AI Arrived Before the Evidence: Scientists Map How Universities Can Change Under Uncertainty

Generative artificial intelligence has done something no educational technology has done before: it transformed classrooms around the world before anyone could study whether it should. That inversion, according to a new framework published in the International Journal of STEM Education, breaks the fundamental logic that has guided decades of science education reform, and the authors argue that universities need an entirely new playbook for leading institutional change when the evidence base simply does not exist yet.

The paper, written by David Perl-Nussbaum and Noah D. Finkelstein of the University of Colorado Boulder, builds on a distinction drawn by education researchers Justin Reich and Jeffrey Dukes between adoption technologies and arrival technologies. Adoption technologies, such as Peer Instruction, Tutorials in Introductory Physics, the Learning Assistant program, and interactive simulations like PhET, followed a familiar and well-studied sequence: researchers designed them with learning goals in mind, evaluated them rigorously, and only then made them available to faculty alongside evidence of their effectiveness. Generative AI followed the opposite path. ChatGPT entered schools through sudden, widespread, largely spontaneous use by students and teachers, promoted heavily by technology companies, before institutions had formed any response at all and before researchers could assess its risks and benefits for learning.

The authors acknowledge that earlier outside technologies did disrupt classrooms, but they note that calculators and mobile phones appeared on roughly ten-year iteration cycles, giving educators time to adapt between generations of hardware. Generative AI offers no such interval. Frontier models are gaining new capabilities on timescales of months rather than years, and the 2026 AI Index report finds that AI capabilities are still accelerating rather than plateauing. AI systems also exhibit what researchers call a jagged frontier: they handle some tasks well and fail at others that seem similar, and the boundary between the two shifts with every model release. A tool introduced in the first week of a semester may behave in meaningfully different ways by week fourteen.

This creates what the authors identify as the central dilemma of AI-era reform. Institutions cannot wait for best practices to emerge, because student learning is already being reshaped by AI use, campus policies, and individual faculty experimentation, with or without deliberate institutional involvement. Yet neither can they responsibly scale innovations that remain unvalidated. Banning generative AI is not tenable, and neither is uncritically embracing it. Existing models of institutional change, the authors argue, presuppose exactly the kind of stable, well-evidenced innovation that arrival technologies lack. Rogers’ classic Diffusion of Innovations theory, which underpins many STEM change efforts, explicitly describes adoption as an uncertainty reduction process, and it works only when uncertainty can actually be reduced through accumulated evidence.

To address this gap, the framework identifies six dimensions along which prior change models must be adapted, three concerning the tools at the center of reform and three concerning the people involved. The first tool dimension is the evidence base. Because direct evidence for effective AI-based pedagogy does not yet exist in recognizable form and may not for years, the authors call for humble inquiries grounded in intellectual humility: surfacing what faculty are already trying, documenting what appears to work and for whom, and sharing emerging cases without overclaiming their generalizability. Institutions should fund experimentation cycles and provide access to AI with guardrails, without making long-lasting pedagogical or financial commitments. The authors also flag an equity concern, noting that scaling unevidenced practices risks widening existing gaps, since learners with greater resources and digital literacy are best positioned to benefit from early-stage efforts.

The second dimension, rate of change, leads to one of the framework’s most consequential design implications: reform should be organized around pedagogical approaches rather than specific tools. Where past initiatives like the Science Education Initiative could reasonably build repositories of stable course materials, AI-driven change should build repositories of approaches and example cases that survive the replacement of any particular model or platform. These approaches should draw on the durable evidence base from interactive engagement research, including what is known about student engagement, authentic disciplinary practice, and epistemology, none of which becomes obsolete when a new model is released. This reframing also shifts institutional investment away from tool procurement, such as particular grading systems or tutoring agents, toward sustained support for faculty inquiry and iteration. Course policies, the authors add, should be developed through ongoing dialogue between instructors and students and revisited as tools evolve, though they caution that iteration cannot excuse institutions from coherent principles around ethics and academic integrity.

The third tool dimension concerns scope. Unlike clickers or tutorials, which were purpose-built by educators with learning theory embedded in their design, generative AI is a general-purpose technology developed outside education and appropriated for classroom use. Some of its most common applications, producing finished work with minimal effort, stand in direct tension with the purposes of learning. AI is also simultaneously a mediator and an object of learning: students must learn to use it productively, and the disciplinary practices being taught are themselves being reshaped by it. The authors argue that faculty therefore need institutional support to educate students beyond disciplinary content, addressing ethics, agency, and the role of AI in students’ professional futures, and that flexible general tools are a more durable strategy than dedicated platforms that carry obsolescence risk.

On the human side of the framework, the authors document a profound shift in faculty agency. In past reforms, faculty opted in voluntarily, motivated largely by intrinsic interest in student learning. With generative AI, the decision to engage has effectively been made for them: their courses have already been transformed by unsupervised student use, whether faculty are aware of it or not. The disruption is also highly uneven, hitting large introductory writing and computational courses far harder than laboratory or graduate courses, so a framework that treats all faculty as equally affected risks irrelevance for some and under-resourcing for others. The authors recommend mapping disruption through structured interviews and investing in support structures analogous to the Science Education Initiative’s discipline-based teaching fellows, but with a crucial redefinition of the role. Because expertise in AI-based educational practice is largely absent, change agents cannot serve as brokers of known best practices; they must instead act as facilitators of collective inquiry. The authors specifically caution against overvaluing AI tool expertise, warning that change agents drawn from industry or vendor contexts may be implicitly oriented toward promoting adoption rather than critically evaluating it, and arguing that pedagogical and disciplinary expertise is more durable.

The final dimension repositions students from the endpoint of reform to genuine participants in shaping it. Survey data consistently shows that students lead in AI use, are actively seeking guidance on responsible use, and often cite respect for their instructors’ policies as a reason for not using AI to produce finished work. The authors found a structured early-semester survey of students’ AI use to be a valuable entry point for collaborative discussions about norms. They illustrate the framework through a case study of a six-session faculty workshop series in the CU Boulder physics department in Spring 2026, involving nineteen faculty across the series. In one session, survey data from roughly 350 students revealed that checking the correctness of solutions was among their most frequent AI uses, prompting faculty to test actual homework solutions in chatbots and examine which prompting strategies encouraged reflection rather than shortcuts. The authors are careful to note the framework’s limitations: it is theoretical and not yet empirically tested, it draws primarily from U.S.-based change initiatives, and the AI landscape itself continues to evolve. They present it not as exhaustive or final, but as a call to action, urging institutions to move forward with humility and intention into an uncertain landscape that has already arrived.

Subject of Research: A framework for adapting STEM higher education institutional change models to generative AI

Article Title: A framework for institutional change in the age of AI

Article References: Perl-Nussbaum, D., & Finkelstein, N. D. (2026). A framework for institutional change in the age of AI. International Journal of STEM Education, 13(1), Article 60. https://doi.org/10.1186/s40594-026-00649-4

Image Credits: AI Generated

DOI: 10.1186/s40594-026-00649-4

Keywords: generative AI, STEM education, institutional change, higher education, arrival technology, interactive engagement, faculty development, change agents, student partnership, educational reform, intellectual humility, pedagogy