Universities across the United States are moving from guidance to binding policy on generative AI, driven by a recognition that students have already adopted the tools and institutional silence is itself a decision. A new multi-institution position paper argues that AI has not changed how expertise is built, while a separate preprint finds that the negative effects of AI on learning may intensify among students who are more confident in their ability to evaluate its output.
The policy shifts came from multiple directions this week. Florida's Board of Governors issued a notice of intent to require public universities to adopt AI policies addressing faculty disclosure, student disclosure, and permitted uses. Only 8 of the state's 40 public institutions had issued any AI guidance, with just two adopting formal policies. The University of Alaska regents reviewed a draft systemwide policy organized around preserving human expertise, protecting data security, and expanding access. The University of Texas at Austin began requiring an AI statement in every syllabus, pushing the decision to the course level where disciplinary differences can be respected.
A widely-covered MIT committee report gave faculty shared language for a phenomenon many had noticed but struggled to name: "cognitive surrender." The phrase describes students reaching for chatbots at the first sign of difficulty, producing what the report called "the illusion of learning." Coverage catalogued a spectrum of institutional responses, from Ohio State integrating AI across majors to the University of Chicago and UC Berkeley Law restricting AI in certain courses.
The expertise question
A position paper from 26 statistics, data science, and computer science educators at 11 liberal arts colleges and Carnegie Mellon University makes a compact claim: generative AI amplifies existing expertise but hinders novice development. Because the tools offer "a largely effortless way to accomplish tasks," they let novices bypass the deliberate practice that builds the automated mental models expertise depends on. The authors write that novices lack the background knowledge to evaluate AI output, lack the triage skills experts use to decide what to check, and are subject to a metacognitive failure in which performance declines while the feeling of learning rises.
"There is simply no substitute for good old-fashioned practice," the paper states. Its recommendations include in-person, AI-free assessments aligned to stated learning goals and active re-encouragement of learning communities as a counterweight to solitary AI interaction displacing office hours and study groups.
A separate preprint from a single institution, surveying 118 students across 12 AI-related courses, found that negative associations with AI use intensified as students' self-reported confidence in judging AI output increased. The finding is correlational and hypothesis-generating rather than causal, but it raises a research question worth pursuing: whether knowing more about a tool's limitations changes how heavily a student leans on it.
Verification at both ends of the ladder
A journal editor's account made the stakes concrete. Nihar B. Shah, an editor-in-chief at Transactions on Machine Learning Research, offered ten authors of desk-rejected submissions a meeting before the decision became final. Seven joined roughly 30-minute video calls. Three could not answer basic questions about their own papers. Two showed almost no substantive understanding of the contents. Shah concluded that when authors cannot answer basic questions about their papers, "it is difficult to see how they could have verified the paper's contents."
At the opposite end of the expertise spectrum, OpenAI announced that an internal model produced a computer-checkable proof for the Navier-Stokes equations, one of the Clay Mathematics Institute's seven Millennium Prize problems. The proof was written in Lean, a language in which a computer can verify every step. Specialists noted the version settled is the one the prize is officially written around, not the version mathematicians most want answered, and the prize has not been awarded. But the specialists agreed the key idea came from people: the strategy the proof relies on is credited to Diego Córdoba and Luis Martínez-Zoroa.
Read alongside the expertise paper, the proof poses a governance question. Nearly every serious proposal for AI oversight - peer review, procurement review, model evaluation - assumes a supply of humans whose judgment is at or above the level of the thing being judged. If AI shortens the path to producing expert-looking output while lengthening the path to holding expert judgment, institutions are drawing down a stock of expertise faster than they are replacing it.
Why this matters for educators and writers
The week's developments converge on a practical question: what can a policy realistically observe and enforce? AI use is difficult to observe passively, so a disclosure requirement is not a monitoring instrument but a self-report instrument. Self-reports respond to how people read the purpose behind them. If disclosure is understood mainly as a way to catch and regulate use, it tends to suppress the very reports it depends on. If it is framed around a shared aim - that institutions are collectively trying to understand these systems well enough to oversee them - candid accounts of real use become more likely.
For faculty, the most actionable takeaway is the shift from detection to verification. Rather than relying on AI detectors, which are unreliable and adversarial, the "verification-first" approach designs assessments whose process is observable: drafts, checkpoints, oral components. The goal is not to catch AI after the fact but to build instruments that capture what a student actually understands. The person most affected by the illusion of learning is not the instructor who receives a misleading essay but the student who receives a misleading estimate of themselves.
Your membership also unlocks:
