New Article: Mesa-Optimization is Destroying Education

The Palladium Letter

SubscribeSign in

New Article: Mesa-Optimization is Destroying Education<br>Educational institutions have learned to optimize for everything except learning. Artificial intelligence could force them to measure what’s important.<br>Jul 17, 2026

16

Share

Kirubakaran Manoharan/Princeton University<br>This article by Matt Duffy was published at Palladium Magazine on July 17, 2026.<br>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.”<br>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.<br>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.<br>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.<br>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.<br>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.”<br>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.<br>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.<br>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.<br>Campbell described a few contemporary phenomena...