# What is an Academic Article For? The Production of Scholarly Work and its Meaning in an Agentic AI World **Authors:** Jason Miklian **Published in:** *Under Review*, 2026 **Canonical URL:** https://miklian.org/papers/what-is-an-academic-article-for-scholarly-work-agentic-ai ## Abstract Agentic AI tools are already embedded in every stage of the social science research lifecycle, yet social science scholars lack a framework for governing this integration. This article argues that productive governance requires answering a question the social sciences generally have left implicit but that scholars will have to confront head-on: what is an academic article for? We conducted two empirical tests to show that AI language increased 151% in social science-affiliated journals from 2022 to 2025, while at the same time almost none of the world’s top universities have coherent guidance for scholars on how to use AI in their work. This article therefore develops a typology of scholarly knowledge work, mapping a gradient from activities where AI capability is already substantial to those where the scholar’s cognitive engagement constitutes the contribution itself. It reframes agentic AI as a rupture in the scholar-knowledge relationship rather than an acceleration of existing pressures, and demonstrates that agentic AI operates differently across epistemological traditions. For positivist research, substance and process are separable, whereas for interpretivist scholarship, they are not. The article proposes three governance principles and identifies three scales of inequality that any framework must address, and closes with guidance on how to navigate this shift responsibly. **Keywords:** agentic AI, knowledge production, epistemology, social science, peer review, academic governance, intellectual formation, generative AI --- ## Full Text What is an Academic Article For? The Production of Scholarly Work and its Meaning in an Agentic AI World Jason Miklian Centre for Global Sustainability, University of Oslo jason.miklian@globe.uio.no Abstract Agentic AI tools are already embedded in every stage of the social science research lifecycle, yet social science scholars lack a framework for governing this integration. This article argues that productive governance requires answering a question the social sciences generally have left implicit but that scholars will have to confront head-on: what is an academic article for? We conducted two empirical tests to show that AI language increased 151% in social science-affiliated journals from 2022 to 2025, while at the same time almost none of the world’s top universities have coherent guidance for scholars on how to use AI in their work. This article therefore develops a typology of scholarly knowledge work, mapping a gradient from activities where AI capability is already substantial to those where the scholar’s cognitive engagement constitutes the contribution itself. It reframes agentic AI as a rupture in the scholar-knowledge relationship rather than an acceleration of existing pressures, and demonstrates that agentic AI operates differently across epistemological traditions. For positivist research, substance and process are separable, whereas for interpretivist scholarship, they are not. The article proposes three governance principles and identifies three scales of inequality that any framework must address, and closes with guidance on how to navigate this shift responsibly. Keywords: agentic AI, knowledge production, epistemology, social science, peer review, academic governance, intellectual formation, generative AI Introduction: Two Futures Imagine a political science department in 2030. Five years earlier, its faculty prohibited the use of artificial intelligence (AI) tools in research and teaching. Most cheered the principled decision: students would learn to think by thinking, to write by writing, to read by reading. The department preserved a model of intellectual formation that had served the social sciences for generations. The faculty are still publishing at the same rate, but whereas five years ago this put them in the top quartile of all departments, they now sit at the bottom. Their students have not fared much better. Graduates entered think tanks, governments, and international organizations where AI-assisted analysis is ubiquitous and using AI was a skill that they were expected to already know. They can’t generate AI-generated briefings or use the synthesis tools their colleagues rely on. They lack the vocabulary to critique AI outputs in their own domains because they were never taught to engage with these tools critically. As they fall further behind their AI-trained cohort, the students rue an increasingly useless degree that saddled them with debt but failed to provide critical skills to succeed. Now imagine the business department next door, which embraced AI without friction. Its faculty publish at twice the rate they did in 2025. But the literature this department generates is conceptually hollow, filled with incremental data mashed, smashed, and pattern-matched to existing frameworks, lacking the interpretive rigor that produces genuine insight. When a curious visiting scholar presses a junior faculty member on the causal mechanism underlying his most-cited paper, the junior struggles to reconstruct the reasoning or make a coherent argument at all. The claims were not his in any meaningful sense. The department did not collapse; to the contrary, central administration is thrilled with the metrics, and holds up the department’s progress as a best practice beacon. Yet, it has simply stopped producing its own knowledge. Its students produce polished and technically competent dissertations in half the time, but struggle to justify basic concepts. They are employed at high rates at companies who ask them to do much of the same work, but, at the first sign of economic trouble, these students are the first casualties. Most of us already see these scenarios bubbling up across the global academy as advances in AI march on while we debate principles or launch polemics. But building a framework for how to ethically engage AI in scholarship requires answering a question that most are content to leave implicit: what is an academic article for? Different scholarly traditions offer different answers, and those answers carry different consequences for how AI relates to the knowledge mission. An article can introduce a novel theoretical framework that reorders how a field understands a phenomenon, or incremental refinement of an existing theory, testing boundary conditions or extending explanatory scope. It can present new empirical data, whether from an original survey, an ethnographic immersion, or a computational analysis of large-scale datasets. It can synthesize a body of evidence for a policy audience, or it can develop methodology. Each of these purposes draws on a different constellation of scholarly activities, and AI's relationship to each is distinct. A literature synthesis, for example, draws heavily on information retrieval and pattern identification, activities with substantial AI capability. But a novel theoretical contribution depends on interpretive judgment, normative reasoning, and the generative friction of writing through an argument, activities where the scholar's cognitive engagement constitutes the contribution. A policy piece condenses technical academic arguments into more accessibility, a different AI strength. Treating all three as instances of a single activity called "research" produces frameworks that are either too restrictive for synthesis work or too permissive for interpretive work. The failure to differentiate is what allows the blanket AI-prohibitionist and celebratory-AI positions to persist and leaves little room for nuance. This article pursues this nuance by grappling with the heterogeneity of scholarly purpose. We do that by asking: for a given kind of knowledge, produced within a given epistemological tradition, aimed at a given audience, what does responsible AI academic engagement look like? This question is urgent. In 2021, AI writing tools were curiosities with limited scholarly utility. By early 2023, GenAI systems could produce passable literature reviews and identify relevant bodies of work across disciplines a researcher might never otherwise encounter, typically accompanied with a heavy moral disdain akin to cheating. By 2025, the most capable systems could engage meaningfully with theoretical frameworks, identify cross-disciplinary connections, and generate feasible synthetic data within methodological parameters. But the capability frontier is already shifting again: first to sustained analytical reasoning across complex, multi-layered arguments, and then to agentic operation, in which AI systems autonomously handle complex goals, execute multi-step research workflows, and adapt when initial approaches fail. Briefly, AI encompasses any system that mimics human reasoning; generative AI (genAI) is the subset that can create new content itself, and agentic AI is the emerging class that acts autonomously to plan, decide, and execute complex tasks. AI reasons about an academic argument and genAI gives feedback as a tool, but agentic AI can independently identify a research gap, survey the relevant literature, and generate hypotheses. It is (whether we like it or not) something closer to a collaborator. The social science community has not yet reckoned with what agentic AI means for the nature of knowledge creation. This compressed disruption of the research lifecycle has unfolded in roughly three years (Frumin, 2026; Dwivedi et al., 2023). The social sciences have not had three years' worth of reflection on what previous advances mean, let alone be prepared for the next wave. The literature remains dominated by practical questions of adoption and implementation (Crompton & Burke, 2023), with less attention to the epistemological and institutional consequences. If scholars do not set the agenda for how AI reshapes research, that agenda will be set for them by platform companies whose business models depend on maximizing adoption, by university administrators operating under ranking logics, and by funding bodies that reward volume over depth (Morriss-Olson, 2024). In short, if we fail to articulate how agentic AI should relate to our knowledge mission, actors with different priorities will fill the vacuum. AI use is a spirited debate. Some scholars embrace AI uncritically as a productivity multiplier that accelerates research. Those who do publish significantly more, use a wider variety of sources, and are much more cited (Liang et al. 2025). Reviewers are using AI to draft evaluations at scale (Potsdam, 2025), younger scholars can improve their productivity by 30% (Brynjolfsson et al., 2025), and those who use GenAI for funding proposals are more successful (Koslov, 2026). But the uncritical embrace ignores what happens to knowledge when the cognitive labor of interpretation and argumentation is outsourced to systems that process language without understanding meaning. It also ignores equity: AI tools are unevenly distributed across institutions, disciplines, and the global academy, and uncritical adoption risks compounding existing stratification (Bozkurt, 2024). Others argue for prohibition, treating any AI involvement as a threat to intellectual integrity. And with good reason. GenAI steals (Bartz vs. Anthropic, 2025). It plagiarizes (Kofinas et al. 2025). It hallucinates (Roig et al. 2026). It destroys the climate (Xiao et al., 2025). Its founders use their new wealth to undermine democracy (Miklian and Hoelscher 2026). But a prohibitionist position cedes agenda-setting power to actors with no stake in the integrity of social knowledge. A discipline that refuses to engage with AI ensures that the reshaping of its environment happens without our input. Most of us lie in the middle somewhere, seeing GenAI’s impressive capabilities and possibilities but wary of a process that morphs thinking into sheer content. But therein lies a fundamental mismatch. Many academics like to think of ourselves as in something akin to a Formula 1 race, finely tuned writing machines competing with our ideas on a tight set of guardrails. In this world, using AI is a dastardly cheating mechanism, an object of scorn. But what we’re actually living is more like It’s a Mad, Mad, Mad, Mad World (or Mad Max, if you prefer); a near-complete free for all where everyone is already playing by slightly different rules. In this real world, AI is a nitrous oxide boost that can either put you in the lead or blow up your engine if used in the wrong way. With AI already embedded in the research lifecycle at every stage, and its capabilities expanding rapidly, any framework premised on guardrail limitations will be obsolete before implementation. AI principles must orient with a moving target in our Mad, Mad, Mad, Mad World, differentiated by the kind of knowledge being produced and the epistemological tradition producing it. Therefore, we propose a framework that incorporates the heterogeneity of scholarly purpose into a typology that maps where AI advances the knowledge mission and where it substitutes output for insight. It examines how AI reshapes the research ecosystem differently for students in formation, scholars under evaluation, and institutions struggling to govern a technology that moves faster than their policy cycles. It also documents how AI exacerbates existing inequalities at global, institutional, and disciplinary scales, proposing three principles for governing AI across the research lifecycle that stay useful as AI capabilities evolve. The Nature of the University and Its Knowledge Mission The question of what AI does to research requires answering first: what is research for? What does it mean to "know" something when the synthesis was performed by a machine? And who gets to answer that question: the scholars who produce knowledge, or the institutions that house and monetize it? Universities operate under multiple, often contradictory, conceptions of their mission. The tension between these conceptions predates AI by decades, but AI forces it into public view because these tools serve competing visions of what scholarship is supposed to contribute. Three models of the university's knowledge mission coexist, and AI's relationship to each is distinct. The first is the Humboldtian ideal: the unity of research and teaching, knowledge as intrinsically valuable, and the university as a space of intellectual freedom and slow, difficult thinking (Tomicic, 2019). Under this model, research serves the discipline and the student. Scholarship advances theoretical frontiers, and the process of conducting it forms the scholar. The articles this model values most are those offering novel theoretical frameworks, deep interpretive arguments, and monograph-length engagements. These are precisely the forms of scholarship where the process of production is constitutive of the intellectual product. AI disrupts both functions simultaneously. If a doctoral student can generate a passable literature review in an afternoon, the formative experience of reading, struggling, and synthesizing across months is bypassed. Worse, AI can become a "cognitive crutch” that weakens the student’s reasonability the more it is used (Baracauci 2026). This can lead to a sudden collapse of general knowledge and innovation (Acemoglu et al., 2026). The volume of production is irrelevant if a scholar can use years of deep thinking to come up with One Great Work that changes a field or how we think about the world. The second is the public good model: research as service to society, knowledge creation as democratic infrastructure (Marginson, 2011; Calhoun, 2006). Here research serves policymakers, practitioners, and the public. The articles this model values most are evidence syntheses, policy translations, and accessible interventions in public debate. Under this conception, AI can accelerate the translation of scholarly findings into accessible formats, synthesize evidence across disciplines for policy audiences, and amplify the reach of public-facing research. It can also concentrate these capacities in well-resourced institutions, widening the gap between universities that produce policy-relevant knowledge and those excluded from the conversation. Here the point is in translation, to help others know what we do as scholars. The third is the managerial university: research as output, knowledge as measurable productivity, the institution as a competitor in global ranking systems (Shore & Wright, 2024; Münch, 2020). Under this model, research serves the researcher's career and the institution's position. The articles this model values are whatever the metrics reward: high citation counts, high publication volume, high journal impact factors. AI is a natural accelerant to all of these at the individual and institutional level, even if it simultaneously stresses the system as a whole. If the system rewards volume, AI will help scholars produce more of it. The managerial model does not ask whether the knowledge produced is sound or significant; it simply doesn’t matter. It asks whether the metrics are rising because volume is the point. These three philosophies sit uncomfortably together in many universities. A faculty member pursuing a slow, interpretive monograph (Humboldtian) does so within an institution that evaluates her by annual publication counts (managerial) and expects her work to demonstrate "societal impact" (public good). AI sharpens these contradictions because it serves the managerial logic most directly. The tools that accelerate output hollow out the formative and interpretive work that the other two models depend on. The collision is between conceptions of the university that have coexisted in separate boxes for decades, now forced into a single operational space. Of course, many of these dynamics predate AI. Publish-or-perish was already producing a literature of unmanageable volume: 3.4 million scientific papers were published in 2025, roughly 47% higher than a decade earlier, far exceeding the growth of active researchers (Hanson et al., 2023). The replication crisis had demonstrated that volume does not entail reliability (Ioannidis, 2005; Open Science Collaboration, 2015). Journal proliferation and the article processing charge economy had created publishing infrastructure in which the financial incentive to accept manuscripts competes with the obligation to reject weak ones, as the mass retraction of compromised articles from Hindawi journals in 2023 illustrated (InterAcademy Partnership, 2024). The "least publishable unit" was standard strategy well before AI made producing one cheaper (Shore & Wright, 2024). Agentic AI does more than just accelerate these pressures. When a scholar delegates literature review to a research assistant, the assistant is a fellow epistemic agent with whom intellectual accountability can be reciprocally negotiated. When a scholar delegates it to an AI, the relationship is different: the system processes language without holding commitments, without understanding what it retrieves, and without any stake in whether the resulting knowledge is sound. A scholar who writes every word of an article through months of sustained engagement with a body of literature stands in a different epistemic relationship to the resulting claims than a scholar who evaluates an AI-generated synthesis and incorporates it into an argument. In short, it is the struggle itself that generates the value, leaving the output itself as almost immaterial. “Almost” because the output and the process itself are inseparable, so the quality and innovativeness of the output itself can collapse (Acemoglu et al, 2026). AI engagement thus changes what it means to "have written" an article, to "know" a literature, or to "hold" a position. Treating AI as merely the latest tool to exploit them understates these new epistemological questions on the relationship between the scholar and the scholarly product, and risks corrosion of the institutional trust on which both teaching and research depend (Eaton, 2023; Cotton et al., 2024). The Research Ecosystem: Formation, Evaluation, and Governance Research happens within an ecosystem of actors, institutions, and incentive structures, and AI reshapes that ecosystem. This section identifies three cross-cutting thematic pressures: the problem of intellectual formation, the crisis in evaluation, and the gap between governance rhetoric and governance reality. First, the apprenticeship model has structured academic training for generations. Doctoral students learn to think by doing the difficult, time-consuming work of the early research lifecycle: conducting literature reviews by hand, coding qualitative data line by line, cleaning datasets, drafting and redrafting until the argument sharpens. AI compresses or eliminates many of these. For example, a literature review that once took three months can now be generated, at passable quality, in an afternoon. AI also creates new formative possibilities. A doctoral student who learns to critically evaluate AI-generated literature reviews, to probe AI-generated interpretations for shallow pattern-matching, and to use AI as an interlocutor while maintaining intellectual independence is developing analytical skills suited to the research environment she will actually inhabit. If collaborative reasoning with AI can serve the generative cognitive function that sustained writing traditionally served, then the challenge is to design new structures for intellectual development, not to preserve old ones for their own sake, because otherwise the market logic of efficiency will simply eliminate formative experiences rather than replace them. As AI handles work that research assistants once performed, the economic justification for hiring junior researchers weakens. A principal investigator who can use AI to code interviews, clean data, and draft preliminary analyses has less reason to fund RA positions. Fewer RAs means a narrowing of the doctoral pipeline at precisely the point where it is supposed to be widest (Frumin, 2026). Institutions have an obligation to treat intellectual formation as a value that AI governance must actively protect. Including AI-critical analysis, extended collaborative reasoning, and traditional interpretive apprenticeship, calibrated to the epistemological demands of each discipline, is necessary. AI also levels access beyond the traditional apprenticeship model, as in how students at under-resourced institutions gain synthesis capacity previously available only at elite research universities: these dimensions drive adoption behavior at least as powerfully as institutional policies (Liang & Tsai, 2024). Second, as most of us are painfully well aware, peer review is under severe strain. Submission volumes have risen sharply, reviewer fatigue is well documented, and the system's capacity to provide substantive evaluation is stretched thin (Potsdam, 2025; Hanson et al., 2023). AI has introduced an acute integrity crisis. A survey of 1,600 academics across 111 countries found that half used AI tools while peer reviewing manuscripts (Nature, 2025), and 21% of submitted peer reviews are fully AI-generated, identified through watermark detection (ICLR, 2026). The accountability structure that peer review is supposed to provide is collapsing. The dominant institutional response of detection is losing. Detection tools applied to submissions already show concerning accuracy limitations: one study found baseline detection accuracy of 39.5% that dropped to 17.4% when simple adversarial techniques were applied (Perkins et al., 2024). Bias compounds the problem: false positive rates of 61% for non-native English speakers, compared to 5% for native speakers (Liang et al., 2023). At least twelve elite institutions have disabled AI detection entirely, citing reliability concerns (Weber-Wulff et al., 2023). A more productive path redesigns peer review around the quality of intellectual engagement rather than the provenance of the text. This means allowing for guardrailed AI use in both manuscript preparation and review, developing explicit standards for what the reviewer's judgment adds beyond what AI can provide, and shifting evaluation criteria from process to substance (Checco et al., 2021). For example, a reviewer evaluating a large-N empirical study is primarily assessing methodological rigor and evidential support, activities where AI-assisted review can improve thoroughness. A reviewer evaluating an interpretive theoretical contribution is assessing the quality of meaning-making, the depth of engagement with existing debates, and the generative power of the conceptual framework, activities where AI-assisted review are more likely to muddle rather than clarify as the reviewer's situated expertise is the evaluative instrument. Third, the institutional governance of AI in research is characterized by a structural disconnect. Administrators impose AI-use policies that are at odds with the incentive structures the same institutions maintain. Humboldtian logic governs teaching policy, managerial logic governs research evaluation, and AI exposes the contradiction (Azevedo, 2025; Jiang, 2025). Today, 90% of institutions now report active AI initiatives, while only about 20% have published formal governance frameworks (EDUCAUSE, 2025). The gap is between institutions that are integrating AI into their operations at speed and institutions that have articulated principled governance for that integration. The former group is far larger, and it is proceeding without waiting for the latter. The external regulatory landscape adds pressure. The 2026 EU AI Act represents the first comprehensive legal framework for AI governance. Its obligations around AI literacy, transparency, and risk classification apply to educational and research contexts, directly affecting European universities and any institution collaborating with European partners. Whether the Act proves adequate to academic knowledge production remains open, but its existence means institutional AI governance can no longer operate in a purely self-regulatory space. Faculty governance of AI policy is essential: top-down administrative mandates consistently lag behind researcher realities and miss discipline-specific considerations. A policy adequate for a computer science department, where AI tools integrate naturally into existing workflows, may be entirely inadequate for an interpretive sociology program. Governance structures that treat "AI in research" as a single policy domain will continue to produce frameworks that researchers ignore in practice while affirming in principle. The Nature of Knowledge Creation: A Typology AI's impact upon knowledge creation varies substantially across scholarly work, but we lack vocabulary for mapping this variation, encouraging blanket positions to persist. This section proposes a typology to provide that vocabulary by accounting for epistemological pluralism. We identify seven distinct activities that constitute the research process, each with a different relationship to current AI capabilities (see Table 1). TABLE 1: A typology of scholarly work against AI capability, human contribution, and the labor shift. The first activity is information retrieval and synthesis: scanning literatures, identifying patterns across bodies of work, and summarizing the state of a field. AI capability here is high and improving rapidly. Current systems can process thousands of articles and generate structured summaries that would take a human researcher months to compile. The human contribution that remains essential is selection: deciding what to synthesize, evaluating source quality, and assessing relevance. The labor shifts from finding to judging, carrying a well-known risk. AI systems routinely fabricate citations: plausible-sounding references to nonexistent studies, complete with invented authors and DOIs, that have already appeared in published, peer-reviewed work (Walters & Wilder, 2023). If the scholar evaluates AI-generated syntheses that include fabricated sources detectable only through manual verification, then any efficiency gain is illusory. The judgment required is not merely intellectual (is this source relevant?) but forensic (does this source exist?), and from our own experience it is natural for a reader to consider the entirety of an academic article suspect if such hallucinations are found, not least because one doesn’t know which ideas were the authors and which were generated by the plausibility machine. The second is conceptual development: building new theoretical frameworks, coining concepts, identifying generative tensions between existing approaches. AI capability here is emerging and substantial. Systems can propose novel combinations of existing ideas, and collaborative exchanges can produce new framings that neither scholar nor system would have reached independently. The question is whether the scholar can evaluate which connections are generatively productive and which are only superficially plausible: an ability typically born of years of immersion in a field's debates and dead ends (Ide & Talamàs, 2025). The scholarly contribution shifts toward curatorial judgment, but it remains unclear what sort of pre-existing expertise is needed to use this tool in a valuable manner. Interpretive judgment is the third activity. This involves reading context, weighing competing accounts, and applying situated knowledge to ambiguous evidence. Interpretation requires what Polanyi (1966) called tacit knowledge: an embodied, contextual understanding developed through experience. Agentic AI systems can deliver interpretations, but not the situated commitments that ground interpretive claims in a specific intellectual tradition, a specific fieldwork experience, or a specific disciplinary debate (Hadjimichael et al., 2024). This terrain is contested. Some argue that language models are fundamentally incapable of meaning (Bender & Koller, 2020), while capabilities scholars suggest that sufficiently complex systems develop behaviors that blur the line between pattern-matching and understanding (Wei et al., 2022). Floridi and Chiriatti (2020) offer a middle path where AI systems operate in an informational space that requires new epistemological vocabulary rather than forced analogies to human cognition. Interpretive judgment remains, for now, a domain where the scholar's cognitive engagement constitutes the contribution. The fourth activity is normative reasoning: deciding what matters, why, and for whom, and making value commitments explicit. Agentic AI can map normative positions with sophistication, identifying ethical dimensions and laying out competing frameworks. The scholar's contribution is to stand on the landscape of values and defend the choice. One risk is in the sophistication of AI arguments and the risk of creating echo chambers by which a scholar’s repeated engagement with the AI leads them to believe something truly groundbreaking is taking shape, when instead what is happening is the AI’s inherent capacity for telling the prompter what they want to hear. Give it enough prompting and any AI will eventually tell you that even your worst ideas are nothing short of brilliant. The fifth activity is writing as thinking: the generative role of prose composition in clarifying and developing thought. Composing text is a cognitive process through which ideas are discovered, tested, and refined (Flower & Hayes, 1981). AI produces fluent text, but the cognitive transformation has no analogue in AI processing. The writer is changed by writing in ways that prompting an AI does not replicate (Zheng et al., 2024). Losing this process means losing a primary mechanism of intellectual development. Whether extended collaborative reasoning with AI can serve an analogous generative function is an open empirical question. The sixth is collaborative reasoning: sustained dialogue in which agentic AI functions as an interlocutor, offering extended exchanges in which the AI debates, surfaces counterexamples, and proposes alternative framings. This mode of engagement is closer to intellectual exchange with a colleague than to delegation of a task. If generative cognitive processes can occur through dialogue with an AI interlocutor (or generate multiple interlocutors in a “society of thought” model (Evans et al. 2026)), then the concern is about generative cognitive engagement rather than about writing. When a scholar refines an argument through dialogue with an AI system, and the result is stronger than what either could have produced alone, the intellectual product is the result of a collaborative reasoning process with a non-human participant, and this complicates traditional notions of scholarly authorship. It occupies a third space between the productivity framing (AI as tool) and the integrity framing (AI as threat), and it is the form of engagement scholars report as most intellectually valuable and least captured by current disclosure norms. The seventh activity is autonomous research agency, where multi-step workflows independently identify tasks, execute them across systems, and adapt when plans fail, without continuous human prompting. In agentic operation, the system sets intermediate goals and pursues them. The human contribution increasingly shifts from direction to oversight: evaluating completed outputs rather than guiding the process. The scholar's relationship to the resulting knowledge is mediated by post-hoc assessment rather than by dialogue, and the epistemological distance between the human and the knowledge produced grows accordingly. Whether this constitutes a new form of knowledge creation or a more efficient form of delegation is an open question, but ignoring it because the capability is recent would be a mistake the framework cannot afford. Understanding the Gradient The typology reveals a gradient (Figure 1 below). At one end, information retrieval and synthesis are automatable, and AI's contribution is substantial. At the other end, interpretive judgment, normative reasoning, and writing as thinking remain domains where human cognition is constitutive of the activity itself. In the middle, conceptual development, collaborative reasoning, and autonomous research represent emerging capabilities where AI's contribution is real but dependent on human evaluation, direction, and oversight. Value shifts depending on the epistemological tradition and the purpose of the article. FIGURE 1: AI Capability Spectrum across Scholarly Activities Note. Activities from Table 1 are positioned along a horizontal axis representing current AI capability, from limited (left) to high (right). Vertical placement above or below the axis is for readability and does not represent a second dimension. The seven activities span a gradient from domains where human cognition is constitutive of the scholarly contribution (left) to domains where AI already advances the work directly (right). Consider two scholars working in the same department. The first is a computational social scientist producing a large-N study that synthesizes behavioral data across six countries. Her article aims to identify robust empirical patterns and test a set of pre-registered hypotheses. The second is an interpretive sociologist conducting a three-year ethnography of a single community. Her article aims to develop a new conceptual vocabulary for understanding how that community experiences a particular form of structural change. For the first scholar, AI's contribution is substantial across the research lifecycle. Synthesis, pattern identification, data cleaning, and preliminary hypothesis testing can be advanced through agentic AI. The epistemological tradition is positivist, treating the knowledge claim as separable from the process that produced it: what matters is whether the finding replicates, regardless of who or what generated the initial analysis. For this scholar, AI is a powerful accelerant, and the governance challenge is primarily about quality control, transparency, and avoiding fabrication. The second scholar's contribution is constituted by the interpretive labor: the close reading of context, the sustained engagement with ambiguity, the theoretical vocabulary that emerges from years of immersion. In this hermeneutic epistemological tradition, how the knowledge was produced is constitutive of what the knowledge means. An AI-generated interpretation of fieldwork data lacks the situated commitments that ground the claim. For this scholar, AI is useful for peripheral tasks (literature retrieval, translation, formatting) but the core intellectual contribution is irreducibly hers. Between these poles sit policy syntheses, methodology articles, incremental theory extensions, and interdisciplinary bridging work: articles whose purposes map onto different positions along the gradient. A policy synthesis draws heavily on Activities 1 and 2 and may benefit from AI. A methodology article demands rigorous reasoning about procedures and assumptions that AI can assist but cannot substitute for. An incremental theory extension involves testing boundary conditions, where AI can help identify the relevant literature and propose test cases, but where the argumentative labor of demonstrating why the extension matters remains the scholar's. This differentiation is the framework's core contribution. Treating all scholarly work as occupying a single position will produce policies that are either too restrictive for synthesis work by choking off productivity, or too permissive for interpretive work by enabling displacement of the cognitive labor that constitutes the scholarly contribution. Agentic AI introduces three structural shifts that cut across the typology. The first is the pattern-recognition puzzle. Agentic AI can identify connections across literatures no human could read in a lifetime. Pattern-identification becomes knowledge through human interpretive labor through validation, contextualization, and evaluation of whether the pattern is meaningful or merely statistical. AI surfaces connections, while the scholar's role shifts toward curation and meaning-making (Ide & Talamàs, 2025; Zhao et al., 2025). The second is disciplinary boundary dissolution. AI tools do not respect the boundaries that have organized scholarly knowledge production. A scholar using AI can engage with other literatures in ways that previously required years of cross-disciplinary training (Xu et al., 2025). This is simultaneously an analytical gain and a threat to the disciplinary structures that organize hiring, journals, and tenure committees within often-outdated boundaries. At the same time, a scholar who can suddenly access knowledge across disciplines may lack the contextual knowledge to correctly evaluate what she finds. The third is knowledge obsolescence. If AI can produce a comprehensive literature review in hours, the shelf life of any published synthesis shrinks dramatically. A review article that took two years to compile used to hold value for a decade, but now it risks being outdated before peer review concludes. This temporal compression changes puts particular pressure on the slower forms of scholarly work such as monographs, longitudinal studies, and interpretive projects that unfold across years rather than months. These are precisely the forms of scholarship that the Humboldtian model most values, and precisely the forms most vulnerable to a system that increasingly rewards speed (Hanson et al., 2023). That said, the line between pattern-matching and interpretation is less clear than humanists assume (Shanahan, 2024), and the emergent capabilities literature suggests that qualitative shifts in model behavior can appear unpredictably as systems scale (Wei et al., 2022). The arrival of agentic AI has already pushed the boundary again, so whether the gradient holds is itself an empirical question, one the field should investigate with the same rigor it applies to other epistemological claims. Three Scales of Inequality: Building the Evidence Base AI affects researchers unequally at global, institutional, and disciplinary levels. Reliable internet, institutional AI subscriptions, and computational resources remain concentrated in the Global North, distributed in ways that mirror and deepen existing academic stratification (World Economic Forum, 2023; UNESCO, 2024). Moreover, AI models are trained predominantly on English-language scholarship, and the frameworks, citation patterns, and argumentative conventions that large language models have absorbed are disproportionately Anglophone. When a scholar in Bogotá or Dakar uses an AI system to develop an argument, the system draws on a knowledge base that treats Anglophone social science as the default register of scholarly reasoning. AI may flatten the intellectual diversity the global academy depends on. Distinct scholarly traditions, including French social theory, German philosophical argumentation, and Latin American dependency frameworks, employ different modes of reasoning, different standards of evidence, and different relationships between theory and practice. If AI-assisted scholarship converges towards Anglophone academia, that convergence represents an epistemic loss (Birhane, 2021; Mohamed et al., 2020). The potential result is a two-tier global academy with AI-augmented institutions in the North producing at accelerating volume and AI-excluded or AI-homogenized institutions in the South falling further behind or losing the distinctiveness of their contributions. This framing, however, risks treating Global South scholars as recipients of inequality rather than agents responding to it. To wit, India's national AI Mission has deployed over 34,000 publicly funded GPUs as a "common compute" pool to democratize access for researchers, startups, and public-sector innovators (GoI, 2025). South-South cooperation models increasingly bypass Northern infrastructure dependency, emphasizing knowledge exchange on implementation and governance adaptation (Carnegie, 2026). Global South scholars may have principled reasons to limit AI integration. If epistemological traditions grounded in oral knowledge practices, community-embedded research relationships, or non-Anglophone theoretical vocabularies stand to lose their distinctiveness through AI-mediated homogenization (or are stolen by AI firms), then a deliberate choice to restrict AI's role may represent epistemic self-defense rather than technological deprivation. Within universities, tenured professors with AI budgets and unfunded doctoral students relying on free-tier tools operate under vastly different capability ceilings. Moreover, natural sciences absorb AI tools as extensions of existing computational infrastructure, but many social science departments face material barriers to adoption. A shadow economy has developed in many departments where researchers pay out of pocket for premium AI tools, API access, and computational resources their institutions fail to provide (Bozkurt, 2024). This stratification shapes research capacity over time, especially when universities sign exclusive deals with firms to be their sole approved provider, shifting access gaps from subscription costs to compute infrastructure and technical capacity. Unfortunately, the fields best positioned to critically analyze AI's social implications, including the humanities and qualitative social sciences, are the least resourced to engage with the tools. Humanities departments operate under minimal technology budgets, and many face epistemological resistance to AI that is partly principled and partly a consequence of unfamiliarity (Chatzichristos, 2025), as in concerns that AI pushes qualitative research toward a latent positivism (Davidson & Karell, 2025). For interpretive traditions, the mode of knowledge production is part of what the knowledge means. But if the natural sciences absorb AI tools readily and publish faster, and if reward systems continue to equate volume with value, social science scholars face compounding disadvantages through less institutional support, less AI capacity, slower output, and declining status. This demands deliberate cross-disciplinary capacity building by creating spaces where critical analysis of AI and practical engagement with it inform each other. To build an evidence base, we analyzed 2750 full-text open-access articles in 15 social science journals from 2022-2025, spanning psychology, political science, education, public health, sociology, communication, and interdisciplinary social science. The density of AI-associated linguistic markers doubled in this period: the mean frequency of 32 AI-associated vocabulary items rose from 8.2 per 10,000 words in 2022 to 20.5 in 2025: a 151% increase in three years. The sharpest acceleration occurred between 2023 and 2024, coinciding with the mass adoption of ChatGPT and its successors. The marker trajectories reveal that AI-influenced academic articles are more likely to describe findings as “crucial,” frameworks as “comprehensive,” approaches as “holistic,” and analyses as “nuanced.” It is more likely to “foster” outcomes, “navigate” complexities, and “leverage” resources. This is a detectable shift toward the specific language that LLMs produce when prompted to write academic prose. See Appendix 1 for full data and methodology. Figure 1: AI-associated marker density across 2,750 articles, 2022–2025. Shaded area shows interquartile range. Figure 2: Individual AI-associated markers across 2,750 articles, 2022–2025. Almost all journals showed a substantial increase in AI vocabulary between 2022 and 2025. Humanities and Social Sciences Communications leads at +210%, followed by Frontiers in Political Science (+206%), and BMC Public Health (+191%). Two outliers are noteworthy. American Political Science Review (+28%) and Business Ethics Quarterly (-29%) were least affected. These two journals share a common feature: they publish heavily theorized, argumentative scholarship where the voice of the author is central to the intellectual contribution. Their resistance to AI vocabulary suggests that AI’s linguistic fingerprint is strongest where writing functions as synthesis rather than as original argumentation. Figure 4: AI-associated vocabulary frequency by journal. Warmer colors indicate higher frequency. The AI linguistic fingerprint is a robust, cross-disciplinary phenomenon, and journals that publish more synthesis-heavy, empirical work show higher AI marker densities, while journals emphasizing original theoretical argument show lower densities. The AI-influenced social science article in 2025 is more likely to describe findings as "crucial" or "pivotal," frame analysis as "comprehensive" or "nuanced," and claim to "underscore" or "illuminate" patterns, reflecting AI’s stylistic preferences across fields that have historically maintained distinct rhetorical traditions. Figure 5: AI-associated vocabulary, 2022 vs. 2025, with percentage change. n = 2,750. These findings provide empirical support for two claims. First, AI is already embedded in the research lifecycle, as in the doubling of AI-associated vocabulary across social science articles. Second, AI is flattening intellectual diversity via epistemic homogenization, at the level of the linguistic medium through which ideas are expressed. Journals that publish more synthesis-heavy, empirical work show higher AI marker densities. The journals with the strongest interpretive and normative traditions in the sample show the lowest. This is consistent with the gradient: AI’s linguistic influence tracks the typology’s prediction about where AI capability is highest. To test if institutional governance of AI in research is characterized by a structural disconnect, we next coded AI policies at 70 of the world’s top-ranked universities across 16 countries. None of the 70 institutions differentiated AI policy by discipline, epistemological tradition, or type of scholarly work. None addressed AI use in peer review, or included a revision or sunset clause. Of the 63 institutions with publicly findable AI policies, zero systematically differentiate expectations by discipline or epistemological tradition. Every institution treats “AI in research” as a single policy domain; the typology of scholarly work proposed in the article has no analogue in any institutional governance document we examined. Current institutional practice treats AI governance as a document to be written once and posted, but given that AI capabilities have shifted dramatically every 12–18 months since 2022, policies codified even as recently as 2024 are already governing a technology landscape they were not designed for. See Appendix 1 for more. An Accountability Framework: Three Principles Taken together, this baseline evidence reflects the deep need for a new approach. Therefore, we propose three principles for governing AI in the research lifecycle. They are designed to remain useful as capabilities evolve, yet are admittedly uncomfortable or even provocative principles for many scholars to engage with. Effectively, they are principles designed for us to engage with the AI landscape as it has already become, in a frontier that moves faster than policy cycles. Rules often face obsolescence before implementation, as the tools they target improve and the detection mechanisms they rely on degrade (Perkins et al., 2024). The three principles are described below and summarized in Table 2. TABLE 2: Three Governance Principles for AI in the Research Lifecycle Principle 1: Accountable knowledge, evaluated on substance This principle fuses two related commitments. The first is that research outputs should be evaluated by the quality of the intellectual contribution. A brilliant, AI-assisted insight represents better scholarship than a mediocre purely human one. We fully recognize that this statement may be uncomfortable or even a provocation to scholars uncomfortable with the thought of AI as a contributor. Next, reviewers should assess arguments on their merits, with AI use as disclosed context. The current fixation on process (was AI involved?) distracts from substance (is the knowledge sound?). Every major publisher now prohibits AI from authorship on accountability grounds (COPE, 2023). We agree with the accountability rationale while arguing that the fixation on process extends well beyond it, into a general suspicion of AI-assisted work that is neither epistemologically justified nor sustainable as AI improves. If process does not matter, what stops a researcher from submitting pure AI output as scholarship? The second commitment offers an answer: the author must be able to defend their work under sustained scrutiny. If no human can reconstruct the reasoning and respond to challenges, the work fails regardless of whether AI was involved. A scholar who uses AI extensively and can defend every claim has met the standard. A scholar who writes every word by hand and cannot defend the argument has not (Hosseini & Resnik, 2024). This principle works especially for epistemological traditions where the knowledge claim is separable from the process that produced it. For these traditions, evaluating "substance" means evaluating the finding, the method, and the argument on their own terms. For interpretive social sciences, the relationship between substance and process is more intimate. A claim produced through two years of ethnographic immersion carries epistemic weight that is partly constituted by the process of immersion itself. The same propositional content generated through AI synthesis would lack this grounding. In hermeneutic and interpretivist traditions, how the scholar arrived at a claim is part of what the claim means. For synthesis-heavy and empirical work, "evaluate on substance" means evaluating the quality of the evidence, the rigor of the method, and the soundness of the argument, with AI involvement as disclosed context. For interpretation-heavy work, "evaluate on substance" must include evaluating the epistemic process, because the process is constitutive of the substance. A reviewer evaluating an interpretive contribution is assessing the quality of meaning-making, and meaning-making is irreducibly tied to the situated commitments of the scholar who produced it, making explicit what "substance" means. This principle extends to evaluation: if reviewers are not held to the same standards of intellectual responsibility as authors, then the system is itself unaccountable. As a result, a disclosure culture in which honesty invites suspicion is worse than no disclosure at all. Disclosure should describe what AI contributed to the intellectual work, e.g. brainstorming, synthesis, collaborative reasoning, and/or drafting. Current disclosure norms tend to demand categorical statements ("AI was used/not used") when the reality is a continuum of engagement that varies by research stage (Lund et al., 2023). The field must normalize comprehensive disclosure rather than incentivizing strategic omission for fear that the binary “use” disclosure will imply the worst and most extensive possible use of AI. Principle 2: Protect formation The typology and the ecosystem analysis converge on the compression of intellectual formation. If the tasks by which junior scholars develop analytical capacity are automated, and if the economic justification for hiring the research assistants who would otherwise perform those tasks weakens, then the scholarly pipeline is at risk. Accountability governs the quality of knowledge, but formation governs whether there will be scholars capable of producing it (Acemoglu et al., 2026). Therefore, scholarly communities have an obligation to treat intellectual formation as a value that AI governance must protect. A doctoral student who learns to critically evaluate AI-generated literature reviews, to identify confabulated citations, and to use AI as an interlocutor is developing analytical skills suited to the research environment she will actually inhabit. If collaborative reasoning with AI can serve the generative cognitive function that sustained writing traditionally served, then the formation challenge is to design new structures for intellectual development, calibrated to the epistemological demands of each discipline, not to preserve old structures for their own sake. The commitment is that formation is a design problem requiring institutional investment through funded positions for junior researchers, AI-critical curriculum requirements, and mentorship structures that integrate AI engagement with traditional apprenticeship. AI also democratizes formative experiences that the traditional apprenticeship model distributed unequally. A first-generation doctoral student at an under-resourced institution gains synthesis and analytical capacity that partially compensates for the absence of extensive mentorship networks. The formation principle requires that institutions audit how AI adoption affects intellectual development, and ensure that efficiency gains in research do not come at the cost of developmental losses in training. Principle 3: Equitable and adaptive governance Institutions and funders have obligations to ensure AI tools are accessible across the inequality scales described in Section 5. Policies that restrict AI use without providing equitable access produce regressive outcomes and compound existing disadvantages. A university that prohibits students from using AI tools it cannot afford to provide has imposed a prohibition whose costs fall disproportionately on the scholars least able to bear them. Governance structures must also include mechanisms for revision, which is not a separate principle but important when crafting these frameworks. Any framework adequate to 2026 will require updating by 2028. Sunset clauses, regular review cycles, and cross-institutional learning processes are constitutive features of governance designed for a moving target. This principle extends to evaluation reform which will need to happen in parallel. The metrics by which research is assessed must evolve alongside the tools by which it is produced. Engagement metrics, reward structures that value integrative work, and preprint cultures enabling post-publication evaluation represent practices better suited to an AI-integrated research lifecycle than volume-based assessment (Hanson et al., 2023; Checco et al., 2021). Discussion: Owning the Agentic AI Transition Scholars must shape how AI transforms knowledge creation, because the alternative is having that transformation shaped by actors with no inherent stake in the integrity of knowledge. AI companies deserve the same critical scrutiny the social sciences apply to any other form of corporate influence on public institutions. Free API credits and research access programs create dependency relationships, normalizing particular tools and locking workflows into proprietary ecosystems. Academic partnership programs, such as OpenAI's NextGenAI consortium, provide universities with infrastructure that carries implicit expectations about adoption (OpenAI, 2024). The cumulative effect is a soft infrastructure of influence shaping what questions get asked and what norms get established, without the visibility that would attend formal policy advocacy (Whittaker, 2021). In 2025, major AI companies each individually spent more on federal lobbying than the entire independent AI safety research field received in grants (Silicon Canals, 2025). The asymmetry mirrors familiar patterns of corporate influence on academic research (Slaughter & Rhoades, 2004). The political economy of academic AI deserves the same analytical attention. Who profits from AI integration in the research lifecycle? What governance structures can protect the public-good function of research from capture by actors whose primary obligations are to shareholders? For example, many predict that AI will improve scholarship in synthesis-heavy work and degrade it in interpretation-heavy work. That prediction is testable. Do AI-assisted papers produce more or fewer novel findings? Large-scale bibliometric analysis, combined with expert evaluation panels assessing intellectual contribution rather than counting citations, could provide baseline data against which future shifts can be measured. Also, how does training with AI shape the intellectual development of the next generation of scholars? The cohort divide offers a natural research design. Longitudinal studies tracking analytical capacity and intellectual independence across successive doctoral cohorts would provide evidence the current debate, conducted almost entirely through assertion and anecdote, urgently needs. A subsidiary question follows from the analysis of collaborative reasoning: can extended AI interlocution serve the generative cognitive function traditionally served by writing, and if so, under what conditions? Academia also needs rigorous, independent research on who profits from AI integration in the research lifecycle and what norms are being shaped by corporate investment. This research must be funded independently of the companies whose influence it examines, which is itself a governance challenge. If the social sciences cannot articulate what an academic article is supposed to contribute, they cannot govern how AI participates in producing one. The framework proposed here argues that articles serve different purposes within different epistemological traditions, and AI's appropriate role varies accordingly. The social sciences have always housed multiple epistemological traditions, and the attempt to impose a single standard of evaluation or a single model of knowledge production has always served some traditions at the expense of others. AI threatens to accelerate this imposition by privileging the forms of scholarship most compatible with computational processing. The framework resists that acceleration by insisting that governance must be differentiated by the kind of knowledge being produced. These principles will require revision as capabilities evolve and empirical evidence accumulates. AI excels at processing truth-claims: identifying patterns, verifying consistency, surfacing evidence. What it does not do is the meaning-making that transforms information into knowledge worth having. Whether social science protects that meaning-making capacity, or quietly abandons it in favor of efficiency, will depend on the choices made in the next few years. Moreover, the scholars best equipped to write a framework for AI in research are the ones already deeply entangled with AI. We invite the reader to apply the framework's own standard: evaluate the argument on its merits, note how AI may have been used in its production, and ask whether the resulting knowledge is sound. We add that the great irony of the debate today is that while it is primarily qualitative scholars that are fearing the rise of AI-assisted work, it is much more likely that quantitative scholars will first feel the pinch, as AI can reliably generate regression analyses and build datasets quickly in ways that it cannot (and perhaps can never) build real human-based fieldwork. For scholars who take seriously the knowledge mission of the academy, the question of how to engage with AI is serious intellectual work, and all scholars will benefit from direct engagements on the merits. We prefer to engage with this transition openly rather than pretend it is not happening, and to hold our contributors to standards that protect what makes interdisciplinary social inquiry valuable in the first place. --- ## Citation Jason Miklian. "What is an Academic Article For? The Production of Scholarly Work and its Meaning in an Agentic AI World." *Under Review*, 2026. ## About the Author Jason Miklian is Senior Researcher at the Centre for Development and the Environment (SUM), University of Oslo. ORCID: [0000-0003-1227-0975](https://orcid.org/0000-0003-1227-0975). Google Scholar: [profile](https://scholar.google.com/citations?user=RHlevGEAAAAJ&hl=en).