In 1976 the social scientist and cyberneticist Donald T. Campbell developed a principle, now known as Campbell’s Law, that would later become one of the most reliable diagnostics of institutional decay. When any quantitative indicator is used for social decision-making, he warned, it becomes a target for manipulation. As the pressure to meet that target grows, the measure itself begins to lose its connection to the underlying reality it was meant to evaluate, which “will distort and corrupt the social processes it is intended to monitor.”
American education has become Campbell’s most dramatic case study. Today’s gifted high school graduates will be greeted by a rude awakening when they sit down for their university’s mathematics placement exams. Instead of acing the test, as they are accustomed to, they are likely to discover that the problems presented are completely alien to them, and find themselves placed in a remedial math program to relearn concepts they should have mastered in middle school. Advanced students are increasingly discovering that they are unprepared for college-level studies and struggling to succeed in subjects where they believed themselves to be proficient.
At UC San Diego, a 2025 report found that roughly one in eight incoming freshmen tested below high school level in math, with nearly 12 percent of the entering class enrolling in remedial courses. At the same time, the average high school math GPA among those placed in the lowest remedial track stood at an A-minus. UCSD is a prestigious enough school that we can be quite certain that if it’s happening there, it’s at least a somewhat representative signal for the rest of the university ecosystem. An intervention is necessary to prevent this trend from worsening.
In UCSD’s case, these results came after standardized testing was dropped as a student assessment measure, removing a tool that counteracted grade inflation in high schools—a widespread problem even before the appearance of LLMs. While artificial intelligence is certainly used by students to complete coursework at the cost of their own learning, the same technology underlying AI may provide the foundation for a more transparent and effective educational system.
How Signal Gets Distorted
One of the most interesting and important concepts to come out of AI alignment research is mesa-optimization. The term comes from a 2019 paper, “Risks from Learned Optimization in Advanced Machine Learning Systems.” Notably, two of its writers came from the Machine Intelligence Research Institute (MIRI) founded by Eliezer Yudkowsky. “Mesa-optimization” was coined as a deliberate inversion of “meta”: where a meta-optimizer sits one level above the system it tunes, a mesa-optimizer sits one level below.
A training process known as the “base optimizer” searches for a model that scores well on an objective. Sometimes the model it finds is itself running a search internally, pursuing its own objective; that internal objective, which the training process never specifies directly and cannot easily inspect, is the “mesa-objective.” The gap between what the model was trained to do and what it has actually learned to pursue is what the authors call the “inner alignment problem.”
Complex systems designed to optimize for some goal can create “mesa-optimizers,” which are subsystems that optimize for things that are correlated with the original goal but are not actually aligned with it. When this mesa-objective diverges from the base objective, the system continues producing outputs, but those outputs no longer match its original purpose.
The process of natural selection, for example, helps select for beneficial traits that eventually propagate across the system. Natural selection created humans, and humans independently select for many subjectively beneficial behaviors—some of which are aligned with the process of natural selection, and a whole host of behaviors which are not, such as celibacy, contraception, or adoption. Optimizing for reproduction, the “base process” produced a new class of agents—humans—who possess independent goals unrelated to reproduction itself.
In any large human institution we can identify a base objective, or what it was designed to do, and a subsequent web of downstream mesa-objectives that emerged through pressure from incentives and survival constraints. Campbell’s Law is relevant here in the sense that left unchecked, mesa-optimization can fundamentally distort the base objective.
Campbell described a few contemporary phenomena that illustrated the widespread institutional failures of the 1970s, which saw soaring crime rates and helicopters evacuating American embassy personnel from the rooftops of Saigon. He cited research on how plea bargaining was misused by prosecutors to charge criminals with unsolved crimes in exchange for a more lenient sentence, improving the clearance rate of their district. He pointed out that during the Vietnam War, the success of U.S. anti-guerrilla warfare was measured by the number of enemy combatants killed in action, which led to commanders inflating those numbers and even including civilian dead as KIA. Both these interventions to meet performance benchmarks worsened the system they were meant to measure.
Today, the same institutional pressures that led UCSD to drop standardized testing have also led to a degradation of standardized measurements like the SAT. Long reading passages taken from literary texts, for example, have been replaced by short, digestible excerpts, and vocabulary-based questions have been done away with. The restoration of the test will not necessarily solve the problem of evaluating how students will perform in a more demanding environment.
Education’s Base Objective
Campbell’s Law describes how corruption works, but not what stops it from being corrected. A benchmark resistant to mesa-optimization must be anchored to a correctly identified base objective—but base objectives are not designed in a vacuum. They are socially derived, which means every pressure that corrupts a metric can also corrupt the objective behind it. Yet the incentives for grade inflation are as old as grades themselves; something used to resist it.
In 1958 Hannah Arendt argued that the crisis in American education was not an education problem at all, but a fragment of a modern crisis of authority. Adults no longer felt entitled to present the world to the young and take responsibility for it. America, in her account, was the extreme case—the country where mass education had made the problem most severe, and where adults had completely surrendered their own judgment to the reigning theory of the day:
Nowhere have the education problems of a mass society become so acute, and nowhere else have the most modern theories in the realm of pedagogy been so uncritically and slavishly accepted. Thus the crisis in American education, on the one hand, announces the bankruptcy of progressive education and, on the other, presents a problem of immense difficulty because it has arisen under the conditions and in response to the demands of a mass society.
Arendt’s conclusions scandalized progressive educators of the time, and would scandalize them now. Fashionable pedagogical theories centering students’ perceptions relieved the adults in the room of the need to exercise judgment and authority. Reformers, she claimed, posited that there exists a “child’s world and a society formed among children that are autonomous and must insofar as possible be left to them to govern. Adults are only there to help with this government.” Instead, she asserted that educators must be guardians that protect the world from the newness that comes with each generation, and each generation’s newness from the world.
Historically, education was a virtuous pursuit venerated by elites, and mostly confined within elite circles. Harvard, founded in 1636, existed largely to train clergy and the sons of the gentry in a classical curriculum, and as late as 1900 only about 3.5 percent of college-age Americans were enrolled in higher education at all. Even as the economy became more complex, college remained a mostly elite pursuit, enrolling only around 7 percent of college-age Americans by 1940.
The GI Bill of 1944 then sent millions of returning veterans onto campuses that had never planned for them, and the Higher Education Act of 1965 opened the federal student-aid spigot that has underwritten mass enrollment ever since and driven ballooning tuition costs. Expansive institutions were built to spread education. This multiplication of the settings in which students were to take instruction from teachers happened at the very moment when the prevailing theory of reformers like John Dewey said adults had no right to compel this, for school was a place of “adjustment learning” and student self-discovery. To Arendt, this meant that “the authority that tells the individual child what to do and what not to do rests with the child group itself—and this produces, among other consequences, a situation in which the adult stands helpless before the individual child and out of contact with him.”
The next sixty years of progressive education validated Arendt. Mass education led to massive inconsistency in student preparation and ability. Failure rates and differential educational access became politicized during the Civil Rights era. Unequal outcomes, combined with financial pressures, pushed institutions to reduce attrition. The need for steady tuition revenues, the combined pressure of social and parental expectations, and a later focus on “student satisfaction” slowly shifted universities away from pedagogy and towards customer-service-based objectives. By the late 1980s a new vocabulary had taken hold among administrators, who spoke increasingly of “success rates” as though it were a base objective.
College GPAs began rising at roughly a tenth of a point per decade, and the A has become the most commonly awarded grade—close to 43 percent of the total—where a generation earlier it was far rarer. High schools faced a similar shift in educational standards. With colleges as the gatekeepers for “good jobs,” there was intense administrative, financial, and parental pressure for high schools to increase college admissions rates, and the only reliable way to do so was by inflating GPAs. Between 2010 and 2021 the average high school GPA climbed from 3.17 to 3.36 even as average ACT scores fell to their lowest point of the decade.
In 1995, the College Board “recentered” the SAT scale to bring the national average back to 500 in each section, instantly adding roughly eighty points to the typical verbal score and twenty to the math. Gamed disability accommodations, like extended test time, drove further score improvements. By the 2020s the test-optional movement had spread to the great majority of four-year colleges, with the University of California system dropping the SAT and ACT entirely. The result is declining fidelity of educational institutions to actual competence. Each is an adjustment that avoids judgment of the students in favor of either changing measurements or ignoring the measurements that might drive inconvenient, judgmental conclusions.
In response, employers struggling to discern high-quality candidates from a pool of contrived credentials started relying on other factors. How many internships did the student have and where? What actual work did they publish on their GitHub or in other public channels? The only reason these became necessary signals is because the information embedded in things like GPAs and SAT scores became so unreliable. When the primary signal collapses, the job market inevitably turns to expensive, high-variance alternatives. Employers look to internships and project portfolios, and admissions offices turn to essays, extracurriculars, philanthropy, and other socioeconomically correlated indicators. These costly signals replace the original, cheap signals that have stopped carrying information.
What all of these decisions have in common is twofold: administrative pressure leading to a drop in rigor, and a collective avoidance of uncomfortable discussions with students that don’t make the grade—particularly students who clearly try hard or who have powerful parents advocating on their behalf. At each step we have built in incentives to soften the feedback to a student, for fear of weakening their future prospects, which in turn weakens the signal to the next institution in the education chain. At every critical juncture the only required disruption is an adult who knew better saying so.
If the people moving through these education systems are becoming less educated because the system is focused on generating statistical outcomes over meaningful ones, then slowly our human capital will degrade. This has been commonly described as a stratification process—those who are the most intelligent or the best at what they do will finally see outsized rewards as they outstrip the hoi polloi. But weaker performance across the board is more likely. Intelligent, outlier students will be rewarded for a mediocre level of output that is sufficient to surpass that of their peers. Top universities and companies will still obtain and retain this “top talent,” but that will not correspond to a diffusion of competence into the broader workforce.
How AI Can Restore Signal
Built properly, AI educational systems could focus on the base objective rather than optimize for mesa-objectives. Without external pressure, however, we will get incremental “AI integration” from established ed-tech companies—the same vendors that profit from the current system’s dysfunction. They have no incentive to expose the gap between credentialed and actual competence. But AI can repair the assessment layer.
Recent disruptions in the educational space show it’s possible to build AI-native school systems that break the cycle. Alpha School is an Austin-based private school where instruction and assessment are primarily conducted by a proprietary AI system. Models generate endlessly varied assessments, preventing gaming, and require 90 percent accuracy before unlocking new material at the student’s own pace. It can evaluate answers uniformly across classrooms and districts, eliminating the idiosyncrasies of teacher judgment. Its scoring distributions, item difficulty levels, and error patterns can be published and audited. It is capable of enforcing the base objective, not the institutional mesa-objectives, in a way that is not as leaky as the current system.
Alpha School is not without its problems. Under this AI-first system, teachers are replaced by “guides,” who oversee student learning. In 2025 they were revealed to mostly consist of remote workers from the Philippines, which is something no one would want for their child. In this way Alpha solved the measurement problem while completing the abdication of responsibility Arendt spoke of—no adult in the room stands behind any judgment at all. Alpha students perform extremely well on Measures of Academic Progress (MAP) tests that are used nationally to evaluate student progress. But relying solely on this metric may risk succumbing to Campbell’s Law as well, and it will take some time before we see how these students perform in higher education and the workforce.
The relevant question, however, is whether or not AI is more resistant to the specific failure modes that corrupted the current system. AI is more resistant, precisely because the gaming moves up a level, and at that level it becomes more visible to parents, teachers, and other interested parties.
An administrator choosing an AI assessment system can be selective about difficulty settings, subject coverage, and competency thresholds. These choices can be gamed in some ways; an administrator might select easier settings to boost pass rates. However, unlike a teacher inflating grades behind a closed door, these choices are auditable. State-mandated model cards could publish difficulty calibrations, the underlying standard teaching corpus, and expected score distributions for median students at each level. When a district chooses an easy setting, everyone could see that choice.
This transparency also enables upward mobility that the current system blocks. A student in a low-rigor district who aces every assessment can have parents advocate for a difficulty upgrade for that child specifically, without changing schools. The student’s capability becomes visible against a common standard, rather than hidden behind a local grade from a known weaker school that no one outside the district can interpret.
Under the current system, a struggling teacher can plausibly claim that they inherited weak students from the prior grade. This excuse disappears when we have continuous assessment data. At the least, AI-based micro-quizzing woven into daily instruction can generate a time series for each student. We can see what a student knew when they entered a classroom and what they knew when they left. We can measure inflection points. A teacher who consistently produces upward trajectories looks different from one whose students stagnate or decline, regardless of where those students started. Teachers can also leverage micro-assessment as a pedagogical tool, to enable course correction very early on. It’s another way to remove a place where abdication hides.
What Alpha School gets right is that none of this can be bolted on to existing systems. It should not be an AI augmentation of current testing regimes or a learning management system add-on. The current system’s distortions would simply absorb and neutralize it. It must be a strategic commitment from a state education system, implemented as a replacement for existing assessment infrastructure.
The natural implementing body is the state, but education under this system would also resemble pre-industrial “schoolhouse” models in some ways. If teachers are no longer responsible for assessment and students advance through material at their own pace, the teacher’s role will become defined by observation, mentorship, and instilling academic discipline. Interacting with a sycophantic AI model can be damaging for a growing mind; human feedback becomes more necessary to promote the habits of mental discipline and prevent AI psychosis. Alpha’s use of “guides” is not sufficient, but it is also unsurprising, given their incentives. That does not invalidate the concept behind their technology. The public education system will need to perfect the model, as tall an order as that is.
Planners need to accept that honest measurement will reveal that a large portion of those we’ve certified as educated are not. GPAs will crater and failure rates will spike. This short-term cost falls on whoever implements it, but the long-term benefit diffuses across the entire economy. It requires a state with unusual political cover—a governor with political capital to spend, or a jurisdiction desperate enough to try something because they are so far behind in the current regime. It might not even happen in the U.S. first.
Even a short period of measured decline, as the real outcomes of the education system are brought to light, is worth the restored signal downstream. It means students will know where they actually stand. It means employers and graduate programs will trust credentials again. And it means breaking the expensive and humiliating signaling arms race, in which over a third of undergraduates at elite academic institutions have disability accommodations.
We set up many human systems with clear, prosocial objectives, which are then corrupted by thousands of small mesa-optimizations that distort the outcomes. We can apply this framework to some of our other intractable problems like community safety, healthcare, and purportedly unbiased information systems like journalism. We ought to examine those systems to see where unbiased measurement signals can restore overall signal propagation.
Too much present-day hype around AI focuses on sci-fi level potential, or risks, rather than the more modest ways it could help us solve some of our intractable, everyday problems. Human systems do not move back toward truth on their own; they need, at minimum, a measurement regime that cannot be gamed. AI gives us the first chance in ages to restore the base objective of education and make measurements honest again. The rest was never a measurement problem.