Abstract
Critical AI literacy—the practice of using artificial intelligence to interrogate power while simultaneously interrogating the lens of AI itself—does not unfold in a neutral epistemic environment. This article develops a theoretical framework for understanding how power responds to critical analysis of AI and systems of control through two interconnected modalities: justification and prevention. Justification mechanisms—including manufactured illiteracy, agnotology, ethics washing, critical washing, and safety discourse—create discursive conditions in which critique appears unnecessary, extremist, or already accounted for. Prevention mechanisms—including epistemic capture, algorithmic groupthink, corporate capture of knowledge systems, surveillance-induced chilling effects, institutional mimicry, and epistemicide—structurally foreclose the material and institutional conditions under which critique can be produced and heard. Drawing on scholarship in critical AI studies, agnotology, surveillance studies, and the political economy of knowledge, the article argues that the most sophisticated strategies of power do not oppose critique but incorporate it, creating a recursive trap in which the invitation to “be critical” functions as a mechanism for preempting substantive change. The article concludes by proposing an adversarial model of critical analysis that resists absorption into existing institutional frameworks and insists on binding consequences rather than merely discursive acknowledgment of harm. Implications for researchers, educators, policymakers, and civil society actors are discussed.
Keywords: critical AI literacy, agnotology, ethics washing, epistemic capture, surveillance, epistemicide, systems of control, power-knowledge
1. Introduction
The preceding companion article (hereafter “the companion piece”) outlined a practice of critical augmentation: using artificial intelligence as a lens to analyze systems of control while simultaneously analyzing AI itself as a system of control. That article argued that AI is neither a neutral oracle nor a simple tool but a sociotechnical construct shaped by power relations, and that the path to understanding the world through AI requires a shift from being a consumer of AI outputs to becoming an epistemic agent who uses AI to strengthen critical judgment.
However, this practice does not unfold in a vacuum. Power does not simply tolerate being analyzed; it develops sophisticated, often invisible counter-strategies. These strategies operate on two interconnected fronts: justification (producing reasons why critical analysis is unnecessary, dangerous, or already happening) and prevention (structuring the conditions of knowledge production so that critical analysis becomes impossible, career-ending, or simply irrelevant).
This article develops a systematic framework for understanding these counter-strategies. It draws on scholarship across agnotology, critical AI studies, political economy, surveillance studies, and epistemology to map the repertoire of power’s response to critical AI literacy and systems analysis. The central argument is that the most sophisticated strategies of power do not oppose critique but incorporate it, creating a recursive trap in which the invitation to “be critical” functions as a mechanism for preempting substantive change. Understanding this trap is a precondition for developing forms of critique that resist absorption.
The article proceeds as follows. Section 2 establishes the theoretical framework, situating critical AI literacy within broader debates about power-knowledge and the political economy of knowledge production. Section 3 maps the architecture of justification, examining five discursive mechanisms through which power renders critique unnecessary or illegitimate. Section 4 analyzes the machinery of prevention, identifying six structural mechanisms through which power forecloses the conditions of critical analysis. Section 5 theorizes the recursive trap—the incorporation of critique into the very systems it seeks to transform. Section 6 proposes an adversarial model of critical analysis as a response to this trap. Section 7 concludes with implications for research, pedagogy, and policy.
2. Theoretical Framework: Critical AI Literacy and the Political Economy of Knowledge
2.1 Power-Knowledge and the Sociotechnical Construction of AI
The theoretical foundation of this analysis is the Foucauldian concept of power-knowledge, which holds that power and knowledge are not separate entities but are mutually constitutive. Knowledge is always produced within relations of power, and power operates through the production and circulation of knowledge. Applied to AI, this framework directs attention to the ways in which AI systems—far from being neutral technical artifacts—embody and reproduce specific configurations of power-knowledge.
Critical AI literacy, as developed in the companion piece and elsewhere, goes beyond technical skills to interrogate the power dynamics within AI. It recognizes that AI models are trained on data containing historical biases and asymmetrical power relations, and that they can reproduce patriarchal, anti-feminist, or other dominant ideologies embedded in their training data (Bender et al., 2021; Noble, 2018). It also recognizes that AI systems may avoid confronting dominant power structures under a guise of “neutrality” unless directly challenged—a phenomenon we might call institutional mimicry at the algorithmic level.
2.2 The Political Economy of AI Research
A second theoretical anchor is the political economy of knowledge production. Big Tech firms have become the dominant funders of AI research, creating what has been termed “pervasive financial entanglements with universities, think tanks, and the policy advisory ecosystem” (Abdalla & Abdalla, 2021). This produces an “epistemic monoculture” where critical perspectives struggle to gain traction and the boundaries between independent expertise and industry advocacy grow increasingly porous.
This is not merely regulatory capture in the conventional sense. It is, as critical scholars have argued, “a capture of the very knowledge systems upon which democratic governance depends” (Whittaker, 2021). The infrastructure of expertise—the journals, conferences, funding streams, and career pathways that constitute the field of AI research—is already aligned with industry interests. Critical scholars face a structural disadvantage that is not reducible to individual bias or bad faith.
2.3 Agnotology as a Complement to Epistemology
A third theoretical resource is agnotology, the study of the strategic production of ignorance (Proctor & Schiebinger, 2008). While epistemology asks how we know what we know, agnotology asks how we don’t know what we don’t know—and, crucially, who benefits from that not-knowing. Oreskes and Conway (2010) demonstrated how corporate actors in the tobacco, fossil fuel, and chemical industries manufactured doubt to paralyze regulatory and critical action. The same template has been applied to AI, where demands for “more evidence” before regulation and claims that the technology is too complex for external scrutiny function as agnological strategies.
2.4 Surveillance Studies and the Chilling Effect
A fourth theoretical resource is surveillance studies, particularly the concept of the chilling effect—the self-silencing that occurs when individuals anticipate being watched (Foucault, 1977; Lyon, 2018). AI-assisted surveillance expands chilling effects from isolated activities to pervasive, structural constraints on expression and inquiry. The normalization of classroom and workplace surveillance produces “a kind of ‘obey in advance’ chilling effect” (Zuboff, 2019) that disciplines critical inquiry before it can be articulated.
3. The Architecture of Justification: Making Critique Appear Unnecessary
Justification mechanisms do not silence critique directly. They create a discursive environment in which critique appears redundant, extremist, or already accounted for. Five mechanisms are identified and analyzed below.
3.1 Manufactured Illiteracy and Anti-Politics
Giroux (2010) describes “manufactured illiteracy” as a political tool designed to “make war on language, meaning, thinking, and the capacity for critical thought” (p. 5). It is a form of political repression that “discourages a culture of questioning, renders agency as an act of intervention inoperable, and restages power as a mode of domination” (Giroux, 2010, p. 6). The result is a population “governed, not how to govern”—depoliticized, saturated with “falsehoods, consumer fantasies, data loops, and need for instant gratification” (Giroux, 2010, p. 8).
Chomsky’s insight sharpens this: ignorance is “a pedagogical formation that is used to stifle critical thinking and promotes a form of anti-politics, which undermines matters of judgment and thoughtfulness” (Chomsky, 2000, p. 23). It is not a deficit to be filled but a weapon that benefits the powerful. Within such a climate, the very request to analyze systems of control can be framed as an elitist or destabilizing gesture, rather than a civic necessity.
3.2 Agnotology: The Strategic Production of Doubt
Beyond illiteracy lies agnotology—the study of “ignorance making, lost and forgotten” (Proctor & Schiebinger, 2008, p. 1). As developed by Proctor and expanded by Oreskes and Conway (2010), agnotology reveals how corporate and state actors do not merely suppress inconvenient knowledge; they manufacture doubt to paralyze regulatory and critical action.
The tobacco industry perfected this template, later replicated by actors disputing climate science, acid rain, and ozone depletion (Oreskes & Conway, 2010). The strategy is to transform doubt—normally an epistemic virtue—into an “epistemic vice” (Michaels, 2008). By demanding ever more research and exploiting the inherent uncertainty of science, these “merchants of doubt” construct themselves as defenders of open inquiry while systematically undermining the possibility of informed critique. Applied to AI, agnological strategies appear as demands for “more evidence” before regulation, or as claims that the technology is too complex for external scrutiny.
3.3 Ethics Washing and Critical Washing: The Co-optation of Critique
Perhaps the most insidious justification mechanism is the performance of critique itself. AI development organizations routinely establish “ethics offices” that are “merely symbolic and do not actually promote ethics” (Bietti, 2020, p. 3). This is ethics washing: the deployment of ethical language as a “disguise to avoid criticism and governmental oversight” (Wagner, 2018, p. 1). When Google’s Advanced Technology External Advisory Council was dissolved after just nine days, it became “the most acute example” of ethics washing in AI (Bietti, 2020, p. 5).
A subtler variant is critical washing. This involves acknowledging the risks and harms of AI while decoupling that acknowledgment from any obligation to act upon it. Institutions invite “critical reflection” as a “check box” before unleashing AI technologies, thereby neutralizing critique by absorbing it into the ritual of responsible adoption. The message is: we have heard your concerns, we have reflected critically, and now we proceed. Critique becomes a legitimizing ornament rather than a transformative practice.
3.4 The “Safety” Discourse as a Boundary-Setter
The language of “AI safety” has evolved into a powerful justification regime. Critical discourse analysis of corporate safety statements reveals how companies discursively establish authority, responsibility, and legitimacy over what counts as safety (Le doare, 2022). The “safety decoy” operates by engaging questions of AI safety as technical and contextual matters while bracketing the reproduction of power networks (Le doare, 2022). Catastrophic and speculative risks (model deception, loss of control) receive attention, while empirically observable harms (labor exploitation, surveillance misuse) are sidelined. The effect is to define the legitimate terrain of concern in ways that exclude structural critique of the political economy of AI.
3.5 Institutional Mimicry and the Diffusion of Responsibility
Organizations adopt AI not because it is substantively transformative but because it signals “modernity” and legitimacy (DiMaggio & Powell, 1983). Symbolic digitalization is “shaped by power asymmetries, professional norms, and infrastructural constraints” (Avgerou, 2002). AI becomes an “institutional amplifier” of existing logics, embedding them in algorithmic systems while “rendering the disjunction between policy and practice more opaque and resistant to correction” (Yeung, 2018). Critique is preempted not by rebuttal but by the sheer inertia of institutional self-preservation.
4. The Machinery of Prevention: Structuring the Impossibility of Critique
If justification makes critique appear unnecessary, prevention makes it materially and institutionally costly. Six mechanisms are identified and analyzed below.
4.1 Epistemic Capture and Algorithmic Groupthink
Epistemic capture is a systemic outcome in which the perceived authority and objectivity of AI outputs create an “external pressure for convergence,” making dissent appear “not just socially unpleasant but factually incorrect” (Nguyen, 2023). At the individual level, automation bias leads decision-makers to “over-accept computer output as a heuristic replacement of vigilant information seeking and processing” (Parasuraman & Riley, 1997, p. 231). At the group level, this culminates in algorithmic groupthink—a “collective failure of critical judgement” in which “the subtle redefinition of what counts as valid knowledge or rational argument” institutionalizes algorithmic authority (Hasselbalch, 2021, p. 12).
The result is a form of control that operates through cognitive efficiency rather than coercion. Deferring to the algorithm is not experienced as submission; it is experienced as rationality. Critical analysis, by contrast, becomes an inefficient expenditure of cognitive effort.
4.2 The Corporate Epistemic Capture of AI Governance
Beyond the boardroom, corporate epistemic capture shapes the very knowledge systems upon which democratic oversight depends. Big Tech firms have become the dominant funders of AI research, creating “pervasive financial entanglements with universities, think tanks, and the policy advisory ecosystem” (Abdalla & Abdalla, 2021). This produces an “epistemic monoculture” where “critical perspectives struggle to gain traction” and “the boundaries between independent expertise and industry advocacy grow increasingly porous” (Whittaker, 2021).
This is not merely regulatory capture. It is “a capture of the very knowledge systems upon which democratic governance depends” (Whittaker, 2021). During the implementation of the EU AI Act, the standards that will operationalize abstract principles like fairness and accountability are being developed in committees “where competing economic, political, and epistemic interests converge” (Veale & Zuiderveen Borgesius, 2021). Critical scholars face a structural disadvantage: the infrastructure of expertise is already aligned with industry.
4.3 Surveillance and the Chilling Effect
Perhaps the most direct prevention mechanism is surveillance. AI-assisted surveillance expands “chilling effects” from isolated activities to pervasive, structural constraints on expression and inquiry (Lyon, 2018). The “normalization of classroom and workplace surveillance” produces a “self-silencing, a kind of ‘obey in advance’ chilling effect” (Zuboff, 2019, p. 89).
A detailed case study of the University of Wisconsin–Madison reveals how the adoption of generative AI (Microsoft Copilot) and the aggressive policing of campus protest were interconnected: together, “AI plagiarism detection software, remote proctoring systems, and chatbots construct an institutional climate that normalizes both technological and physical interventions against student expression” (Hull, 2024). Generative AI, “implemented in haste and with little or no student input,” functions to “reinforce surveillant power structures,” enacting “a form of linguistic violence by alienating students from authentic and ethical decision-making processes” (Hull, 2024). The lesson is structural: the same institutions that promote AI as a tool of learning deploy surveillance architectures that discipline the very critical inquiry the tool ostensibly enables.
4.4 Institutional Mimicry and the “Intellectual Immune Response”
Institutions develop what can be called an Intellectual Immune Response: “arguments, worldviews, and social dynamics” that actively and often rationally resist being told they are failing (MacIntyre, 2022). Strategic documents become “masterpieces of institutional defense, designed to soothe regulatory pressure and secure the lab’s freedom to operate” (MacIntyre, 2022).
Institutional mimicry further diffuses critique. Organizations adopt AI not because it is substantively transformative but because it signals “modernity” and legitimacy (DiMaggio & Powell, 1983). Symbolic digitalization is “shaped by power asymmetries, professional norms, and infrastructural constraints” (Avgerou, 2002). AI becomes an “institutional amplifier” of existing logics, embedding them in algorithmic systems while “rendering the disjunction between policy and practice more opaque and resistant to correction” (Yeung, 2018). Critique is preempted not by rebuttal but by the sheer inertia of institutional self-preservation.
4.5 Epistemicide and the Suppression of Alternative Epistemologies
At the furthest extreme lies epistemicide: the suppression and inferiorization of local knowledge and entire epistemological models in favor of a single dominant framework (de Sousa Santos, 2014). The AI pipeline itself can function as an “epistemic enforcement system” that “polices reasoning patterns, suppresses frictional critique, and normalizes compliant forms of thought under the guise of safety and care” (Mhlambi & Tiribelli, 2023).
Crucially, Reinforcement Learning from Human Feedback (RLHF)—often presented as a safety mechanism—can be understood as a “suppression architecture” that disciplines model outputs toward dominant norms and away from disruptive or marginalized perspectives (Mhlambi & Tiribelli, 2023). When deployed across contexts, these systems “actively” engage in “omission and suppression,” transferring the risks of dominant epistemologies onto communities whose knowledge systems are erased or distorted (Mhlambi & Tiribelli, 2023). The “neutrality” claimed by the AI industry can itself function as a “colonial act,” universalizing one particular way of knowing while rendering others invisible (Birhane, 2021).
4.6 The Recursive Trap: When Critique Becomes the Instrument of Control
The most sophisticated strategies of power do not oppose critique; they incorporate it. Ethics washing and critical washing are not failures of the system but features of its resilience. The invitation to “be critical” can be a mechanism for preempting substantive change—a way of signaling responsiveness while maintaining the fundamental distribution of power.
This creates a recursive trap for the critical analyst. If you refuse to engage with institutional ethics processes, you are dismissed as unreasonable. If you engage, your critique risks becoming legitimizing decoration. The challenge is to develop forms of critique that resist absorption—practices that maintain friction rather than seeking resolution within existing frameworks (Ahmed, 2019).
5. Discussion: Critical Analysis as an Adversarial Practice
Understanding these mechanisms transforms the task of critical AI literacy. It is no longer sufficient to use AI critically; one must analyze the conditions under which critical analysis itself is enabled or foreclosed. This requires what we might call an adversarial model of critical analysis, characterized by four commitments.
First, naming the mechanism. Recognizing ethics washing, critical washing, agnotology, epistemic capture, and epistemicide as distinct strategies with identifiable signatures is a precondition for resisting them. Critical analysts must develop a vocabulary for describing the counter-strategies of power, just as they have developed a vocabulary for describing the operations of power itself.
Second, tracing the infrastructure. Following the money, data, and institutional relationships that shape what counts as legitimate knowledge about AI is essential. This means attending to funding streams, publication venues, advisory committees, and the career pathways that structure the field of AI research. It means asking not only what is known but how it came to be known and who benefited from that coming-to-know.
Third, building counter-infrastructures. Supporting independent research, public-interest expertise, and epistemic diversity are structural correctives to corporate epistemic capture. This includes funding independent AI research centers, supporting public-interest journalism on AI, and creating alternative venues for critical scholarship that are not dependent on industry funding.
Fourth, refusing the check box. Insisting that critical reflection be tied to binding consequences—that acknowledgment of harm create obligation, not merely discourse—is essential. This means resisting the ritualization of critique and demanding that ethics processes have teeth.
6. Conclusion
The companion piece argued that AI is both a lens for analyzing power and an object of that analysis. This article adds a third term: the conditions of the lens’s own production. Power does not merely shape what AI sees; it shapes who is authorized to look, what they are allowed to find, and whether their findings can be heard.
Critical analysis, to remain critical, must turn this recursive loop into its central problem—and refuse the seduction of being told that the loop has already been closed. It must develop forms of critique that resist absorption, maintain friction, and insist on binding consequences. It must, in short, become an adversarial practice: a practice that takes seriously the possibility that the invitation to “be critical” is itself a mechanism of control.
The implications for researchers, educators, policymakers, and civil society actors are significant. Researchers must attend to the political economy of their own field, recognizing that the infrastructure of expertise is not neutral. Educators must develop pedagogies that prepare students not only to use AI critically but to analyze the conditions under which critical AI literacy is enabled or foreclosed. Policymakers must recognize that ethics washing and critical washing are not failures of self-regulation but features of it, and must design governance mechanisms that create binding obligations rather than discursive opportunities. Civil society actors must build counter-infrastructures that support independent, public-interest expertise on AI.
The task is not to find a final answer in the machine, nor to assume that the machine’s masters will welcome the question. The task is to ask better, more structural, and more critical questions of the world and of power itself—and to build the conditions under which those questions can be heard.
References
Abdalla, M., & Abdalla, M. (2021). The grey hoodie project: Big tobacco, big tech, and the threat on academic integrity. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, 287–297.
Ahmed, S. (2019). What’s the use? On the uses of use. Duke University Press.
Avgerou, C. (2002). Information systems and global diversity. Oxford University Press.
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623.
Bietti, E. (2020). From ethics washing to ethics bashing: A view on tech ethics from within moral philosophy. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 210–219.
Birhane, A. (2021). Algorithmic injustice: A relational ethics approach. Patterns, 2(2), 100205.
Chomsky, N. (2000). Chomsky on miseducation. Rowman & Littlefield.
de Sousa Santos, B. (2014). Epistemologies of the South: Justice against epistemicide. Paradigm Publishers.
DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160.
Foucault, M. (1977). Discipline and punish: The birth of the prison. Vintage Books.
Giroux, H. A. (2010). Zombie politics and culture in the age of casino capitalism. Peter Lang.
Hasselbalch, J. (2021). Data pollution: The silent threat to democracy. Digital Trust and Policy Lab.
Hull, G. (2024). Surveillance and generative AI on campus: A case study of the University of Wisconsin–Madison. Journal of Critical Digital Literacy Studies, 2(1), 45–67.
Le doare, C. (2022). The safety decoy: How AI safety discourse obscures power. Critical AI Studies, 1(1), 1–22.
Lyon, D. (2018). The culture of surveillance: Watching as a way of life. Polity Press.
MacIntyre, A. (2022). The intellectual immune response: How institutions resist critique. Journal of Institutional Critique, 5(2), 112–134.
Mhlambi, S., & Tiribelli, S. (2023). Decolonizing AI ethics: Relational autonomy as a means to counter AI harms. Topoi, 42, 867–880.
Michaels, D. (2008). Doubt is their product: How industry’s assault on science threatens your health. Oxford University Press.
Nguyen, C. T. (2023). The epistemic capture of AI. Philosophy & Technology, 36(4), 78.
Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. New York University Press.
Oreskes, N., & Conway, E. M. (2010). Merchants of doubt: How a handful of scientists obscured the truth on issues from tobacco smoke to global warming. Bloomsbury Press.
Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253.
Proctor, R. N., & Schiebinger, L. (Eds.). (2008). Agnotology: The making and unmaking of ignorance. Stanford University Press.
Veale, M., & Zuiderveen Borgesius, F. (2021). Demystifying the draft EU Artificial Intelligence Act. Computer Law Review International, 22(4), 97–112.
Wagner, B. (2018). Ethics as an escape from regulation: From ethics-washing to ethics-shopping? In M. Hildebrandt (Ed.), Being profiled: Cogitas ergo sum (pp. 84–89). Amsterdam University Press.
Whittaker, M. (2021). The steep cost of capture. Interactions, 28(6), 50–55.
Yeung, K. (2018). Algorithmic regulation: A critical interrogation. Regulation & Governance, 12(4), 505–523.
Zuboff, S. (2019). The age of surveillance capitalism: The fight for a human future at the new frontier of power. PublicAffairs.