Universities Don’t Have an AI Problem. They Have a Human-Capacity Problem

Universities are scrambling to respond to artificial intelligence. A recent New York Times article An M.I.T. Report Warns A.I. Is Causing ‘Cognitive Surrender.’ Universities Are in a Bind captures just how unsettled the response has become. Some universities are embracing it; others are restricting it. Some are hot and cold with AI like they are three dates in and still refusing to define the relationship. On a classroom level, some professors require students to use it; others prohibit it. Some avoid it altogether, hoping the latest fad will blow over. Institutions are rewriting academic-integrity policies, redesigning assessments, experimenting with AI-proof assignments, investing in detection tools, and teaching students how to prompt.

All those responses may have a place. But they are circling a more fundamental problem.

Universities are trying to govern the presence of AI when they need to develop the judgment of the humans using it.

That distinction is becoming increasingly difficult to ignore.

In August, an MIT committee examining AI in teaching and learning warned of “cognitive surrender”: students reaching for AI at the first hint of struggle and mistaking access to an answer for learning. Emerging research gives universities reason to take that concern seriously. An MIT Media Lab study found lower cognitive engagement, weaker recall, and less ownership among participants using an LLM for essay writing. A much larger 2026 study of nearly 27,000 secondary students found something even more provocative: AI use raised homework scores while exam performance declined. The learning losses were concentrated among students whose behavior was consistent with outsourcing their homework to AI.

The lesson is not that AI makes students less intelligent. It is that successful completion and learning are not synonymous.

And more importantly, cognitive offloading is not cognitive surrender.

COGNITIVE OFFLOADING IS NOT COGNITIVE SURRENDER

Human beings have always used tools to offload cognitive work. We use calculators rather than doing every calculation in our heads. We write things down rather than relying entirely on memory. We use navigation systems for route guidance and Michelin stars to select restaurants. Search engines, reference books, spellcheckers, maps, databases, other people… the list of tools humans use to offload cognitive work is long. Humans have been outsourcing pieces of our thinking for centuries. Apparently, outsourcing to Michelin is sophisticated, but outsourcing to ChatGPT is a crisis in human cognition.

Offloading cognition is not inherently a problem. The question is what gets offloaded, when, and why.

If AI takes on work that no longer requires an individual’s direct attention and frees them to spend more time analyzing, questioning, creating, or solving a more complex problem, that can be augmentation. But we should not confuse work that can eventually be offloaded with work that never needed to be learned.

A first grader may eventually use a calculator. But first, we ask them to do the computation themselves, not because calculators are bad, but because developing number sense helps them understand what the calculator is doing and recognize when its answer does not make sense. It also gives them the capacity to check the output, because when a calculator gets something wrong, we do not hold the calculator accountable.

Accountability belongs to the human using the tool, a concept parts of the AI industry have occasionally approached with all the enthusiasm of a shrimp salad that’s been sitting in the sun for hours at a neighborhood potluck.

But accountability is only part of what the student gained by doing the work in the first place.

The same is true across learning. Struggling to organize an argument may be how a student discovers what they actually think. Wrestling with an unfamiliar problem may develop persistence and problem-solving ability. Comparing conflicting evidence may develop judgment. Revising a bad first draft may reveal weaknesses in reasoning that were invisible when the idea existed only in the student’s head.

The question, then, is not whether AI can remove friction. Of course it can. The question is whether the friction we are removing leaves the learner more or less capable.

Education still has a fundamental responsibility to cultivate a more capable learner.

That means the educational question cannot stop at, “Can AI do this task?” We also must teach learners to ask: “What is appropriate to offload to AI, and what must I do myself because doing it is how I become capable?” That is a judgment students will need long after graduation.

THIS IS BIGGER THAN CHEATING

Much of education’s response to generative AI began, understandably, with academic integrity. Did the student write the paper? Was AI permitted? How much help is too much? Those questions matter and cheating has long been a cause of concern in education. Students have been finding creative ways not to do work since the invention of homework (here’s looking at you CliffsNotes). Some of us probably spent more time exploring ways to cheat on a test than we would have spent actually studying for it. But cognitive surrender can happen (and has happened) without breaking a single rule. It happens when fluency gets mistaken for expertise, when an immediate answer replaces productive struggle, when a convincing response goes unverified, and when “What does the AI think?” quietly replaces “What do I think?”

A student can follow every AI policy on campus and still surrender the very capacities higher education exists to develop. And that is the problem with trying to policy our way out of a judgment problem.

A policy can tell students what is permitted. It cannot teach them how to decide when nobody is watching and “knowing how to use AI” does not go far enough either.

Students absolutely need to know how to work with AI. But teaching students to use AI is not the same as preparing them to use AI. Knowing how to prompt a model does not tell you whether you should trust the response. Knowing how to automate something does not tell you whether it should be automated. And preventing students from using AI does not solve the problem either. A student protected from every opportunity to misuse AI has not necessarily developed the capacity to use it responsibly. Abstinence is very effective at teaching abstinence. History suggests it is considerably less effective at teaching what to do when abstinence ends.

If our AI strategy depends on the tool making the right decision for the student or the institution making every decision about the tool for the student, we have not developed student judgment. We’ve simply relocated it.

STOP BLAMING THE HAMMER

We ban the hammer in one classroom, require it in another, write increasingly elaborate rules about acceptable hammer use, and offer workshops on advanced hammer technique. Then we send students into the world without ensuring they can recognize whether the job calls for a hammer at all. We obsess over the hammer at precisely the moment when we need to pay more attention to the hands holding it.

Yes, AI is considerably more complicated than a hammer. It can respond, persuade, recommend, simulate expertise, and shape the behavior of the person using it. That makes human judgment more important, not less.

Because eventually students leave the classroom. There will be no syllabus telling them AI is permitted for brainstorming but prohibited for drafting. They will encounter unfamiliar problems and decide whether AI belongs in the workflow. They will receive authoritative-sounding answers and decide whether to trust them. They will have to recognize missing context, questionable assumptions, hallucinated evidence, privacy risks, and moments when efficiency is being purchased at the expense of judgment.

And sometimes they will have to make the hardest call of all: I could ask AI to do this for me. I shouldn’t.

Eventually, the stakes stop being an essay. They become hiring decisions, patient communications, financial analyses, research conclusions, strategic recommendations, and decisions affecting other human beings. The workplace will not grade students on whether they complied with their professor’s AI policy. It will depend on whether they developed the judgment that policy was supposed to protect.

Universities are still trying to define their relationship with AI. Embrace it? Restrict it? Require it? Ban it? The relationship may feel complicated. The responsibility isn’t.

AI can contribute to the work. It can extend capability and it can even change what is possible. But responsibility for judgment cannot be outsourced with the task.

That requires more than knowing how to use AI. It requires skepticism, accountability, judgment, discernment, and an understanding of how interacting with AI can shape our own behavior. Those are human capacities and they must be developed.

FROM AI ADOPTION TO HUMAN CAPACITY

At Fusion Collective, this is the problem we built Fusion Compass to address. Compass starts with a different question: Are people prepared?

Can they question AI, evaluate its output, recognize the psychological pull toward overreliance, determine where it belongs in a workflow, and collaborate with it without surrendering human responsibility? Access does not equal readiness. Prompting does not equal judgment. Adoption does not equal accountability. Universities can keep debating how much AI belongs in education. They should. But that debate cannot substitute for the deeper work.

The future of education will not be secured by keeping AI out of the classroom. Nor will it be secured by putting AI into every classroom. It will depend on whether the humans leaving those classrooms remain capable of thinking without surrendering the responsibility to think.