On a Tuesday morning in September, a composition instructor at a public university in California opened On a Tuesday morning in September, a composition instructor at a public university in California opened her laptop to find thirty-seven student essays on the prompt she'd assigned the previous week: "Discuss the role of symbolism in The Great Gatsby." They were all competent. Most were good. Three were excellent. And she was almost certain that none of them had been written by human hands alone.
This wasn't a crisis of academic integrity—at least, not in the way we've traditionally understood it. The students hadn't cheated, exactly. They had simply done what millions of people now do dozens of times a day: they had asked a large language model to help them think. The question that kept the instructor awake that night wasn't whether her students had used AI—it was whether her assignment had any remaining pedagogical purpose in a world where such tools exist. More fundamentally: what cognitive work remains distinctively human when machines can generate fluent text on demand?her laptop to find thirty-seven student essays on the prompt she'd assigned the previous week: "Discuss the role of symbolism in The Great Gatsby." They were all competent. Most were good. Three were excellent. And she was almost certain that none of them had been written by human hands alone.
This wasn't a crisis of academic integrity—at least, not in the way we've traditionally understood it. The students hadn't cheated, exactly. They had simply done what millions of people now do dozens of times a day: they had asked a large language model to help them think. The question that kept the instructor awake that night wasn't whether her students had used AI—it was whether her assignment had any remaining pedagogical purpose in a world where such tools exist. More fundamentally: what cognitive work remains distinctively human when machines can generate fluent text on demand?
The Medium and the Message
To understand what's at stake in this transformation, we need to think carefully about what N. Katherine Hayles calls "medium specificity"—the recognition that different cognitive systems (human brains, large language models, institutional structures) have distinct capabilities, constraints, and blind spots. An LLM like GPT-4 or Claude is extraordinarily good at certain things: pattern recognition at massive scale, statistical coherence across millions of text samples, rapid synthesis of information from disparate sources, and the production of grammatically fluent prose that adheres to genre conventions. It can write a competent literary analysis of The Great Gatsby because it has processed thousands of such analyses and learned the discursive patterns that make them recognizable as "competent."
But—and this is the crucial asymmetry—an LLM has no embodied experience. It has never stood in an archive at 2 AM, physically exhausted, suddenly understanding how three seemingly unrelated documents connect. It has never felt the visceral resistance of a problem that won't resolve, the cognitive dissonance that signals a conceptual breakthrough is near. It has never made an intellectual commitment that felt morally consequential, never experienced the vulnerability of defending an original argument to skeptical peers, never known the specific texture of intellectual courage. These embodied, affective dimensions of cognition aren't decorative additions to "real" thinking—they're constitutive of what Hayles calls "cognition in the wild," the way human intelligence actually works in complex, ambiguous, high-stakes environments.
In her work on human-AI collaboration, Hayles argues that we're entering an era of "technogenesis"—the co-evolution of humans and technologies in which each shapes the other's capacities. The relevant question isn't whether AI will replace human intelligence but how the cognitive assemblage of human-plus-AI differs from either component alone. What emerges from this assemblage isn't simply augmentation (humans doing the same things faster or better) but transformation (humans doing fundamentally different kinds of cognitive work).
This is where Ethan Mollick's concept of "co-intelligence" becomes useful. In his research on organizational adoption of AI tools, Mollick has documented how the most effective human-AI collaboration doesn't involve clearly delineated roles (human does X, AI does Y) but rather a kind of improvisational dance in which each party's contribution is shaped by the other's presence. A skilled prompt engineer doesn't simply give the AI instructions—she enters into a dialogue where each response refines her understanding of what she's actually trying to accomplish. The AI's output reveals assumptions she didn't know she was making, blind spots in her initial framing, alternative pathways she hadn't considered.
But here's the difficulty: this kind of sophisticated co-intelligence isn't evenly distributed. It requires what Lucy Suchman calls "artful integration"—the ability to recognize when a tool's affordances match a task's demands and when they don't. Suchman's work on human-machine configurations emphasizes that all action is "situated"—embedded in specific contexts, responsive to particular constraints, emerging from the interaction of multiple agents (human and nonhuman) rather than pre-existing in the mind of any individual actor. Plans, in Suchman's account, are resources for action rather than its determinants; effective practice involves constant improvisation and adjustment based on real-time feedback from the environment.
Applied to education, this means that the relevant question isn't "Should students use AI?" but rather "Under what conditions does human-AI collaboration produce genuine learning versus shallow performance?" The answer has everything to do with what Lev Vygotsky called the "zone of proximal development"—that sweet spot between what a learner can do independently (too easy, no growth) and what's beyond their reach even with support (too hard, produces frustration and shutdown). AI, when used well, can scaffold students into that zone. It can take a vague intuition and help articulate it, translate expert discourse into accessible language, generate examples that illuminate abstract concepts. But when used poorly—when students outsource the entire cognitive process to the machine—it eliminates the zone of proximal development entirely, leaving them neither struggling productively nor succeeding independently.
The Paradigm That Isn't There Yet
Thomas Kuhn's The Structure of Scientific Revolutions gives us language for understanding what's happening to education right now. Kuhn distinguished between "normal science"—the everyday puzzle-solving that occurs within an established paradigm—and "revolutionary science," the crisis-driven restructuring that occurs when anomalies accumulate to the point that the existing paradigm can no longer accommodate them. Before a revolution, there's a period of what Kuhn called "extraordinary science": competing frameworks proliferate, practitioners disagree about fundamental questions, and the field loses its consensus about what constitutes legitimate knowledge or valid practice.
Higher education in 2025 is in precisely this pre-paradigmatic state. There's no consensus about what AI-enabled pedagogy should look like, no shared framework for assessment, no agreed-upon standards for AI literacy. Instead, there's a chaotic proliferation of approaches: some instructors banning AI entirely, others embracing it uncritically, most somewhere in between, improvising as they go. The old paradigm—the one built around essays, exams, and credentials as proxies for learning—is clearly failing, but the new paradigm hasn't yet crystallized.
What makes this crisis particularly acute is what Kuhn called "incommensurability"—the difficulty of translation between paradigms. Faculty trained in the old model literally cannot see what the new model's practitioners are doing, because their conceptual frameworks don't include the relevant categories. When a teacher like Jason Gulya asks students to critique AI-generated essays rather than write original ones, colleagues steeped in the old paradigm experience this as a betrayal of academic standards. They're not wrong, exactly—Gulya is abandoning certain traditional standards. But he's replacing them with different standards, ones calibrated to measure capacities the old paradigm didn't recognize as relevant: the ability to identify algorithmic bias, to articulate tacit knowledge the AI lacks, to negotiate the boundary between human and machine cognition.
Here's where the wineskin metaphor becomes more than decorative. In Matthew's Gospel, Jesus uses the image to explain why his teaching can't be contained within existing religious structures. "No one puts new wine into old wineskins," he says, "for the new wine would burst the wineskins and the wine would be spilled." The parable isn't about preserving tradition or embracing novelty—it's about recognizing when a transformation is so fundamental that incremental adaptation won't work. The fermentation can't be stopped or slowed; it can only be accommodated or spilled.
But unlike Kuhn's paradigm shifts—which resolve when a new framework achieves consensus and becomes the new normal science—education's transformation is likely to remain permanently unsettled. This is because education is inherently normative, not just descriptive. We're not just trying to understand how learning works (though that matters); we're trying to decide what kinds of humans we want to cultivate, what capacities we value, what forms of knowledge deserve recognition. These are political and ethical questions that can't be resolved through empirical research alone. Different visions of human flourishing will continue to generate different pedagogical commitments, even after the technical questions about AI's capabilities are settled.
The Magister Ludi and the Problem of Synthesis
Hermann Hesse's 1943 novel The Glass Bead Game (Das Glasperlenspiel) imagines a future centuries hence in which the highest intellectual achievement is not specialized expertise but the ability to synthesize knowledge across disciplines into elegant, aesthetically satisfying patterns. The novel's protagonist, Joseph Knecht, rises through the ranks of Castalia—a semi-monastic order devoted to cultural and intellectual preservation—to become the "Magister Ludi," the Master of the Game. The game itself is never fully explained (this is deliberate), but it involves representing connections between seemingly disparate domains: a mathematical theorem, a musical phrase, a historical event, a philosophical concept. The master player discerns patterns that cross disciplinary boundaries, revealing deep structural affinities invisible to specialists.
The game is, in other words, an aesthetic practice that treats knowledge itself as the medium. This is crucial: it's not about accumulating information or solving practical problems but about perceiving formal relationships, about experiencing the intellectual pleasure of synthesis. Players spend years learning the game's symbolic language, its aesthetic conventions, its subtle gradations of elegance and depth. The highest achievement is a game so perfectly constructed that it produces what the novel calls "a feeling of transcendent harmony, of universal interconnectedness."
But Hesse's novel is not a simple celebration of this rarefied intellectual practice. The entire second half of the book is Knecht's gradual disillusionment with Castalia's withdrawal from the world. He comes to see the Game as a kind of beautiful irrelevance, an intellectual monasticism that preserves culture while abandoning engagement with lived reality. The novel ends with Knecht leaving Castalia to become a tutor to a single student—choosing pedagogical relationship over abstract synthesis, concrete human development over aesthetic perfection. The irony is profound: the Master of the Game abandons the game because he recognizes that synthesis without application, pattern recognition without ethical commitment, is ultimately sterile.
This tension—between synthesis and engagement, between pattern recognition and ethical action, between aesthetic achievement and pedagogical responsibility—is precisely what's at stake in the AI transformation of education. The large language model is, in a sense, a crude version of Hesse's glass bead game. It has been trained on vast corpora of human knowledge and can generate texts that synthesize across domains, identify patterns, make unexpected connections. When you ask Claude or GPT-4 to explain quantum mechanics using musical metaphors, or to connect Foucault's biopower to contemporary data surveillance, or to trace the influence of Cubism on modernist literature, it can produce remarkably insightful syntheses—the kind of interdisciplinary pattern recognition that the glass bead game was supposed to cultivate.
But here's what the AI can't do: it can't tell you whether a particular synthesis matters. It can't distinguish between clever pattern-matching and genuine insight. It has no stake in the outcome, no investment in whether the knowledge is put to use, no ethical framework for determining when synthesis serves human flourishing versus when it becomes mere intellectual performance. These judgments require what Karen Barad calls "intra-action"—a recognition that the observer and the observed, the knower and the known, mutually constitute each other through their engagement. Knowledge, in Barad's account, isn't a representation of pre-existing reality but an enactment that brings certain phenomena into being while rendering others invisible. The choice of what to synthesize, what patterns to pursue, what questions to ask—these choices are not neutral. They reflect and shape relations of power, commitments about what deserves attention, visions of what kind of world is worth making.
Donna Haraway makes a similar point in her insistence on "situated knowledges" as a feminist epistemology. Against the pretense of view-from-nowhere objectivity, Haraway argues for acknowledging the partial, embodied, socially located character of all knowing. The AI's outputs are situated too—trained on datasets that reflect historical biases, optimized for statistical coherence rather than truth, shaped by the political economy of the companies that built them. But the AI can't recognize its own situatedness, can't acknowledge its blind spots, can't engage in what Haraway calls "the embodied nature of all vision." This self-awareness, this capacity for reflexivity about one's own standpoint, remains distinctively human work.
The Division of Research Services and the Infrastructure of Synthesis
At UC Riverside, where I serve as Director of Research and Technology for the Division of Research Services, we're attempting to build infrastructure that takes these theoretical insights seriously. Our division operates at the intersection of multiple technological domains: we have data scientists working with faculty on computational research methods, AI research librarians helping graduate students navigate large language models, makerspaces supporting physical prototyping, robotics labs exploring embodied AI, GIS specialists mapping spatial data, and media technologists creating multimodal scholarship. This isn't a random assemblage of trendy technologies—it's a deliberate attempt to mirror the kind of synthesis that students and faculty now need to develop.
The challenge isn't just technical (making sure the 3D printers talk to the data visualization software) but pedagogical and epistemological. How do you help a political science student understand that spatial analysis might reveal patterns invisible to traditional archival research? How do you convince a literature scholar that building a chatbot that simulates a novel's narrator could deepen their understanding of free indirect discourse? How do you create spaces where a robotics engineer and a disability studies researcher can collaborate on assistive technologies, each bringing forms of expertise the other lacks?
This is synthesis not as abstraction but as infrastructure. It requires what Suchman calls "artful integration"—the ability to recognize affordances across systems and broker connections between communities of practice that don't share common vocabularies. The AI research librarians we've hired aren't traditional reference librarians who happen to use AI tools; they're what I've come to think of as epistemological engineers, people who can move fluidly between computational and humanistic frameworks, who can speak to both the technical capabilities of large language models and their implications for knowledge production in specific disciplines.
Consider what this looks like in practice. A graduate student in environmental science approaches one of our AI research librarians with a straightforward question: she's trying to synthesize findings across hundreds of studies on microplastic contamination, and she wonders if an LLM could help. The librarian doesn't just show her how to use the tool—she helps her think through what kinds of synthesis are meaningful in her field versus what kinds would be methodologically inappropriate. They discuss the limitations of studies the AI was trained on (skewed toward English-language journals, Northern Hemisphere contexts, certain methodological approaches). They explore how to validate AI-generated summaries against domain expertise. They consider how to document the human-AI workflow in ways that maintain research reproducibility. The conversation moves back and forth between technical capabilities (what the AI can do) and situated epistemology (what counts as valid knowledge in this specific research context).
This is very different from what most people mean when they talk about "AI literacy." It's not a set of technical skills—how to write a good prompt, how to fact-check outputs, how to detect AI-generated text. Those skills matter, but they're instrumental. The deeper literacy involves understanding how different cognitive systems (human, computational, institutional) have different strengths and limitations, and how to orchestrate them into productive configurations. It's synthesis as lived practice, not abstract theory.
But here's what keeps me awake at night: this kind of sophisticated epistemological work is extraordinarily demanding. It requires comfort with ambiguity, facility with multiple disciplinary frameworks, willingness to work at the boundaries of expertise where no one is quite sure what the right questions are. These capacities correlate strongly with educational privilege. The students who arrive at UC Riverside with extensive preparation, who grew up in homes where intellectual debate was encouraged, who attended well-resourced schools—these students already have many of the dispositions needed for effective human-AI collaboration. The students who are first-generation college students, who work full-time while taking classes, who are still mastering academic English—these students face much steeper learning curves.
The danger is that AI, rather than democratizing access to intellectual tools, could create new forms of stratification. Not between those who have access to the technology (which is increasingly cheap or free) but between those who can use it to amplify their own intelligence versus those who use it as a crutch that prevents them from developing independent judgment. This isn't a technical problem. It's a problem of pedagogical design, institutional commitment, and resource allocation. And it maps onto existing inequalities in ways that should make us deeply uncomfortable.
The Zone of Proximal Development and the Architecture of Desirable Difficulty
Vygotsky's concept of the zone of proximal development offers a framework for thinking through these challenges. The ZPD is the distance between what a learner can accomplish independently and what they can accomplish with appropriate scaffolding. Effective pedagogy, in Vygotsky's account, operates in this zone—providing enough support that the learner can succeed, but not so much support that the task becomes trivial. Over time, with repeated practice in the ZPD, capacities that required scaffolding become internalized, and the learner's zone of independence expands.
AI has extraordinary potential as a ZPD tool—but only if we design the pedagogical context carefully. An LLM can translate expert discourse into more accessible language, suggest relevant resources the student hadn't encountered, generate examples that make abstract concepts concrete, provide feedback on drafts before high-stakes submission. These are all legitimate forms of scaffolding. But they become problematic when they bypass the student's own cognitive struggle entirely. If the AI writes the essay and the student merely edits surface features, there's no developmental movement. The ZPD has been eliminated, not optimized.
This is where what educational psychologists call "desirable difficulty" becomes crucial. Not all obstacles to learning are productive—some are simply frustrating barriers that should be removed. But certain kinds of difficulty are necessary for deep learning: the cognitive strain of retrieving information from memory rather than simply recognizing it, the challenge of explaining a concept to someone else, the discomfort of confronting evidence that contradicts your initial hypothesis. These difficulties force the kind of cognitive processing that produces durable understanding.
The pedagogical challenge is to design AI-assisted learning environments that preserve desirable difficulties while removing arbitrary barriers. This requires exquisite calibration. Too much friction and students simply circumvent the system (have the AI do everything, hide the fact); too little friction and no learning occurs. The "small frictions" I mentioned earlier—requiring students to document their prompts, identify three points of disagreement with AI outputs, explain their revision choices—are attempts at this calibration. But we're still in early experimental stages, and we don't yet have robust evidence about what works at scale.
Moreover, the optimal balance of difficulty varies by student, by discipline, by developmental stage. What constitutes desirable difficulty for a sophomore learning to write analytical essays is different from what a doctoral student needs when synthesizing a literature review. A physics problem that requires grappling with mathematical formalism involves different kinds of cognitive challenge than an ethics case that requires navigating moral ambiguity. One-size-fits-all prescriptions—"Always require students to write their first drafts without AI" or "Let students use AI freely and assess only final products"—miss this fundamental variability.
This is why the metaphor of the Magister Ludi matters. The role of the teacher isn't to transmit content (AI can do that) or even primarily to model expert practice (though that matters). It's to perceive where each student is in their developmental trajectory and to modulate the cognitive environment accordingly. It's to notice when a student is relying on AI as a genuine learning scaffold versus using it to avoid cognitive challenge. It's to create the conditions where students can develop what Hayles calls "cognitive assemblages"—stable, productive configurations of human-AI collaboration that amplify their capacities rather than replacing them.
The Fortress, the Festival, and the Question of Integrity
In the months following ChatGPT's release, many institutions responded with what can only be described as panic. Syllabi were hastily revised to include stern warnings about AI use. Plagiarism detection companies pivoted to offer "AI detection" services (which, it turned out, were unreliable at best and discriminatory at worst, flagging non-native English speakers at disproportionate rates). Some instructors returned to handwritten exams or oral defenses—a retreat to pre-digital assessment that felt less like innovation and more like capitulation.
This "fortress mentality"—the attempt to keep AI out of the classroom—was understandable but ultimately futile. Students were already using these tools, not because they were lazy but because the tools were genuinely useful. They helped with brainstorming, outlining, overcoming writer's block, translating complex concepts into simpler language. Trying to ban AI from education made about as much sense as trying to ban calculators from mathematics or word processors from composition. The question wasn't whether students would use AI but how they would learn to use it responsibly, critically, and creatively.
The alternative approach—what some have called the "festival" model—embraces AI as a collaborator while building in deliberate obstacles that make it difficult to use AI mindlessly. These obstacles aren't punitive; they're pedagogically generative. They slow down the process enough that genuine learning can occur, create opportunities for metacognitive reflection, and preserve the zone of proximal development.
But there's a shadow question lurking beneath all of this: what happens to academic integrity in a world where the boundary between legitimate collaboration and illegitimate outsourcing becomes impossible to police? The traditional concept of plagiarism assumes a clear distinction between one's own work and someone else's work. But what counts as "one's own work" when the thinking emerges from continuous dialogue with an AI system? If I ask Claude to help me clarify a vague intuition, then push back on its initial formulation, then incorporate elements of its response while rejecting others, then refine my position through multiple iterations—whose work is the resulting text?
Lucy Suchman's work on human-machine configurations suggests that the question itself may be poorly framed. Rather than asking "Who did the work?"—a question that assumes discrete, bounded agents with clear authorship—we might ask "What kind of learning occurred?" and "What capacities were developed?" These questions shift attention from policing boundaries to understanding processes, from detecting cheating to assessing growth. They acknowledge that all intellectual work is collaborative, that our thoughts emerge from conversation with others (human and nonhuman), and that originality consists not in producing ex nihilo but in the specific ways we engage with, transform, and build upon existing resources.
This doesn't mean "anything goes." There are meaningful differences between using AI as a thought partner (where the human remains actively engaged in judgment, revision, and critical assessment) and using it as a ghostwriter (where the human is merely a passive conduit). The difference isn't always visible in the final artifact, but it's profoundly consequential for what the student has learned. This is why process-based assessment—requiring students to document their workflow, explain their choices, identify places where they diverged from AI suggestions—becomes crucial. We're not assessing the text alone; we're assessing the relationship between the student and the text.
The Fermenting Future
If we take the wineskin metaphor seriously, the task before us isn't simply to patch the old containers or to pour the new wine more carefully. It's to construct entirely new vessels—institutional structures and pedagogical practices designed from the ground up to accommodate the fermentation that's already underway.
What might those new structures look like? They would almost certainly be more modular and personalized than our current system of semester-long courses and standardized majors. They would prioritize demonstration of capability over seat time and credit hours. They would treat AI literacy not as a technical skill but as a humanistic discipline—a way of thinking critically about intelligence, agency, and authority in an age when those concepts are increasingly contested.
Most radically, they would abandon the fiction that education can be a one-time event, a four-year investment that yields a lifetime credential. In a world where the capabilities of AI tools double every eighteen months, where entire professional skillsets become obsolete within a decade, the relevant question isn't "What do you know?" but "How quickly can you learn what you don't know?" The new wineskins of education must be flexible, resilient, and endlessly renewable.
At UC Riverside, we're experimenting with what we call "learning constellations"—dynamic, multi-node networks that connect students with appropriate resources (human and computational) based on their current projects and developmental needs. Rather than enrolling in a fixed course with a predetermined syllabus, students propose a project or question they want to pursue. An AI system (guided by human librarians and faculty advisors) suggests relevant resources: datasets, theoretical frameworks, methodological approaches, potential collaborators. The student pursues the project with scaffolding that adjusts in real-time based on their demonstrated capabilities. Assessment focuses on the student's ability to navigate complexity, synthesize across domains, identify and address their own knowledge gaps, and articulate what they've learned in ways that would be useful to others.
This isn't a replacement for traditional coursework—at least not yet. It's a supplement, an alternative pathway for students who have already demonstrated foundational competencies and are ready for more self-directed work. But it points toward what higher education might become: less a gatekeeping mechanism that sorts students into credential tiers and more a genuinely developmental environment that cultivates the capacities needed to thrive in radical uncertainty.
The students I work with—especially the graduate students who come to our division for research support—are already quietly making calculations about the value of traditional credentials. They see AI tools disrupting industry after industry, and they're not sure that a lengthy dissertation on a narrow specialty is the best preparation for an unpredictable future. They want to learn how to learn, how to adapt, how to collaborate with machines without being displaced by them. They're less interested in becoming experts in a single domain than in becoming what I think of as "epistemological nomads"—people who can move fluidly across disciplinary boundaries, who can recognize when a method from one field might illuminate a problem in another, who can orchestrate cognitive assemblages that combine human judgment with computational power.
This is the kind of synthesis that Hesse's glass bead game was supposed to cultivate. But unlike Castalia's withdrawn intellectualism, what we need now is synthesis in service of engagement—pattern recognition deployed for ethical action, interdisciplinary connection that addresses real-world challenges, aesthetic appreciation of knowledge that doesn't become mere contemplative escape. We need, in other words, to reclaim Joseph Knecht's final lesson: that the highest form of intellectual achievement is not abstract game-playing but the concrete work of human development, the messy, improvisational, deeply embodied practice of helping people become more fully themselves.
The Crisis as Opportunity
None of this is inevitable. It's entirely possible that institutions will continue to pour new wine into old wineskins until something bursts—probably not the technology, which will continue to advance regardless, but the credibility and relevance of educational institutions themselves. The fortress mentality persists in many quarters, sustained by legitimate concerns about quality, equity, and the preservation of humanistic values. These concerns shouldn't be dismissed. The danger of AI-driven deskilling is real. The risk of exacerbating inequalities is substantial. The potential for technologies built on exploitative labor practices and extractive data regimes to further entrench unjust power relations is undeniable.
But there's also reason for cautious optimism. The educators experimenting with process-based assessment, the librarians reimagining their roles as epistemological engineers, the institutions treating AI literacy as a core competency—these are the early adapters, the ones constructing new wineskins in real time. Their work is messy and uncertain, full of false starts and mid-course corrections. But it's also generative in the deepest sense: it's creating the conditions for learning that matters, for human flourishing in a world saturated with machine intelligence.
What makes me hopeful is not the technology itself but the conversations it's forcing us to have. For the first time in decades, we're having fundamental debates about what education is for, what forms of knowledge matter, what capacities we're trying to cultivate, what kind of humans we want to become. These conversations aren't new—they're as old as Plato's Academy, as contested as the medieval universities' battles over curriculum, as urgent as Dewey's progressive education movement. But they've been largely dormant in higher education for the past half-century, submerged beneath concerns about enrollments, rankings, and resource allocation.
The AI crisis is reopening questions we had prematurely closed. That reopening is uncomfortable, destabilizing, and necessary. Kuhn noted that paradigm shifts aren't clean or rational processes—they're lived through as crises, characterized by confusion, conflict, and the profound disorientation of having one's most basic assumptions called into question. But they're also moments of extraordinary creativity, when new possibilities become thinkable precisely because the old certainties have been shaken.
We're in that moment now. The ivory towers aren't crumbling. They're being repurposed, brick by brick, into something we don't yet have a name for—something that honors what was valuable in the old model while making space for the strange, turbulent fermentation of the new. Whether that transformation succeeds will depend not on the capabilities of our AI tools but on our collective willingness to reimagine what education itself is for. It will require synthesis not as abstract intellectual exercise but as concrete institutional practice—bringing together the technical and the humanistic, the computational and the embodied, the pattern-recognizing capacities of machines and the ethical judgment of humans into configurations that amplify what's best in both.
The game is still being invented. The rules are not yet written. But the players are gathering, the pieces are in motion, and the stakes have never been higher. This is education's glass bead game—not as Hesse imagined it, sealed off from the world in monastic contemplation, but played in real time, with real consequences, by people who understand that the patterns they're creating will shape not just what we know but who we become.
Ray Uzwyshyn is Director of Research and Technology for the Division of Research Services at UC Riverside, where he oversees the integration of AI, data science, makerspaces, robotics, GIS, and innovative media technologies across research and teaching. He holds a Ph.D. in Media Studies from NYU and is co-editor of "New Horizons in AI for Libraries" (De Gruyter, 2025). His social media posts and articles can be found on his profile on LinkedIn
Annotated Bibliography: The Glass Bead Game and AI in Education
Primary Theoretical Framework
Hayles, N. Katherine. Unthought: The Power of the Cognitive Nonconscious. University of Chicago Press, 2017.
Hayles develops the concept of "cognitive assemblages"—configurations of human and nonhuman actors that collectively produce cognition distributed across multiple agents and substrates. She distinguishes between human consciousness (which involves phenomenological awareness and embodied experience) and the "cognitive nonconscious" (which includes both human nonconscious processing and nonhuman computational cognition). This framework is essential for understanding how human-AI collaboration works at a cognitive level, challenging the assumption that either humans or machines are autonomous cognitive agents. The book provides the theoretical grounding for the essay's analysis of medium specificity—why different cognitive systems (human brains, large language models, institutional structures) have distinct capabilities, constraints, and blind spots. Hayles argues we're entering an era of "technogenesis" where humans and technologies co-evolve, each reshaping the other's capacities.
Hayles, N. Katherine. "Cognitive Assemblages: Technical Agency and Human Interactions." Critical Inquiry 43.1 (2016): 32-55.
This article introduces the concept of technical agency—the capacity of computational systems to initiate actions, evaluate outcomes, and modify their own processes. Hayles argues that technical agency is qualitatively different from human agency (which involves intentionality, consciousness, and ethical responsibility) but is nonetheless consequential. She develops a framework for understanding how human and technical agencies interact in contemporary cognitive ecologies, producing outcomes that neither could generate independently. This work provides the theoretical foundation for the essay's argument that effective human-AI collaboration isn't about humans doing X while AI does Y, but rather an emergent property of the human-AI assemblage. The article is crucial for understanding why we need to move beyond questions of "who did the work" toward questions about what kinds of cognitive configurations produce genuine learning.
Suchman, Lucy A. Human-Machine Reconfigurations: Plans and Situated Actions. 2nd ed. Cambridge University Press, 2007.
Suchman's foundational work argues that all action is "situated"—emerging from real-time interaction with environments rather than from pre-existing mental plans. She challenges cognitive science's traditional model of action-as-plan-execution, showing how plans function as resources for action rather than its determinants. Applied to education, this framework helps explain why prescriptive AI policies ("students must never use AI for drafting" or "students may always use AI freely") fail: effective practice requires continuous improvisation based on the specific situation, task demands, student capabilities, and available resources. The concept of "artful integration"—recognizing when a tool's affordances match a task's requirements—is essential for understanding sophisticated human-AI collaboration in learning contexts. Suchman's work on human-machine configurations emphasizes that agency is distributed and emergent, not localized in individual actors.
Mollick, Ethan R., and Lilach Mollick. "Using AI to Implement Effective Teaching Strategies in Classrooms: Five Strategies, Including Prompts." The Wharton School Research Paper (2023).
The Mollicks introduce the concept of "co-intelligence"—the emergent capability that arises when humans and AI systems collaborate effectively, producing outcomes neither could achieve independently. Their research demonstrates that the most successful human-AI collaboration doesn't involve rigidly delineated roles but rather improvisational interaction where each party's contribution shapes the other's. They document how expert users of AI engage in iterative dialogue where the AI's outputs reveal assumptions, blind spots, and alternative framings the human hadn't considered, while the human provides context, constraints, and judgment the AI lacks. This work is crucial for understanding why AI literacy can't be reduced to technical skills (prompt engineering, fact-checking) but must involve epistemic sophistication—understanding how different cognitive systems have different strengths and how to orchestrate them productively. Their empirical work on classroom implementation provides evidence that thoughtfully designed AI integration can enhance rather than undermine learning.
Vygotsky, Lev S. Mind in Society: The Development of Higher Psychological Processes. Edited by Michael Cole et al., Harvard University Press, 1978.
Vygotsky's concept of the "zone of proximal development" (ZPD) describes the distance between what a learner can accomplish independently and what they can accomplish with appropriate scaffolding from a more capable other. Learning occurs most effectively when tasks are calibrated to operate within this zone—challenging enough to require new cognitive strategies but not so difficult as to produce frustration and shutdown. This framework is essential for understanding how AI can function as a legitimate learning scaffold while also recognizing the danger of eliminating productive struggle entirely. The essay uses Vygotsky's ZPD to analyze what constitutes "desirable difficulty" in AI-assisted learning: certain cognitive obstacles are necessary for deep learning (memory retrieval vs. recognition, explaining vs. comprehending, confronting contradictory evidence), while others are arbitrary barriers that should be removed. Vygotsky's emphasis on social interaction and cultural mediation of learning also grounds the argument that all cognition is distributed—never purely individual—which helps reframe questions about academic integrity in the age of AI.
Paradigm Shifts and Institutional Transformation
Kuhn, Thomas S. The Structure of Scientific Revolutions. 50th Anniversary Edition. University of Chicago Press, 2012 [1962].
Kuhn's analysis of how scientific paradigms change provides the structural framework for understanding education's current crisis. He distinguishes "normal science" (puzzle-solving within an established paradigm) from "revolutionary science" (paradigm-displacing transformation triggered by accumulating anomalies). The pre-revolutionary period is characterized by what Kuhn calls "extraordinary science"—proliferating approaches, fundamental disagreements, loss of consensus about what constitutes valid practice. Higher education in 2025 exhibits all these symptoms: no consensus on AI-enabled pedagogy, competing assessment frameworks, disagreement about what counts as legitimate knowledge in an AI-saturated world. Kuhn's concept of "incommensurability"—the difficulty of translation between paradigms because they organize experience using incompatible conceptual frameworks—helps explain why faculty trained in traditional models literally cannot see what process-based assessment practitioners are doing. The essay uses Kuhn to argue that education requires not incremental adaptation but structural transformation, while also noting that education's normative character (disputes about values, not just facts) prevents the clean resolution Kuhn observed in natural sciences.
Hesse, Hermann. The Glass Bead Game (Magister Ludi). Translated by Richard and Clara Winston, Picador, 2002 [1943].
Hesse's novel imagines a future intellectual practice—the Glass Bead Game—that synthesizes knowledge across all disciplines into aesthetically satisfying patterns. The protagonist, Joseph Knecht, rises to become Magister Ludi (Master of the Game) but grows disillusioned with Castalia's withdrawal from engagement with lived reality. He ultimately abandons the Game to become a tutor to a single student, choosing concrete pedagogical relationship over abstract intellectual synthesis. The novel is crucial for the essay's argument about AI and education on multiple levels: (1) Large language models are crude versions of the Glass Bead Game—they can generate sophisticated syntheses across domains by recognizing patterns in training data, but they lack the ethical judgment to determine whether a synthesis matters or serves human flourishing. (2) The highest intellectual achievement isn't pattern recognition for its own sake but synthesis in service of engagement—exactly what education must cultivate in an AI age. (3) Knecht's realization that teaching is more valuable than game-mastery mirrors the essay's argument that educators' role shifts from content transmission to cultivation of human capacities that remain distinctively valuable: ethical judgment, embodied understanding, willingness to be vulnerable in intellectual commitment.
Feminist Epistemology and Situated Knowledge
Barad, Karen. Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning. Duke University Press, 2007.
Barad develops the concept of "intra-action" to describe how entities don't pre-exist their relations but emerge through their mutual engagement. Applied to knowledge production, this means that knowing isn't representing a pre-existing reality but rather enacting specific phenomena—bringing certain aspects of the world into focus while rendering others invisible. Barad's framework is essential for the essay's argument that the choice of what to synthesize, what patterns to pursue, what questions to ask are not neutral technical decisions but ethical and political commitments that shape what becomes thinkable. AI-generated syntheses reflect the datasets they were trained on, the objectives their designers optimized for, and the relations of power embedded in their architecture. But AI cannot recognize its own intra-actions—cannot acknowledge how its "knowledge" emerges from specific configurations rather than mirroring objective reality. This capacity for reflexivity about one's own situated position remains distinctively human work. Barad's quantum physics background also lends credibility to her critique of Cartesian subject-object duality, strengthening the philosophical foundation for distributed cognition frameworks.
Haraway, Donna J. "Situated Knowledges: The Science Question in Feminism and the Privilege of Partial Perspective." Feminist Studies 14.3 (1988): 575-599.
Haraway argues against the pretense of "view from nowhere" objectivity, insisting that all knowledge is produced from specific embodied positions within particular social relations. Rather than seeing this situatedness as a limitation to be overcome, Haraway proposes it as the foundation for a more honest, accountable epistemology. She advocates for "situated knowledges" that acknowledge their partiality while maintaining commitment to accuracy and truth. Applied to AI in education, this framework helps explain why large language models' outputs are never neutral—they're situated in training data that reflects historical biases, designed by teams with particular demographics and worldviews, optimized for metrics that privilege certain kinds of performance. The crucial asymmetry is that AI cannot acknowledge its own situatedness, cannot engage in reflexive critique of its own blind spots, cannot take responsibility for the knowledge it produces. This self-awareness and epistemic humility remain distinctively human capacities that must be cultivated through education. Haraway's work grounds the essay's argument that students need to learn not just how to use AI but how to recognize and resist the erasure of situatedness that AI's fluent outputs encourage.
Contemporary AI and Education
Mollick, Ethan. Co-Intelligence: Living and Working with AI. Portfolio/Penguin, 2024.
This book synthesizes Mollick's empirical research on AI adoption across multiple sectors, documenting patterns of effective and ineffective human-AI collaboration. He argues that we're in an "age of co-intelligence" where the relevant question isn't whether AI replaces humans but what new capabilities emerge from human-AI assemblages. The book provides practical frameworks for understanding when to involve AI in workflows (when pattern recognition at scale is valuable) and when to resist it (when embodied judgment, ethical reasoning, or creative risk-taking are essential). Mollick emphasizes that AI doesn't eliminate expertise—it changes what expertise means, shifting emphasis from knowledge possession to knowledge curation, from execution to judgment, from individual production to orchestration of cognitive resources. The book's case studies of education sector adoption provide empirical grounding for the essay's claims about process-based assessment and the importance of metacognitive reflection in AI-assisted learning.
Gulya, Jason. "Rethinking Assessment in the Age of AI." ASCD Education Update (June 2024).
Gulya, a high school English teacher, describes his experimental approach to process-based assessment where students begin by generating sophisticated AI-written essays and then demonstrate learning through critique and annotation. His framework treats the AI artifact not as an end product but as an "aperture"—an opening into deeper analysis. Students must identify where the AI makes logical leaps, reproduces conventional wisdom without interrogation, or generates authoritative-sounding prose that lacks genuine insight. Gulya documents that students who can create "distance" between their own thinking and the AI's output—who can see where ChatGPT is intellectually lazy or making unwarranted assumptions—demonstrate real understanding, while students who can't distinguish their judgment from the machine's reveal concerning gaps in critical thinking. This practical application of medium specificity theory provides concrete evidence that assessing the student-text relationship rather than the text alone can preserve academic rigor in an AI-saturated educational environment.
Selwyn, Neil. "What's the Problem with Learning Analytics?" Journal of Learning Analytics 6.3 (2019): 11-19.
Though focused on learning analytics rather than generative AI, Selwyn's critical analysis of data-driven educational technologies provides important context for understanding institutional responses to AI. He argues that technological solutionism in education often obscures rather than addresses underlying structural inequalities, creating new forms of surveillance and control while claiming to optimize learning. Selwyn's framework of "critical data studies" emphasizes that educational data is never neutral—it's produced through specific measurement regimes that reflect particular values about what counts as learning, success, or progress. This work grounds the essay's skepticism toward "AI detection" tools and fortress-mentality responses to ChatGPT, showing how technological responses often intensify rather than solve pedagogical challenges. Selwyn's insistence on centering questions of power, justice, and human agency in discussions of educational technology aligns with the essay's emphasis on equity concerns and the danger that AI could exacerbate existing educational stratification.
Warschauer, Mark, and Paige Tate. "AI in Education: Leveling Up or Scaling Down?" Educational Researcher 53.2 (2024): 96-105.
Warschauer and Tate's empirical research documents how AI tools are being adopted differentially across educational contexts, with well-resourced schools treating AI as a cognitive amplifier while under-resourced schools use it primarily for remediation and efficiency. They demonstrate that the "digital divide" is no longer primarily about access to technology but about access to pedagogical sophistication in using technology—what they call the "implementation divide." Students from privileged backgrounds are more likely to attend schools where teachers have time, training, and institutional support to design thoughtful AI integration, while students in under-resourced contexts face either AI bans (leaving them unprepared for AI-saturated workplaces) or uncritical AI adoption (producing dependency rather than capability). This research provides crucial empirical support for the essay's argument about equity concerns: the danger isn't AI itself but unequal access to the kind of epistemological sophistication required for effective human-AI collaboration.
Supporting Theoretical Works
Bjork, Robert A., and Elizabeth Ligon Bjork. "A New Theory of Disuse and an Old Theory of Stimulus Fluctuation." From Learning Processes to Cognitive Processes: Essays in Honor of William K. Estes 2 (1992): 35-67.
The Bjorks' concept of "desirable difficulties" describes learning obstacles that temporarily reduce performance but enhance long-term retention and transfer. Examples include retrieval practice (forcing memory recall rather than re-study), interleaving (mixing practice of different skills rather than blocking), and generation (producing answers rather than passively receiving them). This framework is essential for the essay's argument about AI scaffolding: not all obstacles to learning are beneficial, but certain types of cognitive struggle are necessary for deep learning. The challenge is designing AI-assisted learning environments that preserve desirable difficulties (forcing students to retrieve, generate, and articulate their own thinking) while removing arbitrary barriers (providing language support, resource access, technical assistance). The Bjorks' work provides cognitive science foundation for process-based assessment approaches that require students to document their thinking and identify places where they diverge from AI outputs.
Bowker, Geoffrey C., and Susan Leigh Star. Sorting Things Out: Classification and Its Consequences. MIT Press, 1999.
Bowker and Star's analysis of classification systems demonstrates how seemingly neutral organizational schemes embed values, create inclusions and exclusions, and have material consequences for who counts as legitimate, whose needs get met, and what becomes thinkable. They introduce the concept of "infrastructural inversion"—making visible the usually invisible systems that organize information and coordinate action. Applied to educational AI, this framework highlights how training datasets, architectural choices, and optimization metrics in large language models constitute classification systems that privilege certain forms of knowledge and marginalize others. The book grounds the essay's argument that AI literacy must include critical awareness of how AI systems categorize, what those categorizations make possible or impossible, and whose perspectives get encoded as normal versus deviant. Their emphasis on the ethical and political dimensions of seemingly technical decisions supports the essay's insistence that education must cultivate judgment about whether to use AI, not just how to use it.
Dreyfus, Hubert L., and Stuart E. Dreyfus. Mind over Machine: The Power of Human Intuition and Expertise in the Era of the Computer. Free Press, 1986.
The Dreyfus brothers argue that expert human performance fundamentally differs from computational rule-following. True expertise involves embodied intuition developed through extensive experience—pattern recognition that operates below conscious awareness and resists formalization into explicit procedures. They describe five stages of skill acquisition (novice, advanced beginner, competent, proficient, expert) where beginners follow rules consciously while experts respond fluidly to situational demands without deliberative reasoning. This framework challenges the assumption that AI, by processing more data or applying rules faster, can replicate expert judgment. Applied to education, it suggests that what students most need to develop—embodied judgment, contextual sensitivity, creative response to ambiguity—is precisely what resists algorithmic capture. The essay draws on this work to argue that AI can assist with information processing but cannot substitute for the kind of intuitive expertise that comes from sustained engagement with complex domains. Education's task is cultivating this human expertise while acknowledging AI's complementary strengths.
Historical and Philosophical Context
McLuhan, Marshall. Understanding Media: The Extensions of Man. MIT Press, 1994 [1964].
McLuhan's dictum "the medium is the message" establishes the framework for analyzing how different media technologies reshape human cognition and social organization. His distinction between "hot" media (high definition, low participation) and "cool" media (low definition, high participation) provides vocabulary for thinking about AI's cognitive temperature. Large language models are "hot" in some respects (delivering polished, high-definition text that demands little interpretive work) but "cool" in others (requiring active prompting, iteration, and judgment about output quality). McLuhan's insistence that media are not neutral tools but environments that alter perception and cognition grounds the essay's argument about medium specificity—why different cognitive technologies (human brains, writing systems, computational models) have distinct affordances that shape what's thinkable. His concept of "rear-view mirror" thinking—understanding new media through old categories—helps explain why fortress-mentality responses to AI (trying to ban it or police its use) fail to recognize that the educational environment has fundamentally changed.
Ong, Walter J. Orality and Literacy: The Technologizing of the Word. Routledge, 2002 [1982].
Ong analyzes how the transition from oral to literate cultures transformed human consciousness, memory, and social organization. He argues that writing is a technology that restructures thought—enabling abstract analysis, logical categorization, and decontextualized knowledge that oral cultures cannot easily sustain. Applied to AI, Ong's framework suggests we're experiencing another fundamental cognitive transition: from text-based knowledge work that humans performed alone to hybrid human-AI cognitive assemblages. Just as writing made possible new forms of thought (systematic philosophy, cumulative science, legal codification) while eliminating others (formulaic memory techniques, face-to-face dialectic, communal knowledge production), AI will enable new cognitive capacities while atrophying others. The relevant question isn't whether to resist this transition (literacy couldn't be resisted either) but how to cultivate the distinctively valuable forms of human cognition that remain essential. Ong's work grounds the essay's argument that we're not choosing whether to use AI but rather what forms of human capability we want to preserve and develop alongside it.
Dewey, John. Democracy and Education. Free Press, 1997 [1916].
Dewey's pragmatist philosophy of education emphasizes learning as active reconstruction of experience rather than passive reception of information. He argues that genuine education involves solving problems that matter to learners, testing ideas through action, and reflecting on the consequences of one's choices. Dewey's critique of "education as preparation"—the idea that schools should train students for future roles—remains relevant: in a rapidly changing world, the most valuable education cultivates adaptability, curiosity, and capacity for continuous learning rather than fixed knowledge or predetermined skills. Applied to AI in education, Dewey's framework suggests that the relevant question isn't what students should learn about AI but how they should learn with AI in ways that develop judgment, creativity, and agency. The essay draws on Dewey to argue for experiential, project-based learning where students encounter genuine problems that require orchestrating multiple resources (including but not limited to AI) and reflecting on what they learned through the process. This grounds the "learning constellations" model as aligned with progressive education's emphasis on student-directed inquiry.
Research Methods and Assessment
Wiliam, Dylan. Embedded Formative Assessment. 2nd ed. Solution Tree Press, 2018.
Wiliam's research on formative assessment emphasizes that effective feedback must be timely, specific, and actionable—focused on process rather than just outcomes. He distinguishes assessment of learning (summative evaluation) from assessment for learning (formative feedback that guides improvement). Applied to AI-assisted work, this framework suggests that traditional end-product assessment becomes nearly impossible (because AI can generate high-quality artifacts), but process-based assessment becomes more valuable. Wiliam's "five key strategies" for formative assessment (clarifying learning intentions, engineering effective discussions, providing feedback that moves learning forward, activating students as instructional resources for each other, activating students as owners of their learning) all remain possible and important in AI-saturated environments—but they require making students' cognitive processes visible in new ways. The essay draws on Wiliam to argue that requiring students to document their prompts, explain their revisions, and identify points of disagreement with AI outputs creates opportunities for formative feedback on exactly the capacities that matter: judgment, critical thinking, and metacognitive awareness.
Wiggins, Grant, and Jay McTighe. Understanding by Design. 2nd ed. ASCD, 2005.
Wiggins and McTighe's "backward design" approach begins by identifying desired learning outcomes, then determines what evidence would demonstrate those outcomes, and only then designs instructional activities. Applied to AI integration, this framework forces clarity about what capacities education aims to develop: if the goal is generating polished essays, AI obviates the need for student work. But if the goal is developing critical thinking, synthesis across sources, and ability to construct and defend arguments, then AI becomes a tool for revealing and developing those capacities through critique and annotation. The essay's discussion of process-based assessment draws on backward design's insistence that assessment methods must align with learning goals. In an AI age, this means assessing students' relationship to knowledge (how they evaluate sources, integrate perspectives, recognize limitations) rather than their ability to produce artifacts that AI can now generate.
Additional Recommended Reading
For readers interested in exploring the essay's themes further, the following works provide valuable context:
Benjamin, Ruha. Race After Technology: Abolitionist Tools for the New Jim Code. Polity, 2019. [On how technical systems encode and amplify racial inequality]
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. [On the material and political economy of AI systems]
Noble, Safiya Umoja. Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press, 2018. [On bias in algorithmic systems and its social consequences]
O'Neil, Cathy. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown, 2016. [On how algorithmic decision-making harms vulnerable populations]
Pasquale, Frank. The Black Box Society: The Secret Algorithms That Control Money and Information. Harvard University Press, 2015. [On opacity and accountability in algorithmic systems]
Turkle, Sherry. Alone Together: Why We Expect More from Technology and Less from Each Other. Basic Books, 2011. [On how digital technologies reshape human relationships and expectations]
Zuboff, Shoshana. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. PublicAffairs, 2019. [On the political economy of data extraction and prediction]