I’ve spent the last couple of years arguing that pedagogy, not prohibition, is the answer to AI in education. I still believe that. But every now and then a paper comes along that forces me to zoom out and ask whether the conversation I’m having is big enough. Hartzog and Silbey’s essay “How AI Destroys Institutions” (2025) is that kind of paper. Two law professors at Boston University make the case that AI’s deepest threat runs straight through the institutions that hold democratic life together: the rule of law, higher education, the free press, and civic governance itself.
The argument is sweeping, and parts of it are as convincing as they are uncomfortable. All of it is worth reading carefully, especially for educators who tend to think about AI as a classroom problem when it is, increasingly, a structural one.
Three Ways AI Threatens Institutions
Hartzog and Silbey organize their argument around what they call the “destructive affordances” of AI. These are are baked into how AI systems function.
AI undermines expertise. That’s the first affordance. People stop doing the hard cognitive work once AI takes over, and the skills institutions depend on gradually erode. Hartzog and Silbey point to the back-end labor this creates: someone still has to catch the hallucinations, verify the outputs, and clean up the errors. That labor is invisible, and it produces no new knowledge. This tracks closely with the research I’ve covered on cognitive surrender (Shaw and Nave, 2026), where the concept describes how users progressively hand over reasoning to AI systems without recognizing the cost. Hartzog and Silbey are making essentially the same argument, but at institutional scale.
Their second affordance is the short-circuiting of institutional decision-making. Institutions run on rules that can be debated, revised, and challenged. AI systems hide those rules inside black-box models, making decisions opaque and difficult to contest. The authors argue that this also outsources moral choices to machines with no moral capacity, which is a strong claim but one that resonates when you think about AI being used for bail decisions, hiring, or insurance approvals.
Then there’s isolation, the third affordance. Institutions require human interaction, disagreement, friction, even awkwardness, to function. AI replaces all of that with smooth, sycophantic interfaces that feel like connection but aren’t. Hartzog and Silbey cite research showing that people who discover their colleagues used AI start viewing them as less creative (54%), less capable (50%), less reliable (49%), and less intelligent (37%). I covered similar findings through Claessens, Veitch and Everett (2026), whose work on social perceptions of AI use pointed to the same pattern. The trust damage is real, and it compounds over time.
The Higher Education Problem
The higher education section of the paper is the one that’ll land hardest with most readers of this blog, and it’s the section I find most compelling. Hartzog and Silbey make a four-part case that AI is corrosive to the university as an institution.
First, AI offloads the very cognitive tasks that make learning happen. This is ground I’ve covered extensively. Fan et al. (2025) showed that students using ChatGPT produced better essays but showed no additional learning gains compared to students who wrote without it. The AI improved the product without improving the process. Hartzog and Silbey are making a version of that same argument but connecting it to institutional legitimacy: if universities can’t credibly claim they’re developing expertise, they lose their justification for existing.
Second, AI produces homogenized, mediocre content that depresses intellectual risk-taking. The authors quote a striking observation: “Higher education is about learning how to learn as much as it is about learning specific content and skills. We should not be complacent about AI’s effect on attitudes to, and capacities for, knowledge acquisition, and on the willingness to take intellectual risks” (p. 27).
I find this compelling because it names something specific. AI doesn’t just make students lazier. It changes what feels worth attempting. If every student paper reads the same and every research question gets funneled toward what AI can answer, the culture of intellectual risk that universities are supposed to protect starts to dissolve.
Third, AI shifts research questions from qualitative mysteries to quantifiable puzzles. The authors argue this marginalizes the humanities and social sciences by implying that all knowledge is reducible to numbers. I’d complicate that claim, because quantitative research has its own rigor and plenty of qualitative researchers are finding productive ways to use AI (I covered Anis and French’s 2023 work on AI in qualitative research). The real risk is the assumption that whatever AI handles well is the only thing worth studying.
Fourth, AI erodes student-faculty trust. When generative AI replaces professors, students lose faith in what they’re learning and in the institution that charges them tuition to learn it. Hartzog and Silbey reference the Northeastern University case, where professors reportedly used ChatGPT and students objected. That erosion of trust feeds directly into the broader political attacks on higher education the authors describe, giving ammunition to those who already want to dismantle universities.
I agree with Hartzog and Silbey’s core diagnosis. The three affordances are real. The institutional analysis is sharp. But the paper reads, at points, like AI is a single force acting uniformly on all institutions at once. It’s not.
The essay treats AI as inherently destructive, full stop. “The affordances of AI systems are like a cancer in our struggling democracies,” the authors write (p. 40). That’s a powerful line, but it’s also totalizing in a way that leaves no room for the educators, journalists, lawyers, and civic organizers who are already figuring out how to use AI without surrendering their expertise or their institutional responsibilities.
I’ve watched teachers use AI to build more responsive assessment systems, researchers use it to accelerate literature reviews without replacing their own analysis, and students use it to draft and then critically revise their own thinking. Those uses don’t show up in this paper. The authors acknowledge that AI is “just a refracted mirror of humanity” (p. 39) but don’t follow that thread far enough. If AI reflects us, then how we deploy it matters enormously. And that’s a pedagogical question, not just a legal one.
The paper also doesn’t account for what Kosmyna et al. (2025) found at MIT: students who thought independently before engaging with AI showed stronger neural engagement than those who started with AI. The cognitive damage depends on the sequence, the task design, and the pedagogical intent. That nuance is missing here.
What Educators Should Take from This
The value of Hartzog and Silbey’s essay for educators is in the frame, not the prescription. Most of us think about AI at the level of the assignment, the lesson, the semester. This paper forces a harder question: what happens to the institution when AI becomes the default? If expertise atrophies across an entire faculty, if students stop trusting the process, if research narrows to what AI can easily handle, the university doesn’t collapse overnight. It just becomes something different. Something smaller.
Hartzog and Silbey warn that “the more AI systems are deployed, the less durable and adaptable institutions become. As a result, the institutions will become increasingly ossified and delegitimized” (p. 21). That’s worth taking seriously. But institutions have fought off ossification before. The antidote has always been intentional design, transparency about what the institution is for, and a willingness to adapt.
For educators, that starts with being explicit about why certain learning activities exist and what role AI should or should not play in them. Courses need to be built so the cognitive work stays with the student, so AI remains a tool and never a substitute for thinking. And none of that works if we ignore the institutional trust that makes all of it possible.
References
- Anis, S., & French, J. A. (2023). Efficient, explicatory, and equitable: Why qualitative researchers should embrace AI, but cautiously. Business & Society, 62(6), 1139–1144. https://doi.org/10.1177/00076503231163286
- Claessens, S., Veitch, P., & Everett, J. A. C. (2026). Negative perceptions of outsourcing to artificial intelligence. Computers in Human Behavior, 177, 108894. https://doi.org/10.1016/j.chb.2025.108894
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
- Hartzog, W., & Silbey, J. (2025). How AI destroys institutions [Draft]. Boston University School of Law. https://scholarship.law.bu.edu/faculty_scholarship/4179