Generative AI in Education and the Erosion of Critical Democratic Citizenship

Synthesis: Muñiz and Castro Samayoa (2026) argue that the rapid uptake of Generative AI in education threatens democratic citizenship, not because students cheat but because the tools remove the practice that Critical Thinking requires. Drawing on Dewey's account of thinking as an "active, persistent and careful consideration" that improves only through exercise, and on Burbules's claim that democracies become "vulnerable to demagoguery and outright deception" when citizens stop practicing critical evaluation, they name three mechanisms of erosion: the disappearance of productive friction from learning, automation bias, and an efficiency frame that treats the point of education as task completion. Their remedy is a pair of literacies and a design commitment — critical AI and media literacies, and learning practices that prioritize process over product — so that the student, rather than a chatbot, remains "the chief epistemic authority and creative agent." The essay reports no empirical data.

Key Findings

1. Critical thinking is a practice, and generative AI removes the practice. Following Dewey and Haber, the authors insist that becoming a critical thinker means "putting that knowledge to use regularly." Outsourcing inquiry, analysis, evaluation, and synthesis to a model therefore does not merely save time; it stops the exercise the capacity depends on, leaving what they call a "faulty, weakened foundation" for later skills.

2. Three mechanisms erode it, and only one is about accuracy. First, generative AI "diminishes the valuable friction in the learning process" — the frustration, doubt, slowness, and indeterminacy that Allison and DeRewal locate inside knowledge production, and that desirable difficulties research treats as the mechanism of durable learning. Second, many users believe "in the inherent impartiality of technology," the automation bias Gerlich documents. Third, the technology arrives with harms students are rarely asked to weigh: its environmental cost and its reliance on exploited labor.

3. The output is homogeneous and poorly recalled, and the summaries encourage skipping the source. The authors cite Kosmyna and colleagues' findings on work produced with generative AI, and DeRewal's probe of a research assistant tool, to argue that autogenerated summaries are unreliable and that their convenience teaches students to bypass careful reading of original sources — undercutting the purpose of student research.

4. Efficiency talk carries a false premise: that the point of schoolwork is to finish quickly. This is Biesta's "learnification" of education taken further, where process is subordinated to product. The authors concede that students have always sought shortcuts, from ghostwriters to dissertation mills, and note that before computers many men outsourced writing to wives and amanuenses. What distinguishes this moment is the scale plus the moral framing that speed is the goal.

5. The sequencing claim is the paper's most usable pedagogical argument. Elementary students learn arithmetic before calculators; college students learn citation conventions before a reference manager automates them. Generative chatbots reverse that order: they reduce the time spent researching, reading, and forming one's own ideas at exactly the stage when those practices are being built.

6. The political stakes are argued, not measured. Authoritarian governments, the authors write, flourish when people cannot evaluate truth claims, and regimes have historically suppressed literacy to that end. They marshal evidence from elsewhere — a correlation between generative AI use and weaker critical thinking, strongest among participants aged 17–25; cognitive debt among learners writing with a model; consultants "falling asleep at the wheel" with high-quality automation; and workers with higher confidence in the tools exerting less critical thinking effort — alongside surveillance and manipulation findings from Feldstein and Coeckelbergh's account of technofascism.

7. The remedy is two literacies and a change in how learning is judged. Media and information literacy, a critical AI literacy that teaches how these models work and why they fail, and assessment that rewards visible process: probe the tools as a researcher rather than use them as a customer, keep the student as the author of the thinking, and treat the technology as something to be overseen rather than trusted.

Why critical thinking is treated as democratic infrastructure

The essay's foundation is a definitional chain: Dewey's "active, persistent and careful consideration," Haber's emphasis on autonomous systematic thinking, and Burbules's pairing of cognitive skills with values and commitments. Democracy enters not as decoration but as the reason the practice must be maintained: participation requires both skepticism and openness, and a citizen who cannot test competing claims is available for manipulation. The authors add an empirical-feeling claim of their own — that generative AI cannot distinguish truth from falsehood — and connect it to the disinformation literature, where false or unverifiable content generates epistemic cynicism.

Their honest concession matters for how the page should be read. They do not argue for eliminating educational technology, they allow that better tools co-designed with educators may exist, and they acknowledge the enthusiasts' case. Their claim is narrower: at this moment, with these systems, the dominant educational use pattern trades away the practice democratic citizenship needs.

How the erosion actually works

Three causal stories are braided together, and separating them is useful because each implies a different remedy. The friction story says difficulty is productive and its removal is a cost, not a saving; the remedy is designing for effort. The automation story says people over-trust automated output, including fluent output; the remedy is teaching the tool's failure modes explicitly. The framing story says the vocabulary of efficiency redefines what education is for; the remedy is assessment and institutional messaging that value the process.

A fourth story sits underneath: the capacities Aoun calls robot-proof — critical thinking, systems thinking, and cultural agility — are exactly the ones that practice builds and delegation does not. The authors treat that as a reason developmentally appropriate sequencing matters, not as a reason to ban tools.

What this means for practice

• Introduce the tool after the skill, not before it. The calculator and the reference manager are the authors' own analogies; use them as a design rule when deciding at what point in a course a model enters.
• Protect friction deliberately. If an assignment's value lies in the reading, the draft, or the struggle with an ambiguous problem, say so and design the sequence so the model cannot stand in for that step.
• Teach the mechanism, not just the policy. Students should be able to explain how large language models generate text, why they are prone to error, and what it means that they are "probabilistic mimics" rather than retrieval systems.
• Have students probe the tool as researchers. The Design Justice Labs recommendation the essay cites — assessing models' strengths, weaknesses, and limitations through experiments of their own design — turns a use policy into a lesson.
• Assess process artifacts alongside products, and treat mandatory adoption as a policy choice with costs. The process-oriented assessment literature in this knowledge base carries the same argument from the measurement side; this essay supplies the civic rationale for it. When institutions instead require teachers to use these tools, they remove the latitude to weigh climate and labor concerns, and they foreclose the sequencing judgment the authors argue for.

Limitations

• The essay is an argument, not a study: no sample, no instrument, no measured outcome. Its empirical footing is entirely secondhand, and several of the studies it leans on are not appraised for design quality.
• The strongest evidence cited is correlational (use of generative AI associated with weaker critical thinking), which cannot distinguish atrophy from prior disposition or from the selection of students who already prefer shortcuts.
• "Generative AI" is treated as one thing. Differences between models, between prompted and unprompted use, and between a scaffolded classroom task and an open chatbot are largely left aside.
• The democratic-erosion claim is normative and unmeasured: no data in the essay links classroom use of these tools to civic outcomes, turnout, or susceptibility to disinformation. That is the same gap this knowledge base records for civic and political education generally.
• Competing evidence is acknowledged in a sentence rather than weighed, and the possibility that guided use improves critical thinking is not tested against the authors' position.

Citation

Muñiz, R., & Castro Samayoa, A. (2026). Generative AI in Education and the Erosion of Critical Democratic Citizenship. Critical AI, 4(2).