CU Boulder researchers warn higher education AI change models are already out of date

A new paper from David Perl-Nussbaum and Noah D. Finkelstein argues that universities cannot treat generative AI like previous classroom technologies because students are already using it before evidence, policy, and practice have caught up.

Noah Finkelstein, Distinguished Professor at the University of Colorado Boulder, has co-authored a paper on how higher education institutions can respond to AI adoption.

University of Colorado Boulder researchers David Perl-Nussbaum and Noah D. Finkelstein have published a new framework for how higher education institutions should respond to generative AI, arguing that existing models for educational change no longer fit the speed and uncertainty of AI adoption.

The paper, titled "A Framework for institutional change in the age of AI", focuses on STEM higher education but raises a wider issue for universities, EdTech providers, faculty development teams, and AI strategy leads: generative AI is not being introduced through the usual route of testing, evidence-building, and planned rollout. It is already in classrooms, coursework, assessment, and student workflows.

Finkelstein, Distinguished Professor at the University of Colorado Boulder, shared the paper on LinkedIn and said current scholarly models of change were built around "proven technologies that we choose to adopt." AI, he wrote, has changed that pattern because "AI educational practices are arriving to our classes before they are proven and often without our choice."

AI is arriving before the evidence base

The paper draws a line between older forms of education technology adoption and the arrival of generative AI.

Traditional STEM education reforms, including interactive engagement approaches such as Peer Instruction, Tutorials in Introductory Physics, Learning Assistant programs, and simulations, were generally developed, evaluated, and then scaled. In those cases, universities could point to evidence before asking faculty to change practice.

Perl-Nussbaum and Finkelstein argue that generative AI does not follow that sequence. The paper describes AI as an "arrival technology", meaning a technology that enters learning environments before institutions fully understand its impact or have enough evidence to decide how it should be used.

That distinction is the core of the paper. Universities cannot wait years for a settled evidence base because students are already using AI tools. But they also cannot responsibly scale AI-based teaching practices as if the benefits, risks, and learning effects are already proven.

The authors reject both simple bans and uncritical adoption. Their argument is that higher education needs a change model for uncertainty, not another round of rushed tool procurement or policy documents that assume the classroom is still controllable from the top down.

Six dimensions for institutional change

The proposed framework identifies six areas where existing institutional change models need to be reconsidered for the AI era.

Three relate to the tools themselves: the evidence base, the rate of change, and the scope of use. The paper argues that AI tools are different from many previous education technologies because the evidence base is still forming, the tools change quickly, and their use extends beyond a single course, discipline, or platform.

That creates a practical problem for universities. A generative AI tool can affect homework, assessment, feedback, academic integrity, student writing, coding, research, tutoring, and career preparation at the same time. It is not a single classroom intervention that can be neatly piloted in one module and then scaled unchanged.

The other three dimensions focus on people: faculty agency, the role of change agents, and the role of students.

Faculty agency is different because many instructors did not choose AI adoption. Their courses have been affected because students, vendors, campus policies, and individual colleagues are already using or responding to generative AI. The paper says some faculty may not yet recognize how far that disruption has gone, while others are already redesigning assignments and classroom expectations around it.

Change agents also need a different role. In previous education reform work, change agents often helped faculty adopt established practices. In the AI context, the authors argue that their role should shift toward facilitating shared inquiry, helping departments compare experiments, discuss uncertainties, and learn from local evidence.

Students are the final piece of the framework. The paper argues that they should not be treated only as recipients of reform, because they are already shaping AI practice through their own use of tools. The authors call for students to be involved as partners in institutional change, including conversations about what they are doing with AI, why they are doing it, and what support they need.

CU Boulder case study puts the framework into practice

The paper also includes a case study from a University of Colorado Boulder Physics Department workshop series designed to support instructors responding to AI.

The six-session workshop series covered AI policy development, student dialogue around AI use, and specific ways AI might appear in coursework, including checking homework or generating physics simulations. The structure gives the framework a practical footing rather than leaving it as a purely theoretical model.

The paper’s design implications are deliberately cautious. It calls for "humble inquiries" into local AI use rather than premature claims about best practice. It also suggests that institutions should organize reform around teaching approaches and learning goals, not specific AI tools that may change or disappear.

The paper does not argue that universities should wait until the sector has perfect evidence. It argues that institutions should document what is happening, involve faculty and students, and avoid pretending that one policy or platform can settle the issue.

Perl-Nussbaum and Finkelstein also acknowledge limits to the framework. It has not yet been empirically tested against alternative approaches, and it is drawn mainly from U.S.-based change initiatives. The authors also note that the AI landscape is still moving quickly, which means the most important dimensions may shift as the technology and classroom practice evolve.

The paper is now available as an arXiv preprint. Its central question is one many universities are already facing in real time: how to lead AI change when the tools are moving faster than the evidence, and students are not waiting for permission to use them.