Which Universities Are Adapting and Which Are Dying Slowly

In September 2026, Hampshire College announced that it was eliminating traditional letter grades across its undergraduate curriculum and replacing them with a portfolio assessment system in which students demonstrate competency through documented work products. The announcement got some press. Hampshire is small — about 1,000 students — and has always been eccentric; it’s the kind of place that has never had required courses and has been doing portfolio assessment in some departments since the 1970s. Easy to dismiss as the usual Hampshire weirdness.

But Hampshire is not doing this because of tradition. They’re doing it because they looked at AI and concluded, correctly, that a letter grade produced through an AI-assisted process proves almost nothing about what the student can do. A portfolio of documented work, where the student can be asked to explain and extend any piece of it, is much harder to fake without understanding. The assessment form follows from the technological reality.

You can disagree with the specific approach. Plenty of people do. But you cannot disagree that Hampshire is asking the right question: what does an assessment actually need to prove, given that AI can generate plausible-looking outputs in any domain, and given that the student may or may not have understood what was generated? Most universities are not asking this question. They’re writing honor codes with new paragraphs about AI.

The institutions that are genuinely adapting share several characteristics that are easier to see in retrospect. First, they have leadership that was willing to be publicly wrong about specific things. Arizona State’s Michael Crow has been making large, sometimes wrong, sometimes right bets about higher education’s future for twenty years. The institutional culture he built tolerates being publicly wrong in a way that most university cultures do not. MIT’s faculty governance, which runs a significant portion of curriculum decisions with real faculty authority, moved toward oral and practical examination partly because the faculty with actual research in AI were better positioned than administrators to understand what AI could and couldn’t do.

Second, the adapting institutions generally have something other than the credential to sell. MIT has research affiliation. ASU has scale and access. The institutions that are struggling tend to be the ones whose primary value proposition has always been the credential itself — the piece of paper from a name that employers recognize. When the credential’s signaling value weakens, they have nothing to fall back on.

Third — and this is the uncomfortable one — the adapting institutions tend to be the ones that were already financially stable. Redesigning curriculum, retraining faculty, building new assessment infrastructure: these things cost money, and they cost time during which revenue may decline. Institutions with substantial endowments or state support can absorb that transition cost. Institutions operating on thin margins cannot.

The tuition-dependent private liberal arts college occupying the middle of the prestige spectrum — not Harvard, not a directional state university, but something like Gettysburg College or Hiram College or Wheaton College in Illinois — is in the most dangerous position in American higher education right now. These schools typically have enrollments between 1,500 and 4,000 students, tuition in the $45,000-$55,000 range before aid, endowments of $50 million to $300 million, and educational models built around the faculty-led seminar, the close-read discussion, and the writing-intensive curriculum. Everything they do is precisely what AI is most capable of simulating.

The writing-intensive curriculum assumes that students will struggle to produce coherent arguments in writing and will develop the ability to do so through practice, feedback, and revision. This pedagogical model has a lot of evidence behind it. Writing is genuinely one of the best ways to develop clear thinking. The problem is that AI eliminates the productive struggle. The struggle is where the learning happens. When you can produce a well-organized five-paragraph argument in thirty seconds, the exercise no longer produces the cognitive development it was designed to produce. The faculty at these schools know this. They don’t have an agreed-upon solution.

Some of them are moving toward more oral assessment, more in-class writing, more performance-based evaluation. These adaptations are real and positive. But they run directly against the economic model: you cannot run a writing-intensive, discussion-based, highly individualized education at scale, and these schools need scale to survive. The tuition revenue per student is fixed — actually declining in real terms as aid packages get more aggressive. The cost of faculty-intensive education is rising. The middle path is closing.

Purdue University is an interesting counter-case. It’s a large flagship with engineering and agriculture as its anchors — fields where practical, hands-on assessment has always been central. You cannot write an AI essay about how you designed and tested the structural joint. You have to design and test the structural joint. Purdue has been expanding its practical, project-based curriculum in response to AI not because it’s pedagogically innovative but because it turns out that what they were already doing is precisely what AI cannot do for students.

Purdue’s AI in Education task force, which published its report in January 2027, made a recommendation that seemed obvious in retrospect: all Purdue courses should have at least 40 percent of their assessment weight in activities that require physical presence, direct observation, or real-time demonstration. This doesn’t require eliminating AI from courses. It requires that the bulk of what earns a grade cannot be produced by AI alone. The implementation is uneven across departments, but the direction is clear.

Compare this to a prominent liberal arts university I am declining to name, which spent the first half of 2025 developing an AI use disclosure form, spent the second half of 2025 debating whether the form was FERPA-compliant, and had not implemented any substantive assessment redesign as of April 2027. Their faculty senate is currently considering a motion that would require each department to form a subcommittee to study AI’s implications for departmental assessment practices and report back to the full senate within eighteen months. Eighteen months. The technology they are studying has a capability curve measured in weeks.

The international dimension of this is underappreciated. American higher education exports prestige as much as it exports education — the American degree carries a premium in labor markets in India, China, Southeast Asia, and Latin America that has been sustained for decades. The question is whether that premium is defensible as the credential’s domestic signaling value weakens.

The early evidence suggests it’s weakening internationally faster than domestically, because in the most competitive international labor markets, the credential from an American university was always primarily a signal about global competence, and AI is specifically disrupting the knowledge-based components of global competence. The Indian tech sector, which has been the largest driver of American university prestige exports in terms of enrollment volume, is already shifting hiring toward demonstrable skill over credential — coding assessments, portfolio review, practical tests — in ways that the US domestic market hasn’t fully made yet.

Chinese universities have moved aggressively toward AI integration in ways that prioritize their own credential system. The top Chinese technical universities — Tsinghua, Peking, Zhejiang — are producing graduates who have worked with AI tools throughout their education in structured, assessed ways. Whether this makes them better prepared for AI-augmented professional work than American graduates who used AI to avoid doing their own work is an empirical question. My guess is yes, but the data is early.

What would it look like for an American university to genuinely adapt its model for the AI era? I’ve been piecing together the elements from the places that are getting it right. Start with a commitment to assessment that requires demonstrated understanding, not just produced output. Build in more oral examination, practical demonstration, and real-time interaction. Use AI as a tool in the curriculum explicitly and deliberately, teaching students to work with it critically rather than pretending they won’t. Shift the faculty role from content delivery — which AI does very well — toward mentorship, feedback, and the cultivation of judgment. Focus on the things a university can provide that are genuinely not replicable by a subscription: peer cohort, relationships with practitioners, physical infrastructure for fields that need it, and the social experience of extended engagement with a community of learners.

None of this is impossible. None of it requires abandoning the university model entirely. What it requires is institutional will to change faster than the institutional culture wants to change, and financial flexibility to absorb transition costs. The schools that have that combination are adapting. The schools that lack either one are dying slowly, many of them while publishing their strategic plans for excellence.