The School That Got AI Right

High Tech High is a network of charter schools in San Diego. It is not a technology company or an AI lab. It has 16 schools, roughly 5,500 students, and has been operating since 2000. Its pedagogy is built around project-based learning: students work on extended projects that require genuine engagement with real-world problems, produce public exhibitions of their work, and are assessed primarily on their demonstrated capability to do real things rather than their performance on standardized tests.

It was, almost by accident, well-prepared for AI.

When ChatGPT launched in November 2022, High Tech High teachers found that most of their assessments were already relatively resistant to AI substitution. A student working on a three-month project to redesign a component of their school’s HVAC system, who needs to interview facilities staff, take measurements, research energy efficiency standards, and present findings to the principal, this student cannot fully outsource the project to AI. They can use AI as a research assistant, a writing aid, a sounding board. They cannot hand off the project. The learning is embedded in doing, and the doing requires physical presence and genuine engagement with specific people and specific problems.

This is not to say High Tech High is a utopia or that project-based learning is universally applicable. High-quality project-based learning is expensive: it requires teachers skilled enough to design genuinely demanding projects, organizational infrastructure for presentations and exhibitions, and enough administrative tolerance for the messiness of extended open-ended work. It works better in small schools with consistent teaching cultures than in large, fragmented institutions. And it produces students who are sometimes less prepared for the specific genre of standardized testing that colleges use for screening, even if they are better prepared for the actual work colleges purport to prepare students for.

The more general point is that High Tech High’s approach embodies several specific design principles that any school can adapt:

Assessment design that requires demonstration, not just production. A student who presents their work to an audience and defends it against questions is demonstrating something that is difficult to fake. The presentation can use AI-generated slides; the defense cannot.

Projects that require physical engagement. A student who builds a physical prototype, conducts original observations, or interviews real people is generating data that AI cannot fabricate. The analysis of that data can be AI-assisted; the data collection cannot.

Iteration under feedback. Assigning extended projects with multiple revision cycles, in which a teacher sees the evolution of the student’s thinking, not just the final product, creates a record of process that AI-generated polish cannot produce.

None of these require abandoning subject knowledge or curricular content. They require embedding that content in work that can’t be offloaded.

The Mastery Transcript Consortium is a different kind of example. Founded in 2017 by a group of independent schools who were frustrated with GPA-based transcripts as measures of genuine capability, it has developed an alternative credentialing system in which students earn “mastery credits” in specific competencies, analytical thinking, creative expression, design, quantitative reasoning, demonstrated through portfolios of actual work. The transcript shows not what grades a student received but what they can actually do, with evidence.

This system was designed before AI. It turns out to be much more AI-robust than the conventional GPA transcript, because the competency claims are backed by work portfolios that tell the story of the student’s actual engagement. An admissions reader looking at a Mastery Transcript is evaluating the quality and growth of a student’s work over time, not a number generated by an algorithm that, in the current environment, is contaminated by unknown AI contribution.

As of 2025, more than 300 schools were participating in the Consortium, and several selective universities had begun formally accepting Mastery Transcripts in admissions. The growth is modest compared to the scale of the conventional credentialing system: 300 schools out of roughly 130,000 secondary schools in the United States. But the direction is right, and the rate of growth is accelerating as AI makes the conventional transcript less reliable.

The Singapore Ministry of Education released a Framework for AI in Education in October 2024 that is, frankly, the clearest policy document produced by any government on this question, and is underread outside of Southeast Asian education circles. Its core argument is that AI competence for students has three distinct components that require different instructional approaches: AI literacy, understanding what AI is and how it works, AI collaboration, using AI effectively as a tool for genuine work, and AI resilience, the ability to perform critical cognitive tasks without AI assistance when required.

The third component is the one that most Western education policy documents omit. AI resilience is the argument that there remain contexts — professional, civic, interpersonal — where humans need to exercise judgment and capability without AI mediation, and that education should develop those capabilities, not just optimize for AI-assisted performance. A doctor who cannot evaluate a patient without consulting an AI, a lawyer who cannot form a judgment before running it through a model, an engineer who cannot assess a safety risk without an AI prompt, these are professional failures that AI educational integration, done poorly, produces.

Singapore’s framework explicitly preserves unassisted assessment for specific competencies, clearly separated from assessments where AI use is permitted and expected. This is not the same as “ban AI.” It’s a discriminating judgment about which capabilities matter enough to measure in their unassisted form. It’s more sophisticated than the American policy response to date, which has largely been either blanket bans or blanket tolerance depending on which institution you ask.

The schools that are handling AI well share a characteristic that is pedagogically obvious but institutionally difficult: they have clear answers to the question “what are we actually trying to develop in students?” Not vague answers — specific, defensible answers that connect to genuine human capability, to research evidence, and to the world students will actually inhabit. High Tech High’s answer involves real-world problem-solving, collaboration, and public demonstration. Singapore’s framework distinguishes AI-assisted and unassisted capabilities. The Mastery Transcript Consortium’s answer is demonstrated competency over time, documented in work portfolios.

These answers are different, but they share a common structure: they identify what matters and build assessments that actually measure it. The schools that are struggling with AI are the ones that never had to answer this question clearly, because the answer was implicit in the structure of conventional assessment: we teach what’s on the tests, and the tests measure what we teach.

That circular structure is now broken. AI can pass the tests. The schools that had a coherent answer to “what are we developing?” are fine. The schools that were essentially teaching to tests are discovering that they need a coherent answer in a hurry.

The case for optimism, which I hold cautiously, is that the disruption AI is creating in education is forcing institutions to confront questions they should have been forced to confront much earlier. What is a grade, really? What does a college degree certify? What skills are worth 12 years of education to develop? What does “learning” mean when information is freely available?

These are not new questions. They were being asked by educational philosophers before AI, by Dewey in 1938, by Freire in 1968, by Holt in 1964. What AI has done is make them operationally urgent rather than philosophically interesting. An institution that cannot answer them clearly is not just vulnerable to AI disruption in some abstract sense — it is failing to defend the value of what it does in terms that parents, students, and policymakers can evaluate.

The schools that get AI right are mostly the ones that got education right before AI arrived. They had clear answers to the fundamental questions, built structures that matched those answers, and developed teachers who could operate in those structures. The challenge isn’t discovering new educational principles — those exist, and they’re well-documented. The challenge is building the institutional will and the funding to implement them at scale, in systems with very different starting conditions, before the competitive pressure from AI alternatives makes the status quo untenable.

That window is shorter than most school systems think.

The International Baccalaureate provides another example worth studying. The IB Diploma Programme, which serves approximately 1.1 million students globally, has always assessed students through a combination of written work, oral presentations, experimental investigations, and a substantial independent research project (the Extended Essay). The Internal Assessment components, which make up a significant fraction of each student’s final grade, are evaluated through presentations and process documentation, not just final products. When IB schools surveyed their teachers in 2024 about AI’s impact on assessment integrity, the results were more optimistic than comparable surveys of conventional school systems: IB teachers reported that AI had created fewer assessment integrity problems because the assessment design already required process documentation and oral defense.

This isn’t a coincidence. The IB’s design was shaped by a deliberate philosophy about what learning outcomes matter — intellectual curiosity, international-mindedness, research skill, oral communication — that happened to produce assessments resistant to AI substitution. The philosophy predates AI by 60 years.

What this suggests is that the question “how do we adapt our assessments to AI?” has a cleaner answer for institutions that have clear educational philosophies than for institutions that have been optimizing primarily for efficient credential production. The former can ask, “which of our assessments still serves our stated educational goals, and which doesn’t?” The latter has a harder question to answer first.

There is no shortcut around that harder question. Buying AI detection software doesn’t answer it. Updating the honor code doesn’t answer it. Adding “AI literacy” units to the curriculum doesn’t answer it. The question is: what human capabilities are we responsible for developing, and are our current structures actually developing them? Any school that can answer that question honestly is already better positioned than most. The schools that get AI right will be the ones that use the disruption as an opportunity to answer it clearly, and then build backward from the answer to the assessments, curricula, and daily classroom structures that serve it. The technical tools exist. The hard part was always institutional, and it remains so.