72% of What K-12 Teaches Is Now Automatable. Here's What Should Replace It.

The Audit Nobody Asked For

In March 2026, a single AI agent running on a hobbyist's home setup published 172 articles across seven websites with sixteen different journalist voices, six parallel editorial critics, and real citations. The operation was not running at capacity. A week earlier, a developer clean-room rewrote 512,000 lines of leaked proprietary code in two hours. GitHub reported that Copilot opened 88,943 pull requests in 30 days. Klarna eliminated 3,104 customer service positions and saw resolution quality improve.

These are not projections. These are measurements from the past 90 days.

We took the same workforce displacement methodology we use for our labor automation research and applied it to the standard K-12 curriculum. For each major subject area and skill category, we asked: Can a commercially available AI system (Claude, GPT-4o, Gemini, Copilot) perform this task at or above the level expected of a graduating high school senior? The answer, for 72% of tested competencies, was yes.

What's Already Dead

These are not skills AI will eventually replace. These are skills AI replaced before this article was published.

Traditional Skill	AI Performance (vs. 12th Grader)	Evidence
Factual recall (history, science, geography)	Exceeds	GPT-4 scores 90th percentile on AP exams (OpenAI, 2023)
Formulaic essay writing (5-paragraph, persuasive)	Exceeds	AI-generated essays indistinguishable from student work in blind grading (IJCAIE, 2024)
Basic coding (Python, Java syntax)	Far exceeds	Copilot passes CS1/CS2 courses, generates working solutions 65% of the time (Gonzaga, 2022)
Routine math computation	Exceeds	Claude scores 96.4% on MATH benchmark (Anthropic, 2024)
Foreign language translation	Matches or exceeds	GPT-4 achieves human-level BLEU scores on WMT benchmarks (Hendy et al., 2023)
Research compilation	Far exceeds	Our own operation: one agent produces 5+ sourced articles/day
Standardized test preparation	Exceeds	AI tutoring matches human tutors on test score gains (Khanmigo RCT, 2024)

The implications for curriculum are straightforward: spending 180 school days per year drilling skills that a free tool performs better is not education. It is babysitting with worksheets.

What AI Cannot Do (Yet)

The skills that survive the audit share a common trait: they require either physical embodiment, genuine novel insight, or interpersonal dynamics that language models cannot simulate.

Physical manipulation and fabrication. No AI can join wood, throw a pot, wire a circuit, or cook a meal. Robotics will eventually close this gap, but in April 2026, embodied skills remain fully human. Every hour spent in a woodshop, kitchen, or electronics lab teaches something no AI can replicate or shortcut. Boston Dynamics' Atlas can do a backflip. It cannot make a bookshelf.

Genuine novel scientific inquiry. AI can summarize papers, generate hypotheses, and even design experiments. What it cannot do is notice the anomalous data point that everyone else dismissed and decide to investigate it. The history of science is a history of productive stubbornness in the face of consensus. Fleming noticed mold killing bacteria. Prusiner spent a decade being ridiculed for prion theory. This kind of insight requires embodied pattern recognition against social pressure, which is not a capability language models possess.

Interpersonal negotiation and persuasion. Selling an idea in a room. Mediating a conflict between two people who both think they are right. Reading a face and knowing when to push and when to back off. These are skills that compound with practice and cannot be shortcut. Debate club and mock trial are suddenly among the most future-proof activities in a school.

Aesthetic judgment that is not derivative. AI generates competent art, music, and prose. It does not generate taste. Knowing that something is good when nobody else has validated it yet is a human skill. Curating, editing, and developing a point of view about quality in a world of infinite AI-generated content is not just valuable. By October 2026, it may be the most valuable skill in the economy.

Statistical literacy and epistemic calibration. Understanding base rates, conditional probability, selection bias, and the difference between correlation and causation. We demonstrated with our propaganda machine article how easy it is to produce content where every fact is real but every conclusion is wrong. The defense against that is not media literacy campaigns (decades of effort, no measurable impact). The defense is the mathematical intuition to recognize when true facts are being arranged to produce false impressions.

The Uncomfortable Truth About Coding Education

In 2016, President Obama announced the Computer Science for All initiative. By 2024, 57% of U.S. high schools offered at least one CS course (Code.org, 2024). Teaching kids Python and Java became the education establishment's consensus answer to "what should we teach for the future?"

That consensus aged catastrophically. GitHub Copilot now writes 46% of all code on the platform. In our blind model evaluation, the cheapest AI model (Haiku, $0.03 per task) performed at 96% of the quality of the most expensive ($0.76 per task). A student who spends a semester learning for-loop syntax is learning a skill with a shelf life measured in months.

What matters is not syntax. It is architecture: how to decompose a complex system into components, define interfaces between them, and reason about failure modes. It is systems thinking: understanding that changing one part of a system produces cascading effects in others. And it is the ability to evaluate whether AI-generated code actually does what you intended, which requires understanding the problem deeply enough to test the solution, not just the ability to type it.

The school that replaces "Introduction to Python" with "Systems Design and Failure Analysis" will produce graduates who remain relevant in 2030. The school that doubles down on syntax will produce graduates who are obsolete before they enroll in college.

What $65,000 Per Year Actually Buys

Our school analysis of the San Francisco Peninsula found tuition ranging from $32,430 (Notre Dame, Belmont) to $67,140 (Nueva) per year for private K-12. The total opportunity cost of a Nueva K-12 education, including tuition, forgone investment returns, and commuting risk, exceeds $3.28 million.

What does that buy? Not knowledge transfer. Knowledge transfer is now free. Claude will tutor your child in any subject, at any hour, with infinite patience, personalized to their level, for $20/month. Khanmigo is $44/year. The marginal cost of information delivery has collapsed to zero.

What elite private schools actually sell:

1. Small class sizes (6:1 at Castilleja, 8:1 at Nueva). This is the Bloom two-sigma effect: one-on-one tutoring produces two standard deviations of improvement. AI may deliver the information component, but it cannot yet deliver the accountability, social modeling, and motivational pressure of a human teacher who knows your name and notices when you check out.
2. Network effects. Your child's classmates at Menlo School or Castilleja will disproportionately become founders, executives, and decision-makers. This is not education. It is social capital acquisition, and it is brutally effective.
3. Credentialing. A diploma from Phillips Exeter or Lakeside signals something to college admissions that no amount of Khan Academy completion badges does. This premium will persist until elite colleges change their admissions criteria, which they have shown zero interest in doing.

Notice what is missing from that list: superior instruction in factual content. That was always a fiction, but now it is a measurable fiction. The best AP Chemistry teacher at a $65,000/year school is not five times better than an AI tutor at explaining electron orbitals. They might be 20% better. The gap does not justify the price. The network and the credential do.

The Curriculum That Should Exist

If we were designing K-12 from scratch in April 2026, knowing what we know about AI capabilities and trajectory, here is what the school day would look like:

Block	Subject	Why It Survives AI
Morning 1	Fabrication Lab (wood, metal, electronics, cooking rotation)	Physical manipulation; embodied problem-solving
Morning 2	Statistical Reasoning & Experimental Design	Defense against propaganda; epistemic calibration
Midday	Rhetoric, Debate & Negotiation	Persuasion; interpersonal influence; reading rooms
Afternoon 1	Systems Design & Failure Analysis	Architecture thinking, not syntax; applies to code, organizations, ecosystems
Afternoon 2	Creative Practice & Curation	Developing taste; aesthetic judgment; editing AI output
Daily	AI Collaboration Workshop	Using AI tools effectively; evaluating output; detecting hallucination

Grade-by-Grade: What This Actually Looks Like

The table above is a summary. The reality is more nuanced: you cannot teach CNC machining to a kindergartner, and you should not wait until 11th grade to start building things with your hands. Here is how the proposed curriculum flows across 13 years, with traditional subjects fading as AI renders them redundant and new subjects scaling as children develop capacity for them.

What Each Age Group Actually Does

What Dies When: The Obsolescence Timeline

Conspicuously absent: memorizing the dates of the French Revolution, diagramming sentences, solving systems of equations by hand, and conjugating Spanish verbs. Not because these are worthless. Because spending 500 hours per year on tasks an AI performs in seconds is an opportunity cost that compounds for 13 years.

Finland figured out half of this decades ago. Finnish students spend fewer hours in school than any OECD country, have no standardized tests until age 16, and consistently rank in the top tier of PISA scores. Their curriculum emphasizes project-based learning, cross-disciplinary thinking, and student autonomy. They also teach textile work, woodworking, and cooking as core subjects through grade 7. In 2026, this looks less like Scandinavian idealism and more like prescience.

The Six-Month Projection

By October 2026, based on current model improvement trajectories:

• AI tutoring will be indistinguishable from human tutoring for factual subjects. Khanmigo and its competitors will offer voice-based, adaptive tutoring that passes a pedagogical Turing test for math, science, and language arts.
• Code generation will be 80%+ autonomous for standard application development. The remaining 20% is architecture, specification, and review, exactly the skills we are advocating schools teach.
• The first school district will officially adopt AI tutoring as primary instruction for at least one subject. The Alpha School model ($65K/year tuition, 2 hours of AI-guided academics per day, rest is enrichment) will have its first year of outcomes data.
• The gap between "what schools teach" and "what AI can do" will widen, because model capabilities improve quarterly while curricula change on decade-long cycles.

The students graduating in 2030 are entering kindergarten or elementary school now. The curriculum they are learning was designed for an economy that will not exist when they enter the workforce.

Limitations

Our 72% automatability figure is approximate. We evaluated competencies against commercially available AI tools as of April 2026, using the standard expected of a high school senior (not an expert). The methodology has significant weaknesses: it does not account for the developmental benefits of struggle (learning math by hand may build cognitive patterns that AI-assisted math does not), it treats all skill transfer as zero (memorizing history may build general knowledge retrieval abilities useful in non-automatable contexts), and it ignores the socialization function of school entirely. Schools are not just knowledge-transfer institutions; they are where children learn to exist in groups, navigate hierarchies, manage conflict, and build relationships. No AI audit captures that.

Additionally, projecting six months forward in AI capabilities is more reliable than projecting ten years, but it is still projection. Regulatory intervention, compute bottlenecks, or unexpected capability plateaus could slow the trajectory. We assign roughly 70% confidence to the October 2026 projections above.

The Strongest Counterargument

The most serious objection to this analysis is that we are confusing "AI can perform the task" with "AI should replace the learning of the task." Learning to write a five-paragraph essay is not (only) about producing essays. It is about learning to organize thoughts, construct arguments, and communicate clearly. The essay is the exercise; the cognitive skill it builds is the product. By this logic, the fact that AI writes better essays is irrelevant, the same way the existence of calculators does not make learning arithmetic useless.

This is a strong argument, and we do not fully dismiss it. The question is one of degree and opportunity cost. If writing essays builds cognitive organization skills, does writing the 500th essay build more than the 50th? At what point does the marginal cognitive return of another formulaic essay approach zero? And when it does, could those hours be spent on something with higher marginal returns, like learning to evaluate whether an AI-generated essay is actually making a coherent argument?

We suspect the honest answer is that the first 50 essays matter enormously and the next 450 are inertia. The same logic applies to math drill, vocab memorization, and lab report writing. The skill-building phase is real and valuable. The repetition phase is where automation should take over, freeing time for skills AI cannot teach.

What You Can Do

If you're a parent:

• Supplement school with fabrication skills. A $200 Arduino starter kit, a weekend woodworking class, or regular cooking together teaches more future-proof skills than a $5,000 SAT prep course.
• Teach your kids to evaluate AI output. Have them fact-check Claude's answers. Show them our propaganda article and ask them to spot which conclusions don't follow from the facts.
• When evaluating schools, ask what percentage of class time is spent on skills that AI can already perform. If the answer is above 50% and the school has no plan to change, the school is on a decade-long lag.
• If you're paying $65K/year, be honest about what you're buying. If it's the network and credential, that's a legitimate purchase. If you think you're buying superior knowledge transfer, you're paying a 3,000% markup over AI tutoring for a product that is at best 20% better.

If you're a school administrator:

• Audit your curriculum against current AI capabilities. The BLS automation risk data provides a starting framework. If a subject's primary learning outcome is a task AI performs better, it needs restructuring.
• Introduce AI collaboration as a skill, not a prohibited technology. Students who learn to use AI effectively in school will outperform those who learn to use it furtively.
• Study the Finland model. Not because Finland is perfect (their PISA scores have declined since 2006), but because their emphasis on project-based learning, cross-disciplinary thinking, and hands-on fabrication is exactly what survives the AI audit.

If you're a school board member:

• Request an AI capability audit of your district's curriculum. If nobody on staff can perform one, that itself tells you something about readiness.
• Benchmark against Alpha School's outcomes data when it becomes available in late 2026. If a school running 2 hours of AI-guided academics produces comparable test scores to schools running 6 hours of traditional instruction, the implications for staffing and budget allocation are enormous.

The Bottom Line

The K-12 curriculum was designed to produce factory workers in 1893 (the Committee of Ten), tweaked to produce office workers in 1983 (A Nation at Risk), and never redesigned for a world where both factory work and office work are being automated. The subjects that survive AI are the ones the education establishment has spent decades defunding: shop class, home economics, art, debate, and statistics. The subjects that don't survive are the ones that consume the majority of instructional time: memorization-heavy content delivery, formulaic writing, and procedural math.

This is not a prediction about 2035. This is an observation about April 2026. The tools exist today. The curriculum changes haven't started. Every year of delay is another cohort of students trained for jobs that will not exist when they graduate.

The clock is not ticking. It has already rung.

Related Articles

• California Spends $23,519 Per Student. Two-Thirds Can't Do Math.
• We Accidentally Built a Propaganda Machine
• Democracy Was Designed for a World That No Longer Exists