The University as Credential Machine: What Happens When AI Breaks the Lock

The bar exam has a pass rate of around 56 percent for first-time takers in California. GPT-4 passed it in February 2023 at the 90th percentile. By 2025, Claude 3.5 Sonnet was scoring in the 95th percentile consistently, regardless of jurisdiction. As of early 2027, every frontier model passes every professional licensure exam in the country without meaningful effort.

Nobody in legal education wants to say this out loud. They are still running bar prep courses, still charging $3,000 for Barbri, still treating the exam as the meaningful gateway it was when a human mind was the only thing that could pass it. The silence is institutional. The denial is financial.

What actually happened to universities in the two years since AI tools became genuinely widespread — not the chatbot novelty of 2023, but the embedded, always-on, deeply capable systems that showed up in earnest in late 2024 and early 2025 — is something most commentary refuses to address directly. The institutions themselves are in defensive crouch. The critics are either too optimistic (AI will make education better!) or too apocalyptic (degrees are dead). Neither camp is looking at the actual data.

Here is what the data shows. First-year enrollment at four-year private universities declined 7.2 percent between fall 2025 and fall 2026, according to the National Student Clearinghouse. Community college enrollment went up 11 percent in the same period. Online-only programs — the ones that cost $12,000 total for a “degree” — are growing at roughly 22 percent year over year. The students who are opting out are not the ones who could never afford college. They’re the ones who could afford it but are doing the math and deciding it doesn’t pencil out.

The math is not complicated. A four-year degree at a mid-tier private university costs, all-in, somewhere between $220,000 and $280,000 at 2027 prices. The median starting salary for a new graduate with a business degree is $58,000. If you spend $250,000 to earn $58,000, you are buying an option on upward mobility, not a guaranteed return. You have always been buying that option. What’s changed is that the option is getting cheaper to replicate elsewhere.

That’s the uncomfortable truth at the center of this: the university was always fundamentally a credentialing and signaling machine, with learning as a secondary function. The people who loved learning the most — the ones who stayed for PhDs — were selecting themselves out of the labor market for five to eight additional years. The people who went to get jobs were going because the credential sorted them into better starting positions. AI hasn’t broken learning. It’s broken the sorting mechanism.

Consider what a degree actually certified, before AI. It certified that you could sit through unpleasant tasks for four years without quitting. It certified some minimum competence in a domain, usually assessed through tests and papers that required either real understanding or skilled performance of understanding. It certified that you had the social capital to get admitted, the financial resources or risk tolerance to attend, and the organizational discipline to complete. These are genuine signals. Employers used them because nothing better existed.

AI hasn’t made any of those underlying qualities easier to acquire. Discipline, social capital, sustained effort — these are not things a language model gives you. But it has made the performance of the credential’s knowledge component trivially reproducible. A student who uses Claude to write every essay, complete every problem set, and prepare for every exam can acquire the credential without acquiring the underlying competence the credential was supposed to certify. This is not a hypothetical. It is happening at scale.

A 2026 survey by the Chronicle of Higher Education found that 71 percent of undergraduate students use AI tools to assist with written assignments. Thirty-eight percent said they use AI for “most or all” of the drafting. Faculty know this. Most have tacitly accepted it. The academic integrity apparatus — TurnItIn, originality checkers, AI detectors — has been almost completely defeated. Detection models have false positive rates high enough to make enforcement against any individual student legally and ethically treacherous.

The institutions that are actually adapting — and there are some — have stopped fighting the behavior and started redesigning around it. Arizona State University, which was already the largest university in the country by enrollment before any of this, made the most aggressive move in January 2026 when it announced that all written assessments would be conducted in AI-assisted environments with students required to demonstrate, verbally and interactively, that they understood what the AI produced. The exam is no longer about output. It’s about being able to explain and extend the output. That’s a fundamentally different cognitive task, and it’s actually harder to fake.

MIT went the other direction, which was predictable for MIT. They stripped assessment back to oral examination and in-person demonstration for most of their core curriculum. You cannot do a problem set with AI if the problem set is you standing at a whiteboard explaining your approach to two professors who are asking follow-up questions in real time. The pass rate dropped 15 percent in the first year. The faculty called that a feature, not a bug. The students who couldn’t explain their own work were, the argument went, the students who shouldn’t be getting MIT degrees.

Both approaches have merit. Both require institutional will that most universities simply do not have. The median American university is not MIT or ASU. It is a place with heavily unionized faculty resistant to assessment redesign, administrative layers that move at geological pace, and a board of trustees whose primary concern is the endowment and the liability exposure that comes from failing too many students.

The student loan system is where the economics get genuinely ugly. Total outstanding student loan debt in the United States crossed $1.9 trillion in March 2027. The delinquency rate on federal loans has risen to 18.3 percent, the highest since the Department of Education began tracking it in its current form. The Biden-era relief programs helped at the margins but didn’t change the structural equation: people borrowed money to buy a credential whose labor market premium is eroding in exactly the fields where AI is most capable.

Lawyers. Accountants. Financial analysts. Radiologists, though the resistance there from the AMA has been ferocious. Entry-level software engineers. Copywriters. The jobs that were supposed to justify the debt are the jobs facing the most direct AI substitution pressure. The jobs that are AI-resistant — plumbers, electricians, physical therapists, surgeons who operate with their hands — generally don’t require the expensive degrees. The mismatch is not subtle.

The cruelest part is the lag. Students who enrolled in fall 2023 or 2024 made their decisions before the capability curve became undeniable. They’re finishing degrees now into a market that has moved. A 2027 JD graduate entering a legal market where AI handles contract review, discovery, and first-draft pleading is entering a different profession than the one they enrolled to join. The law schools took their tuition. The profession changed. Nobody is held accountable for the gap.

What’s actually worth something now, if the traditional credential is weakening? The honest answer is: proof of work that AI cannot produce. The portfolio that demonstrates judgment over time. The publication record that shows original contribution to a field. The relationships — the genuine human ones, built through years of showing up — that create professional trust. These things don’t come from a degree. They come from doing the work.

The paradox is that universities remain potentially excellent places to build all of those things. The library access, the peer cohort, the faculty connections, the structured time — these have real value if you use them deliberately. The problem is that most students aren’t using them deliberately. They’re using AI to complete assignments so they can graduate, which is rational given the incentive structure universities have built, and they’re leaving without the underlying capital the institution could have provided.

That’s where the tragedy lives. Not in the technology. In the failure of institutions to reconceive their purpose before the credential crisis made reconception urgent. The university had two years to get ahead of this. Most of them spent those two years writing AI use policies, convening task forces, and publishing blog posts about embracing AI while changing almost nothing about how they actually assess learning.

The lock is broken. The question is what gets built in its place, and whether the institutions that hold the keys are the ones who get to decide.