Playing by the rules
AI did not break higher education. Grade inflation, credential obsession, and decades of institutional cowardice did that. The students using AI to navigate the result are not the problem.
Yesterday, I graduated from The George Washington University with degrees in international affairs and economics, and I want to be honest about what that means before I say anything else. I have a disclosure to make, which is that my own GPA is not a number I am going to cite, partly because it is not the point of this essay and partly because it reflects choices I made that this essay is about. I took the economics courses that my peers, mostly pre-law students and political science majors, were not willing to take. I sat in econometrics and graduate-level macroeconomics seminars that were routinely under-enrolled because the grade distributions in those rooms were not compatible with what a competitive law school application requires. I am not telling this story to congratulate myself. I am telling it because it is a data point about what the incentive structure of undergraduate education actually produces, not in the abstract but in practice, in classrooms I was in, at a university I attended, among students who were making entirely rational decisions given the rules that had been handed to them.
Those rules are the subject of this essay. The arrival of AI tools capable of producing a competent undergraduate essay in minutes has generated a great deal of hand-wringing about academic integrity, student character, and the future of education. Almost none of it is addressed to the right target. The students using AI to complete assignments they do not care about, in courses they chose partly for their grade distributions, in majors they selected because the alternative threatened their GPA, are not the problem. They are the rational output of a system that was broken long before any of them arrived on campus, and they will graduate into a world that continues to reward the signal those tools help them produce. The blame belongs somewhere else. This essay is an attempt to say where.
Begin with grade inflation, which is the foundation of everything that follows. From 1990 to 2020, four-year college GPAs rose more than 16 percent at public and non-profit universities. At Harvard, 79 percent of grades awarded in 2020-21 were A’s. Yale recorded the same rate in 2022-23. The most common grade at an American university is now an A, which is to say the grade that was once reserved for exceptional work has become the baseline expectation for ordinary work at institutions that have spent decades advertising their own selectivity.
Here it is worth pausing to make an argument in partial defence of grade inflation before dismantling it, because the partial defence has real force. Harvard admits approximately 3 percent of applicants. Those applicants have been filtered through one of the most competitive processes in the world, and they arrive having spent four years in high school performing at the very top of their cohorts. It is not obviously surprising that a room full of people who were valedictorians of their high schools would produce a lot of A-level work. The argument that Harvard’s grade distribution reflects genuine achievement rather than inflated standards is not absurd. If you select the best students in the country and then give most of them A’s, you have not necessarily committed a fraud.
The problem is that this argument, even if true, produces the same practical outcome as grade inflation caused by institutional cowardice. Whether Harvard’s A’s reflect genuine excellence or lowered standards, the result for the admissions officer at a law school or the recruiter at a consulting firm is the same: a GPA of 3.9 from Harvard tells them almost nothing useful about how this applicant compares to the other Harvard applicants in the pile. The signal has been compressed whether or not the underlying performance warrants it. And the institutions downstream have responded accordingly.
This is where the divergence between institutions becomes important, and where the story gets considerably more uncomfortable for the schools that pride themselves on selectivity. MIT and Caltech have not followed the same trajectory. STEM-heavy institutions, where the mastery of technical material is testable in ways that resist the softening of standards, have maintained grading distributions that still carry information. A 3.5 from Caltech means something different from a 3.5 from Yale, and employers and graduate schools in technical fields know it. The divergence is not merely one of institutional culture. It reflects a deeper split between institutions where the work itself provides an objective standard and institutions where the primary product is increasingly the credential rather than the competency. Harvard and Yale produce extraordinary graduates. They also produce transcripts that are increasingly difficult to use as evidence of anything.
The incentive structure this creates downstream is where the damage becomes most visible, and law school admissions is the clearest example. Law schools report their median GPA in their rankings profiles, and US News uses those medians in its calculations. This creates a direct financial incentive for law schools to admit applicants with higher GPAs, which creates a direct incentive for applicants to maximise their GPAs, which creates a direct incentive for undergraduates to select majors and courses with favourable grade distributions. The Law School Admission Council has noted, with admirable directness, that one strategy available to pre-law students is to take easy courses in an easy major. Forum posts on pre-law advice sites say this explicitly. A 3.8 in communications from a school with known grade inflation will routinely outperform a 3.4 in mathematics from a more rigorous institution in the raw GPA calculation that law schools use for rankings purposes, even if every admissions officer knows the 3.4 represents more actual achievement.
The result, played out across hundreds of thousands of undergraduate trajectories, is a quiet curriculum distortion that nobody designed and that the institution has every incentive not to name. Students who might have studied economics or mathematics or chemistry, who might have developed quantitative skills that would have served them well in a career in law or finance or policy, instead major in subjects that offer more GPA protection. The courses that go under-enrolled are not the bad ones. They are often the best ones, the ones that are genuinely hard, that cover material students do not already know, that require them to think through problems they cannot solve by pattern-matching to lecture notes. Those courses represent a competitive risk that the incentive structure does not reward taking. So students do not take them, rationally, and the institution does not notice, or pretends not to.
AI has not created this dynamic. It has completed it. The undergraduate essay was already, in many cases, functioning as a credentialing exercise rather than an educational one. The assignment that asks students to summarise two scholars’ arguments and evaluate them is not, in most instances, teaching students to think. It is teaching them to perform the appearance of thinking in a format that can be graded at scale by an overwhelmed teaching assistant at eleven o’clock on a Sunday night. Students understood this long before AI existed. They have been gaming such assignments with varying degrees of effort for as long as they have existed. AI does not introduce a new behaviour. It lowers the cost of an existing one to near zero, which makes the underlying design failure visible in a way that a plagiarism policy can no longer manage.
What AI has done, if anything, is force the question of what the essay was actually for. If the purpose of a written assignment is to develop a student’s capacity to organise evidence, construct an argument, and think through a problem they have not encountered before, then AI assistance defeats that purpose entirely. The student who uses Claude or ChatGPT to write their political theory paper has not learned to think about political theory. If, on the other hand, the purpose of the assignment is to produce a document that receives a grade that contributes to a GPA that opens a door, then AI assistance is simply efficient. The question of which purpose the assignment actually serves is one that universities have been quietly avoiding for decades, because the honest answer is uncomfortable and the dishonest answer is profitable.
The institutional response to AI in education has been almost entirely focused on detection and prohibition, which is the wrong instinct applied to the wrong problem. Banning AI in a world where it is free, ubiquitous, and increasingly indistinguishable from polished undergraduate prose is not a policy. It is a performance of policy that disadvantages the students who follow the rules and advantages the ones who do not. It also allows the institution to present itself as defending academic integrity without doing anything that costs money or requires honesty about what the actual problem is.
What would actually change the incentive structure is more expensive and more honest than a prohibition policy. Oral examinations, project-based assessments, genuine research supervision, small seminar formats where a student’s thinking is tested in real time rather than through a document produced in private: all of these make AI assistance either irrelevant or immediately detectable. All of them require more faculty time and institutional investment than the lecture-and-essay format. All of them require universities to be honest about what they are producing, which means being honest about the degree to which what they have been producing is a credential rather than an education. That honesty is available. It would cost something. So far, most have preferred to draft an AI policy and send it to students before the semester starts.
The students graduating this spring, my cohort, inherited this system. We did not build it. We arrived to find a set of rules already in place, rules created by administrators chasing rankings, employers using GPA cutoffs as screening mechanisms for overwhelmed HR departments, and graduate schools publishing median statistics as marketing material. We played by those rules, with varying degrees of enthusiasm and varying degrees of willingness to take the courses that the rules made risky. The ones who used AI to produce the work the rules required were not cheating a fair system. They were finding a more efficient route through an unfair one. The institution that wants to blame them for that might begin by asking who designed the route in the first place.