AI and learning: experts flag memory and writing as at risk

A University of Minnesota expert panel concluded that generative AI can disrupt the foundations of memory: encoding, retrieval and information acquisition. The same experts agreed that all 4 motivational processes they rated (agency, engagement, motivation and persistence) are vulnerable. Writing was the only subject-area process that reached consensus, and they rated it high risk.

The report, "Ways generative AI can support and threaten learning in K-20 U.S. education," came out of the College of Education and Human Development in October 2026. Its lead coordinating authors are Panayiota Kendeou, Jeffrey A. Greene and Nia Nixon. They used a Delphi process, preregistered on AsPredicted.org, in which a panel rates and re-rates until positions settle. 30 experts took part in phase one, and 29 rated in phase two. Of 65 learning processes, the panel reached consensus on 42.

What the experts flagged

The pattern holds well beyond memory. On higher-order thinking, 10 of 18 items reached consensus on vulnerability only, with no matching consensus on support. Critical thinking and creativity are on that list. On social learning, 5 collaboration items landed the same way. The processes flagged most were close to the same across elementary, secondary and higher education.

The authors chose their words with care. Their warning begins, "Casual, careless use of general-purpose GenAI can threaten the very learning processes..." I read those 2 adjectives as the heart of the report. Casual and careless describe how a tool gets used, and how a tool gets used is a design decision that sits with the institution.

The report also says where the tool fits: "Within the acquire - practice - apply cycle, GenAI works best in the middle." That sentence is as close to a course-design rule as a consensus report gets.

The limits of the evidence

Two limits belong up front for any provost reading this. The report covers K-20 education as a whole, and none of its findings are specific to college. Its sharpest concern sits in the elementary grades. In the authors' words, "GenAI's promise and pitfalls are most precarious in the elementary grades." Elizabeth Shockman's coverage for MPR News centered on elementary students.

The second limit is evidence. Most of the evidence-strength ratings in the report were low. These are expert consensus judgments, reached through structured rounds of rating, and experimental results would carry a different kind of weight. I treat the report as a risk map for higher education. A risk map tells you where to look first, and this one points at processes every college depends on.

The four-year bet runs on these processes

A university asks a student to spend years on a curriculum. I call that the four-year bet, and I wrote about it when AI majors nearly doubled and again in a piece on the agentic AI conversation at ASU. The question the bet raises is whether the program is built to be rewritten.

This report sharpens that question. Every program in a catalog depends on memory, critical thinking, collaboration and persistence. Those are the processes the experts flagged. A curriculum that lets AI absorb that work in year 1 can graduate a student in year 4 who holds the credential and lacks the habits underneath it. That is the curriculum half-life problem I explore in my forthcoming book, The Great Relearning: what a student learns has to outlast the tools that change around it.

I spent a decade establishing Cisco innovation centers across 14 countries, and I came away convinced that anything built to last has to be built to be revised. A curriculum follows the same rule. The programs that redesign their courses around these findings now will be the ones whose graduates can keep relearning later.

Where higher-ed leaders can start

The report makes 11 recommendations. 2 of them map directly onto decisions a dean or provost controls.

The first is Competency-Gated GenAI Use. As I read the name, AI use opens up after a student has shown the underlying skill. Writing is the natural place to begin, since it is the one subject-area process the experts rated high risk. A first-year writing course could ask students to show they can draft and revise an argument on their own before AI enters the assignment.

The second is Authentic Assessment Redesign. If a general-purpose tool can produce a take-home essay from start to finish, that essay measures something other than what the course claims to teach. Redesigning assessment is slow faculty work, and it is the work that decides whether the four-year bet pays off for the student.

Kendeou's advice in her interview with MPR News fits alongside both.

"First, give the student the opportunity to work independently so their ideas and their agency is preserved." Panayiota Kendeou, lead coordinating author, to MPR News

That sequence lines up with the report's line about the middle of the cycle. I read it this way: students acquire on their own, practice with AI as a partner, and apply on their own again. Agency was one of the 4 motivational processes the experts agreed is vulnerable, and independent work first is how a course protects it. Any program can make that choice this term.

Kendeou put the stakes plainly to MPR News: "With AI in education, we cannot afford to get it wrong." Here is the question I would bring to your next academic council meeting: which course in your catalog would you gate first, and what would a student have to show before the AI turns on? If you want to work through the report's 11 recommendations with your deans and faculty leaders, I'm easy to find, and the rest of my AI in education work is a good place to start.

Here is what makes Alex a credible voice on this topic: Alex Goryachev is Innovator-in-Residence at Tulane University, the WSJ-bestselling author of Fearless Innovation, and shaped a $1.1 billion innovation portfolio at Cisco. He works with university leaders on building programs that can be revised as fast as the field moves.

Talk with Alex about AI and your curriculum →