partnership modes function as “default configurations” that simplify decision-making during use, even though they remain adjustable over time. This helps explain why interactions with AI can feel both stable (within a task) and variable (across contexts or users). What makes these modes educationally important is that they render differences in user experience more systematic, interpretable, and actionable. Rather than attributing variation to vague differ- ences in attitudes or preferences (e.g., “some students like AI more than others”), the framework shows that these differences reflect underlying configurations of trust, responsibility, and cognitive engagement. For example, two students may both report frequent AI use, but one may be operating in a mode of instrumental reliance while the other engages in co-agency collaboration—leading to very different learning outcomes. By making these distinctions visible, partnership modes allow researchers to analyze patterns of use more precisely, link them to learning processes and outcomes, and develop more nuanced measures of AI engagement. Moreover, this framework enables these patterns to be intention- ally shaped through design and pedagogy. Educators can identify which partnership modes are emerging in their classrooms, evaluate their alignment with instructional goals, and design interventions to encourage more productive configurations. For example, prompts, assignments, and assessment criteria can be structured to promote verification, reflection, and dialogue, thereby shifting students away from authority displacement and toward co-agency collaboration. Similarly, clear policies and disclosure norms can stabilize the Ac- countability Contract, reduce ambiguity, and support responsible use. In this way, partnership modes serve not only as an analytic lens but also as a practical design tool for guiding more effective and equitable human–AI interactions in education. 5 Analytic vignettes: how contracts surface in everyday educational moments To illustrate HAEPT’s explanatory power, consider three analytic vignettes that are not presented as new empirical data but as theo- retically informed composites consistent with patterns documented in the literature. 5.1 Vignette 1. Collaborative learning with GenAI speakers Lee et al. (2023) introduced the CLAIS (Collaborative Learning with GenAI speakers) system represents an instructional use case in which a GenAI speaker is embedded as a peer within small-group collaborative learning, specifically structured through the Jigsaw model. In this setting, 3–4 pre-service teachers work together with a GenAI speaker that can respond to spoken prompts, explain content (e.g., learning theories), and pose questions. The activity proceeds through expert-group and home-group phases, in which both hu- man students and the GenAI take turns explaining assigned content and contributing to the group's understanding. The GenAI speaker ACM, 2026, Zhai participates through natural language interaction, offering expla- nations and answers based on pre-programmed knowledge, while the instructor facilitates the process and intervenes when technical issues arise. In this sense, the use case operationalizes a classroom where GenAI is not merely a support tool but a conversational participant in collaborative knowledge construction. From the perspective of HAEPT, we interpret this use case as a concrete instantiation of a human–AI epistemic partnership in which knowledge work is jointly produced through distributed interaction among learners, AI, and instructional structures. The CLAIS system clearly satisfies the three defining conditions of epistemic partnership: it involves active knowledge production (through explanation and problem-solving), dynamic role distribu- tion (AI as explainer and responder; humans as interpreters and collaborators), and implicit normative consequences (e.g., peer eval- uation of AI and human contributions). What becomes analyti- cally significant is how the three contracts—epistemic, agency, and accountability—are configured in this specific design. The epistemic contract in CLAIS is relatively stable, characterized by high perceived reliability and limited uncertainty. Because the GenAI speaker is trained on curated textbook knowledge, learners tend to treat its outputs as correct, consistent, and trustworthy, as reflected in their high ratings of accuracy and reliability. We see this as a partnership mode where GenAI occupies a quasi-authoritative peer role: it is not formally a teacher, yet it functions as a depend- able source of explanation within the group. The user experience signature here is one of low epistemic friction: learners can quickly access and accept explanations without extensive verification. How- ever, if this epistemic configuration becomes dominant, it might risk the over-trusting fluency. Learners may not engage in critical evaluation or epistemic questioning, especially because the system does not expose uncertainty or alternative interpretations. The agency contract reflects a structured but shifting distribution of cognitive labor. The Jigsaw design ensures that students remain responsible for explaining and integrating knowledge, while the GenAI contributes explanations and answers on demand. In our view, this produces a partnership mode of guided co-agency, where GenAI supports and augments human thinking without fully re- placing it. The user experience signature is enhanced efficiency and interactional flow: students report smoother collaboration and reduced effort in accessing information. At the same time, we notice that GenAI explanations can become the model for group discourse, subtly repositioning students as recipients rather than generators of ideas. If this configuration dominates, the key educational risk is cognitive offloading, in which learners increasingly rely on GenAI for core reasoning processes, resulting in performance gains with- out corresponding conceptual understanding. The accountability contract in this use case remains implicit and under-articulated. Although students evaluate GenAI as a peer, the system does not explicitly address authorship, responsibility, or disclosure of GenAI contributions. Interestingly, students apply different evaluative standards to human and AI peers, being more critical of GenAI while uniformly positive toward human peers, suggesting that GenAI is perceived as accountable in performance but not embedded in social or ethical norms. The user experience signature is, therefore, a form of normative ambiguity: learners interact with GenAI as a contributor without clear expectations regarding responsibility for its outputs. If this contract becomes dominant, it risks diffused accountability, in which students may incorporate AI-generated knowledge without fully owning or justi- fying it, potentially undermining norms of authorship and academic integrity. Taken together, we characterize the dominant partnership mode in CLAIS as structured co-agency with epistemic stabilization and weak accountability. This configuration produces an engaging and efficient collaborative experience, demonstrating the feasibility of human–AI co-participation in the construction of classroom knowledge. However, from a HAEPT perspective, its limitations are equally instructive: the design stabilizes trust and participation but does not sufficiently challenge learners to negotiate epistemic au- thority, retain cognitive ownership, or clarify responsibility. These tensions suggest that future designs should intentionally rebalance the three contracts—introducing epistemic uncertainty, reinforcing human agency in evaluation, and making accountability explicit— to support more robust and educationally productive human–AI epistemic partnerships. Figure 1. Flow chart of the CLAIS-Jigsaw system (adopted from Lee and Zhai, 2025) 5.2 Vignette 2. GenAI facilitates live scientific argumentation Kleiman et al. (2025, November) developed a GenAI multi-Agent system, named ArgueAgent, which orchestrates and actively par- ticipates in live scientific argumentation practices. In this system, students first produce individual explanations or visual models of scientific phenomena. The GenAI multi-agent system evaluates these responses and algorithmically pairs students with differing ideas to stimulate productive disagreement. During subsequent discussion, students are expected to articulate claims, justify rea- soning, critique peers, and respond to counterarguments. When the argumentation process stalls—due to lack of skill, premature consensus, or off-track discussion—ArgueAgent intervenes by tak- ing on roles such as facilitator (prompting participation), mediator (clarifying positions), and challenger (posing counterarguments or probing questions). In this sense, the system does not merely provide content but dynamically shapes the structure and quality of epistemic interaction in real time. From the perspective of HAEPT, we interpret ArgueAgent as a more interventionist and process-oriented epistemic partnership compared to CLAIS (Lee & Zhai, 2025). Here, the AI is not pri- marily a source of knowledge but an active regulator of epistemic processes—guiding the generation, evaluation, and contestation of knowledge claims. The system clearly satisfies the three defining conditions of epistemic partnership: it engages directly in knowl- edge production (through prompting and critique), dynamically re- distributes epistemic roles (e.g., AI as facilitator/mediator/challenger; students as arguers), and introduces normative consequences (e.g., expectations for participation, justification, and critique). What becomes central, then, is how the three contracts are configured to shape users’ experience within this more dialogically intensive environment. The epistemic contract in ArgueAgent is characterized by pro- ductive destabilization of knowledge claims. Unlike CLAIS, where Generative AI User Experience: Developing Human–AI Epistemic Partnership ACM, 2026, Figure 1: Flow chart of the CLAIS-Jigsaw system (adapted from Lee and Zhai, 2025). GenAI outputs are perceived as reliable answers, here GenAI delib- erately introduces tension—by pairing conflicting ideas and chal- lenging students’ reasoning. We see this as a partnership mode in which GenAI functions as an epistemic provocateur rather than ACM, 2026, Zhai an authority. The user experience signatures epistemic friction: students are pushed to justify, defend, and revise their ideas in re- sponse to both peers and GenAI interventions. The activities were configured to highly align with disciplinary practices in science, where argumentation and critique are central. However, if this con- tract becomes overly dominant, there is a risk of over-reliance on AI-generated interventions. Students may begin to treat GenAI challenges as the primary standard for evaluation, rather than de- veloping their own criteria for assessing evidence and arguments. In this sense, epistemic vigilance may shift from internally regulated to externally triggered. The agency contract reflects a distributed but AI-orchestrated division of cognitive labor. ArgueAgent takes responsibility for structuring interaction (e.g., pairing students, initiating prompts, sustaining dialogue), while students are responsible for generating and defending knowledge claims. We interpret this as a partnership mode of orchestrated co-agency, where GenAI governs the process of thinking rather than the content itself. The user experience sig- natures guided engagement: students are kept actively involved through continuous prompts and interventions, reducing the likeli- hood of disengagement or superficial consensus. At the same time, this strong orchestration introduces a subtle shift in control, in which AI determines when and how argumentation proceeds. If dominant, GenAI may risk the process dependency, where students rely on AI to sustain productive discourse and may struggle to self-regulate argumentation in its absence. The development of metacognitive and dialogic skills may thus be constrained if agency is not gradually rebalanced. The accountability contract in this use case is emergent but still ambiguous. On one hand, the system implicitly enforces norms of participation, justification, and critique, thereby strengthening accountability for epistemic engagement. Students are expected to contribute and respond, and GenAI interventions make disen- gagement more visible. On the other hand, the role of GenAI in shaping argumentation raises questions about responsibility: who is accountable for the direction and quality of the discussion—the students or the GenAI that mediates it? The user experience fea- tures a distributed accountability with hidden governance: students experience themselves as responsible participants, yet the GenAI exerts significant influence over the interactional structure. If this contract becomes dominant without explicit clarification, it can risk blurred epistemic ownership, where students may attribute the evolution of ideas or the quality of discourse to the system rather than to their own collective reasoning. Taken together, we characterize the dominant partnership mode in ArgueAgent as AI-orchestrated epistemic engagement with pro- ductive tension but latent dependency, which contrasts meaning- fully with the structured co-agency with epistemic stabilization observed in CLAIS (see Table 2). Whereas CLAIS stabilizes the epis- temic contract by positioning AI as a reliable knowledge contribu- tor, ArgueAgent deliberately destabilizes it to provoke critique and argumentation. Similarly, while CLAIS maintains a more balanced— though still shifting—agency distribution through structured roles in Jigsaw learning, ArgueAgent concentrates greater control in the AI by orchestrating the flow and quality of discourse. In terms of accountability, both systems exhibit ambiguity, but in different forms: CLAIS leaves authorship and responsibility underspecified in knowledge production, whereas ArgueAgent introduces account- ability for participation and reasoning while obscuring who gov- erns the epistemic process itself. From a HAEPT perspective, these differences highlight two distinct trajectories of human–AI partner- ship: one that risks over-stabilizing trust and encouraging passive acceptance (i.e., CLAIS), and another that risks over-centralizing epistemic regulation and fostering process dependency (i.e., Ar- gueAgent). Therefore, we argue that future designs should seek a balance between these modes—combining ArgueAgent's epistemic challenge and engagement with CLAIS's structured participation and clearer role distribution—while intentionally redistributing authority, agency, and accountability back to learners over time. 6 Discussion Although GenAI has become part of educational life faster than most earlier technologies, its uptake has not produced a simple story of acceptance. Students use GenAI widely for brainstorming, summarizing, searching, and drafting, yet they do not uniformly trust what these systems produce (Chan & Hu, 2023; Ravšelj et al., 2025). Instructors show a similarly mixed stance: many are familiar with GenAI concepts, but direct instructional use remains limited, and trust often coexists with distrust rather than replacing it (Lyu et al., 2025). Research adds further depth by showing that students experience GenAI through themes such as immediacy, reassurance, equity, and integrity, while teachers experience it through concerns about assessment validity, fairness, and policy ambiguity(Cotton et al., 2024; Holland & Ciachir, 2025). In other words, users know that GenAI is not experienced as “just another tool,” but it has been less clear how to explain these tensions within one coherent framework. We argue that the central issue is not simply whether GenAI is useful, easy to use, or even engaging. The deeper issue is that GenAI now enters the space of knowledge work itself. It proposes expla- nations, drafts arguments, critiques ideas, and sometimes appears to reason. That changes the user experience in a fundamental way. Instead of only asking whether the system helps complete a task, learners and teachers are also asking: Should I believe this? Who is doing the thinking here? Whose work is this? Existing acceptance models, such as TAM and UTAUT, remain valuable for explaining adoption and perceived usefulness (Davis et al., 1989; Venkatesh et al., 2003), but they are not built to explain these more epistemic questions. By introducing epistemic partnership as the unit of anal- ysis, this paper adds a concept that better fits the realities of GenAI use in education, extending recent work that describes GenAI as a “surrogate knower” or as a new form of epistemic infrastructure (Kasneci et al., 2023; Chen, 2025). Our argument is that the user experience of GenAI is not only relational or conversational; it is epistemic as well. Users are not merely interacting with a respon- sive interface but also negotiating with a system that appears to participate in the making, evaluation, and circulation of knowledge. This study proposed the HAEPT, which offers an integrative structure for explaining how that negotiation unfolds through three linked contracts: epistemic, agency, and accountability. This is im- portant because many of the field’s current concerns map onto one of these contracts but have rarely been theorized together. The Epistemic Contract addresses trust, justification, verification, and the risk of accepting fluent output as authoritative knowledge. Generative AI User Experience: Developing Human–AI Epistemic Partnership ACM, 2026, Table 2: Comparison between CLAIS and ArgueAgent ContractCLAIS (AI Speaker in Collaborative Learning) ArgueAgent (GenAI for Scientific Ar- gumentation) Key Difference EpistemicStabilized, high-trust knowledge environ- ment. AI provides reliable, pre-curated explanations. User experience: low epis- temic friction, smooth and confident use. Risk: over-trust and reduced critical evaluation; AI may become a surrogate knower. Destabilized, contestable knowledge en- vironment. AI provokes critique and con- flicting ideas. User experience: epistemic friction through continuous justification and revision. Risk: overreliance on AI cri- tique; epistemic standards may shift ex- ternally. CLAIS stabilizes knowledge and trust, while ArgueAgent unsettles knowledge to promote critique. Agency Structured co-agency. Humans lead ex- planation and integration; AI supports participation. User experience: efficient, smoother interaction. Risk: cognitive of- floading and reduced generative thinking. AI-orchestrated co-agency. AI structures pairing and discourse; humans argue within it. User experience: guided engage- ment and sustained high interaction. Risk: process dependency and weakened self- regulation. CLAIS distributes cognition with hu- man leadership; ArgueAgent places process control partly in AI. AccountabilityImplicit and underspecified. AI is treated as a peer but not clearly accountable. User experience: normative ambiguity. Risk: diffuse authorship and weak ownership of knowledge claims. Emergent but hidden governance. AI en- forces participation and shapes discourse. User experience: distributed accountabil- ity with hidden control. Risk: blurred epis- temic ownership and responsibility for process. CLAIS lacks clear accountability structures; ArgueAgent embeds ac- countability but obscures who gov- erns it. Overall Partnership Mode Structured co-agency with epistemic sta- bilization and weak accountability. AI-orchestrated epistemic engagement with productive tension but latent depen- dency. CLAIS emphasizes stability and sup- port; ArgueAgent emphasizes ten- sion and co-construction. The Agency Contract addresses how cognitive and metacognitive work is distributed between humans and AI, which helps explain why performance gains do not necessarily translate into learning gains (OECD, 2023). The Accountability Contract brings into the same frame issues that are often treated separately—authorship, disclosure, academic integrity, fairness, and ethical responsibility. Together, these contracts extend distributed cognition and sociocul- tural perspectives by showing not only that cognition is mediated across people and tools, but that GenAI also reshapes the normative organization of knowledge work in educational settings (Hutchins, 1995; Salomon, 1997). This is where we see HAEPT adding some- thing distinct: it does not replace prior theories but makes visible a layer of educational user experience that those theories only partially capture. One reason GenAI research can appear contradictory is that users often report both satisfaction and skepticism, or both trust and unease, in the same study. HAEPT helps explain this not as an inconsistency, but as an expected feature of human–AI epistemic partnership. Users recalibrate over time. They test outputs, accept some, reject others, and gradually adjust the extent of authority, agency, and responsibility they assign to the system. This account builds on research on trust calibration and trust in automation (Lee & See, 2004) and aligns with emerging measurement work showing that epistemic trust in GenAI is multidimensional rather than a single attitude (Pandey et al., 2025). The idea of partnership modes adds a useful middle layer between isolated interactions and broad outcomes. It allows researchers to describe recurrent configurations in ways that are theoretically meaningful and empirically testable. In that sense, HAEPT not only explains why user experiences differ but also how those differences are patterned. This paper also adds to the broader human–AI interaction litera- ture by showing why education cannot simply borrow human–AI teaming models without adaptation. Human–AI teaming research has emphasized transparency, coordination, situation awareness, and the active management of trust over time (Endsley, 2023; Mc- Grath et al., 2025; Schmutz et al., 2024). These insights are highly relevant, but education adds distinct demands. In education practice, the goal is not only effective task completion but also the devel- opment of reasoning, judgment, disciplinary understanding, and learner agency. Teachers remain asymmetrically accountable for assessment, fairness, and the credibility of credentials, even when students use AI independently (Gao et al., 2025; UNESCO, 2023). HAEPT contributes here by showing that educational human–AI interaction is a developmental and institutional case, not merely a collaborative one. This matters because a system that improves immediate performance may still be educationally weak if it re- duces students’ epistemic ownership or encourages uncritical de- pendence. The two vignettes, CLAIS and ArgueAgent, demonstrate clear user experience pictures that are distinct from conventional edu- cational technologies. CLAIS shows a partnership mode in which GenAI is positioned as a reliable peer contributor, producing struc- tured co-agency, smooth interaction, and relatively low epistemic friction, but also the risk of over-trusting AI explanations and under- specifying accountability (Lee et al., 2025). ArgueAgent, by contrast, exhibits a mode in which GenAI destabilizes ideas to stimulate argu- mentation, producing stronger epistemic engagement but also the ACM, 2026, Zhai risk that students become dependent on AI to structure and sustain discourse (Kleiman et al., 2025). The value of this comparison is not only descriptive but also demonstrates that different GenAI designs do not simply create “more” or “less” engagement; they produce different configurations of authority, agency, and responsibility, and these configurations matter educationally. In that sense, the paper also contributes methodologically to how the field might study GenAI going forward. Research on ed- ucational GenAI often reports phenomena such as over-reliance, hallucination risk, cheating, bias, or loss of critical thinking. These are important, but when treated as isolated problems, they can lead to fragmented responses. HAEPT provides a more organized system: Over-reliance can be understood as a shift in the Agency Contract; Uncritical trust can be understood as instability in the Epistemic Contract; Ambiguity about authorship, disclosure, or fairness can be understood as a weak or contested Accountability Contract. This reframing is more than terminological. It suggests that interventions should target the specific contract that is under strain rather than offer generic advice to “use AI responsibly,” a meaningful step for research, design, and policy, because it moves the conversation from broad caution to more precise diagnosis. Current debates can sometimes swing between enthusiasm and alarm: GenAI is framed either as a powerful learning companion or as a threat to integrity and human thinking. HAEPT suggests a more nuanced view. GenAI can indeed support explanation, feedback, creativity, and scientific reasoning, as growing work in assessment and science education already indicates (Cooper & Tang, 2024; Zhai & Krajcik, 2024). But these benefits do not emerge automatically from access or adoption, but depend on how epistemic authority is calibrated, how cognitive work is distributed, and how accountabil- ity is made visible. Put differently, the educational question is not whether GenAI belongs in education, but what kinds of human–AI partnerships education should cultivate. 7 Conclusions and future directions This paper suggests that user experience with GenAI in education is best understood not as simple tool use, but as a form of human–AI epistemic partnership. What users experience when they work with systems such as ChatGPT is not only convenience, speed, or inter- activity. They also experience shifting boundaries of knowledge, thinking, and responsibility. HAEPT captures this by proposing that GenAI use is organized through the ongoing negotiation of three interrelated contracts: what counts as credible knowledge, who is doing the cognitive work, and who remains answerable for the result. This perspective helps explain why the same system can feel empowering in one context, risky in another, and deeply ambivalent in both. It also helps explain why adoption alone tells us very little about educational value. A partnership can be effi- cient but shallow, engaging but overly dependent, or productive but normatively unstable. The quality of the partnership matters as much as the presence of the technology. Future work should examine not only what users say about AI, but also what they actually do with it across time. Looking ahead, although the field has already begun developing instruments for learner–GenAI relationships and epistemic trust (Pandey et al., 2025; Shi, 2025), these efforts remain relatively separate. HAEPT suggests the need for a more integrated measurement architecture. That means combining self-report measures with discourse data, revision histories, prompting patterns, verification behavior, disclo- sure decisions, and classroom observation. Such an approach would make it possible to identify partnership modes empirically, trace how users move between them, and examine which configurations are associated with stronger learning, greater epistemic agency, or more ethical use. Future education systems could make uncertainty more visible to strengthen the Epistemic Contract, require learners to explain why they accepted or rejected AI output to preserve Agency, or embed disclosure prompts and collaborative logs to clarify Account- ability. We argue that if HAEPT is useful, it should support design choices that intentionally reshape the three contracts. Classroom interventions could also test when AI support should be intensified and when it should fade. This is especially important for dialogic and argumentation-based systems. Work on AI-supported scientific argumentation already suggests that design choices can either open space for deeper reasoning or unintentionally centralize control in the AI (Watts et al., 2025). The next step is to study these tradeoffs deliberately rather than treating them as side effects. Future research should also examine how different institutional arrangements—such as disclosure rules, assessment redesign, teacher professional learning, or discipline-specific AI policies—shape part- nership modes in practice. The Accountability Contract is unlikely to stabilize through classroom practice alone. Students and teachers need clearer norms around disclosure, authorship, acceptable as- sistance, and assessment design. Recent guidance has emphasized that institutions are still catching up to the pace of GenAI develop- ment, and this lag leaves users navigating uncertainty with uneven support (UNESCO, 2023). This line of work is especially impor- tant because accountability pressures are not experienced equally. Teachers often carry the burden of policy interpretation and assess- ment legitimacy, while students bear the risks of inconsistency and unclear expectations (Crippen et al., 2026). We believe that future studies should examine how calibration of partnership modes unfolds over weeks, semesters, and years, and how that process varies by discipline, age, prior AI knowledge, lan- guage background, and access to support. Partnership with GenAI is not fixed; it develops. Students and teachers learn from repeated interactions, from failures, from institutional signals, and from each other. This matters not only for learning outcomes, but also for fairness. A partnership mode that appears productive for one group may be inaccessible or risky for another. If GenAI is becoming part of the educational infrastructure, then questions of equity, bias, and differential opportunity must be built into the study of user experience rather than added later as a separate concern. In closing, we see HAEPT as an attempt to offer the field a more precise language for a changing educational reality. GenAI is already altering how learners draft ideas, how teachers plan work, how feedback is generated, and how knowledge moves through classrooms. The challenge is no longer to decide whether a human– AI partnership exists. It does. Instead, the more urgent task is to understand which forms of partnership are educationally worth building, which ones should be resisted, and how pedagogy, design, and policy can help move users toward more reflective, responsible, and intellectually generative forms of working with AI. Generative AI User Experience: Developing Human–AI Epistemic Partnership ACM, 2026, 8 Acknowledgement The research reported here was supported by the National Science Foundation under Grant No. 2101104 (PI Zhai) and the Institute of Education Sciences, U.S. Department of Education, through Grant R305C240010 (PI Zhai). The opinions expressed are those of the authors and do not represent the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Depart- ment of Education. References Bahroun, Z., Anane, C., Ahmed, V., & Zacca, A. (2023). Transforming education: A comprehensive review of generative artificial intel- ligence in educational settings through bibliometric and content analysis. Sustainability, 15(17), 12983. Bai, Y., & Wang, S. (2025). Impact of generative AI interaction and output quality on university students’ learning outcomes: a technology-mediated and motivation-driven approach. Scientific Reports, 15(1), 24054. https://w.nature.com/articles/s41598-025- 08697-6.pdf Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, p. 610–623. Bereiter, C., & Scardamalia, M. (1993). Surpassing ourselves. An inquiry into the nature and implications of expertise. Chicago: Open Court. Bouyzourn, K., & Birch, A. (2025). What Shapes User Trust in ChatGPT? A Mixed-Methods Study of User Attributes, Trust Di- mensions, Task Context, and Societal Perceptions among University Students. arXiv preprint arXiv:2507.05046. Bower, M., Torrington, J., Lai, J. W., Petocz, P., & Alfano, M. (2024). How should we change teaching and assessment in response to increasingly powerful generative Artificial Intelligence? Outcomes of the ChatGPT teacher survey. Education and Information Tech- nologies, 1–37. Bretag, T., Harper, R., Burton, M., Ellis, C., Newton, P., van Haerin- gen, K., Saddiqui, S., & Rozenberg, P. (2019). Contract cheating and assessment design: exploring the relationship. Assessment & Evalu- ation in Higher Education, 44(5), 676–691. Chan, C. K. Y., & Hu, W. (2023). Students’ voices on generative AI: Perceptions, benefits, and challenges in higher education. In- ternational Journal of Educational Technology in Higher Education, 20(1), 43. Chen, B. (2025). Beyond tools: Generative AI as epistemic infras- tructure in education. arXiv preprint arXiv:2504.06928. Chi, M. T. (2009). Active-constructive-interactive: A conceptual framework for differentiating learning activities. Topics in cognitive science, 1(1), 73–105. Coeckelbergh, M. (2026). AI and epistemic agency: How AI in- fluences belief revision and its normative implications. Social Epis- temology, 40(1), 59–71. Cooper, G., & Tang, K.-S. (2024). Pixels and Pedagogy: Examining Science Education Imagery by Generative Artificial Intelligence. Journal of Science Education and Technology, 1–13. Cotton, D. R., Cotton, P. A., & Shipway, J. R. (2024). Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations in Education and Teaching International, 61(2), 228–239. Crippen, K. J., Zhai, X., & Lee, O. (2026). Responsible and Ethical Principles for the Practice of AI-Supported Science Education. In X. Zhai & K. J. Crippen (Eds.), Advancing AI in Science Education: Envisioning Responsible and Ethical Practice. Springer. Davis, F. D., Bagozzi, R., & Warshaw, P. (1989). Technology ac- ceptance model. J Manag Sci, 35(8), 982–1003. Deng, N., Liu, E. J., & Zhai, X. (2025). Understanding university students’ use of generative AI: The roles of demographics and per- sonality traits. International Conference on Artificial Intelligence in Education, p. 281–293. Dillenbourg, P. (1999). Collaborative learning: Cognitive and com- putational approaches. advances in learning and instruction series. ERIC. Dwivedi, Y. K., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, A., Kar, A. K., Baabdullah, A. M., Koohang, A., Raghavan, V., & Ahuja, M. (2023). Opinion Paper:“So what if ChatGPT wrote it?” Multidisci- plinary perspectives on opportunities, challenges and implications of generative conversational AI for research, practice and policy. International journal of information management, 71, 102642. Eaton, S. E. (2023). Academic integrity in the age of artificial intelligence. Endsley, M. R. (2023). Supporting Human-AI Teams: Transparency, explainability, and situation awareness. Computers in Human Be- havior, 140, 107574. Engeström, Y. (2001). Expansive learning at work: Toward an activity theoretical reconceptualization. Journal of Education and Work, 14(1), 133–156. Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Luetge, C., Madelin, R., Pagallo, U., & Rossi, F. (2018). AI4People—An ethical framework for a good AI society: Oppor- tunities, risks, principles, and recommendations. Minds and ma- chines, 28(4), 689–707. https://iris.unito.it/bitstream/2318/1728327/ 1/Article_AI4PeopleAnEthicalFrameworkFor-2.pdf Gao, Y., Zhai, X., Li, M., Lee, G., & Liu, X. (2025). A Multimodal Interactive Framework for Science Assessment in the Era of Gener- ative Artificial Intelligence. Journal of Research in Science Teaching. https://doi.org/10.1002/tea.70009 Gee, J. P. (2000). Chapter 3: Identity as an analytic lens for re- search in education. Review of research in education, 25(1), 99–125. Guo, S., Latif, E., Zhou, Y., Huang, X., & Zhai, X. (2024a). Us- ing generative AI and multi-agents to provide automatic feedback. arXiv preprint arXiv:2411.07407. Guo, S., Shi, L., & Zhai, X. (2024b). Developing and validating an instrument for teachers' acceptance of artificial intelligence in education. Education and Information Technologies, 30, 13439–13461. Guo, S., Wang, Y., Yu, J., Wu, X., Ayik, B., Watts, F. M., Latif, E., Liu, N., Liu, L., & Zhai, X. (2025). Artificial intelligence bias on English language learners in automatic scoring. International Conference on Artificial Intelligence in Education, p. 268–275. Herdiska, A., & Zhai, X. (2024). Artificial Intelligence-Based Sci- entific Inquiry. In X. Zhai & J. Krajcik (Eds.), Uses of Artificial Intel- ligence in STEM Education (p. 179–197). Oxford University Press. Heung, Y. M. E., & Chiu, T. K. F. (2025, 2025/06/01/). How Chat- GPT impacts student engagement from a systematic review and ACM, 2026, Zhai meta-analysis study. Computers and Education: Artificial Intelli- gence, 8, 100361. https://w.sciencedirect.com/science/article/pii/ S2666920X25000013 https://w.sciencedirect.com/science/article/pii/S2666920X25000013? via%3Dihub Holland, A., & Ciachir, C. (2025). A qualitative study of students’ lived experience and perceptions of using ChatGPT: Immediacy, equity and integrity. Interactive Learning Environments, 33(1), 483– 494. Hutchins, E. (1995). Cognition in the Wild. MIT press. Januszewski, A., & Molenda, M. (2013). Educational technology: A definition with commentary. Routledge. Jin, S.-H. (2025). Measures of learner-generative ai relationships. Computers and Education Open, 8, 100258. Johri, A., Hingle, A., & Schleiss, J. (2024). Misconceptions, prag- matism, and value tensions: Evaluating students' understanding and perception of generative AI for education. 2024 IEEE Frontiers in Education Conference (FIE), p. 1–9. Jose, B., Cleetus, A., Joseph, B., Joseph, L., Jose, B., & John, A. K. (2025). Epistemic authority and generative AI in learning spaces: rethinking knowledge in the algorithmic age. Frontiers in Education, Vol. 10, p. 1647687. Kasneci, E., Seßler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Günnemann, S., & Hüllermeier, E. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual differences, 103, 102274. Kleiman, J., Gao, Y., Zhu, Z., Wang, Z., Zhou, Y., Hao, J., & Zhai, X. (2025, November, March 21, 2026). ArguAgent: A dialogue sys- tem for scaffolding argumentative discourse National Conference on AI in Education, Athens, GA, United States) https://ai4genius.org/ arguagent-poster/ Kofinas, A. K., Tsay, C. H. H., & Pike, D. (2025). The impact of generative AI on academic integrity of authentic assessments within a higher education context. British journal of educational technology, 56(6), 2522–2549. Lee, G.-G., Mun, S., Shin, M.-K., & Zhai, X. (2023). Collaborative Learning with Artificial Intelligence Speakers (CLAIS): Pre-Service Elementary Science Teachers' Responses to the Prototype. Science & Education, 1–29. Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human factors, 46(1), 50–80. Lin, L., & Pew Research Center. (2024). A quarter of U.S. teachers say AI tools do more harm than good in K-12 education. https://w. pewresearch.org/short-reads/2024/05/15/a-quarter-of-u-s-teachers- say-ai-tools-do-more-harm-than-good-in-k-12-education/ Lund, B. D., Warren, S. J., & Teel, Z. A. (2026). Measuring Univer- sity Students Satisfaction with Traditional Search Engines and Gen- erative AI Tools as Information Sources. arXiv preprint arXiv:2601.00493. Lyu, W., Zhang, S., Chung, T., Sun, Y., & Zhang, Y. (2025). Un- derstanding the practices, perceptions, and (dis) trust of generative AI among instructors: A mixed-methods study in the US higher education. Computers and Education: Artificial Intelligence, 8, 100383. McCabe, D. L., Butterfield, K. D., & Trevino, L. K. (2012). Cheating in college: Why students do it and what educators can do about it. JHU Press. McGrath, M. J., Duenser, A., Lacey, J., & Paris, C. (2025). Collabo- rative human-AI trust (CHAI-T): A process framework for active management of trust in human-AI collaboration. Computers in Hu- man Behavior: Artificial Humans, 100200. Memarian, B., & Doleck, T. (2023). ChatGPT in education: Meth- ods, potentials, and limitations. Computers in Human Behavior: Artificial Humans, 1(2), 100022. Nazaretsky, T., Cukurova, M., & Alexandron, G. (2021). An In- strument for Measuring Teachers’ Trust in AI-Based Educational Technology. Ng, J., Tong, M., Tsang, E. Y., Chu, K., & Tsang, W. (2025). Explor- ing students’ perceptions and satisfaction of using GenAI-ChatGPT tools for learning in higher education: A mixed methods study. SN Computer Science, 6(5), 476. Nyaaba, M., Shi, L., Nabang, M., Zhai, X., Kyeremeh, P., Ayoberd, S. A., & Akanzire, B. N. (2024). Generative AI as a Learning Buddy and Teaching Assistant: Pre-service Teachers' Uses and Attitudes. arXiv preprint arXiv:2407.11983. OECD. (2023). OECD Digital Education Outlook 2023: Towards an Effective Digital Education Ecosystem. O. Publishing. https://doi.org/ 10.1787/c74f03de-en OECD. (2026). OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education. O. Publishing. Pandey, C. S., Mishra, P., Pandey, S. R., & Pandey, S. (2025). Epis- temic trust in generative AI for higher education scale (ETGAI-HE scale). Ai & Society, 1–14. Ravšelj, D., Keržič, D., Tomaževič, N., Umek, L., Brezovar, N., Iahad, N. A., Abdulla, A. A., Akopyan, A., Segura, M. W. A., & AlHu- maid, J. (2025). Higher education students’ perceptions of ChatGPT: A global study of early reactions. PloS one, 20(2), e0315011. https:// journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0315011& type=printable Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist, 55(1), 68. Sah, R., Hagemaster, C., Adhikari, A., Lee, A., & Sun, N. (2025). Generative AI in higher education: student and faculty perspectives on use, ethics, and impact. Issues in Information Systems, 26(2). Salomon, G. (1997). Distributed cognitions: Psychological and edu- cational considerations. Cambridge University Press. Samala, A. D., Zhai, X., Aoki, K., Bojic, L., & Zikic, S. (2024, 01/25). An In-Depth Review of ChatGPT’s Pros and Cons for Learning and Teaching in Education. International Journal of Interactive Mobile Technologies (iJIM), 18(02), p. 96–117. https://online-journals.org/ index.php/i-jim/article/view/46509 https://online-journals.org/index.php/i-jim/article/download/46509/ 14523 Schmutz, J. B., Outland, N., Kerstan, S., Georganta, E., & Ulfert, A.-S. (2024). AI-teaming: Redefining collaboration in the digital era. Current Opinion in Psychology, 58, 101837. Selwyn, N. (2021). Education and technology: Key issues and de- bates. Bloomsbury Publishing. Shi, L. (2025). Assessing teachers’ generative artificial intelli- gence competencies: Instrument development and validation. Edu- cation and Information Technologies, 1–20. Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive science, 12(2), 257–285. Generative AI User Experience: Developing Human–AI Epistemic Partnership ACM, 2026, Tsiani, M., Lefkos, I., & Fachantidis, N. (2025). Perceptions of gen- erative AI in education: Insights from undergraduate and master’s- level future teachers. Journal of Pedagogical Research, 9(2), 89–108. UNESCO. (2023). Guidance for generative AI in education and research. S. a. C. O. United Nations Educational. Venkatesh, V., Morris, M. G., Davis, G. B., & Davis, F. D. (2003). User acceptance of information technology: Toward a unified view1. MIS Quarterly, 27 (3), 425–478. Wang, Y., Wu, X., Huang, J., Liu, L., Zhai, X., & Liu, N. (2026). BRIDGE the Gap: Mitigating Bias Amplification in Automated Scor- ing of English Language Learners via Inter-group Data Augmenta- tion. arXiv preprint arXiv:2602.23580. Waqas, M., Hasan, S., & Ali, A. (2025). Effectiveness of Gener- ative AI Tools in Computer Science and Engineering Education. International Journal of Artificial Intelligence in Education, 1–32. Watts, F. M., Liu, L., Ober, T. M., Song, Y., Jusino-Del Valle, E., Zhai, X., Wang, Y., & Liu, N. (2025). A Framework for Designing an AI Chatbot to Support Scientific Argumentation. Education Sciences, 15(11), 1507. Wertsch, J. V. (1998). Mind as action. Oxford university press. Zhai, X. (2022). ChatGPT user experience: Implications for edu- cation. SSRN Electronic Journal, 1–13. Zhai, X. (2024). Transforming Teachers’ Roles and Agencies in the Era of Generative AI: Perceptions, Acceptance, Knowledge, and Practices. Journal of Science Education and Technology, 1–11. Zhai, X. (2025). DAIL: Discipline-Based Artificial Intelligence Literacy. Available at SSRN 5745703. Zhai, X., & Krajcik, J. (2024). Artificial Intelligence-based STEM Education. In X. Zhai & J. Krajcik (Eds.), Uses of Artificial Intelligence in STEM Education (p. 3–14). Oxford University Press. Zhai, X., & Nehm, R. (2023). AI and formative assessment: The train has left the station. Journal of Research in Science Teaching, 60(6), 1390–1398. Zhai, X., Pellegrino, J. W., Rojas, M., Park, J., Nyaaba, M., Cohn, C., & Biswas, G. (2026). Science Literacy: Generative AI as Enabler of Coherence in the Teaching, Learning, and Assessment of Scientific Knowledge and Reasoning. arXiv preprint arXiv:2603.06659.

Instant research discovery