A personal knowledge base on generative AI and its impact on K–12 education, with a specific focus on Career and Technical Education (CTE). Built and maintained using Andrej Karpathy's LLM Wiki pattern.
Source documents go into raw/. Claude reads them, discusses key takeaways, then writes and cross-references markdown pages in wiki/ — source summaries, concept pages, comparisons. A single paper typically touches 10–15 pages. The wiki compounds: every new source propagates updates across existing pages, contradictions get flagged, and cross-references already exist when you need them. Querying the wiki draws on curated synthesis rather than re-assembling raw PDFs from scratch.
The human curates sources and asks questions. Claude handles the work that causes humans to abandon wikis: updating cross-references, maintaining consistency, and synthesizing across sources. Start with wiki/index.md for the full table of contents.
Fifty-one sources ingested so far, producing 69 interlinked wiki pages. The picture that emerges is more nuanced — and more actionable — than either the AI-will-take-all-jobs or AI-changes-nothing narratives.
AI completion is not learning. Yan & Gašević's Agentivism framework (2026) and the MIT EEG study (Kosmyna et al.) converge on a hard finding: when AI completes tasks for students before they've built independent capability, measurable neurological harm accumulates over time — up to 55% reduction in neural connectivity. The mechanism is cognitive debt: the brain stops practicing the circuits that need exercise. Task performance goes up; durable understanding goes down.
Students cannot evaluate AI quality — and enthusiasm makes it worse. Abdelghani et al. (2026) studied 63 middle-schoolers doing science tasks with ChatGPT. Prompt quality discrimination: at chance. Answer quality evaluation: at chance. 71.4% of expert-rated low-quality answers were rated "useful" by students. The only protective factor was metacognition — not domain knowledge, not AI familiarity. Positive AI attitudes negatively predicted interaction quality (β = −0.39). The students who felt most comfortable with AI were the most blind to when it was failing them.
AI tutoring has barely worked, and it may be the wrong tool for high schoolers. Khanmigo underperformed its adoption targets because students who most need help are least likely to seek it out. A 2026 meta-analysis of 34 experimental studies (Liu et al.) found near-zero cognitive benefit for grades 10–12 (g = 0.037) while upper-primary students benefited substantially (g = 0.877). The core CTE population is in the grade range where AI educational interventions show the weakest evidence.
The Khanmigo "non-event" now has field-RCT causal confirmation, at the largest scale yet attempted. Oreopoulos & Low's NBER working paper (August 2026) reports a two-year, 18-middle-school randomized trial: Khanmigo raised math achievement by 0.06–0.08 SD/year, but this is statistically indistinguishable from Khan Academy's effect without any AI tutor. Message-level records show why — 96% of students tried Khanmigo once, but the median student engaged it substantively (asking a real question or reasoning step, rather than a bare answer or button-click) in only about 1 message in 7. The paper's own conclusion is nearly a direct quote of Agentivism's thesis: "making AI effective for learning appears to be as much a behavioral challenge as a technological one."
The clearest positive evidence is for teachers, not students. Educator-facing AI tools — especially real-time coaching during lessons — show the most consistent causal benefits, with the largest gains for novice teachers. This is where AI investment in schools is best supported by evidence.
Scattered AI use in learning is worse than no AI — but strategic delegation produces transformative learning. Wang & Zhang (2026), 912 students across China, Europe, and the US, find a U-shaped relationship. Zone 2 (scattered, half-hearted AI assists — the modal deployment condition) adds coordination overhead without freeing cognitive capacity and produces worse outcomes than no AI at all. Zone 3 (committed delegation of entire task categories, freeing capacity for higher-order reflection) produces deep transformative learning — students questioning assumptions, shifting frameworks, restructuring understanding. Crucially, partnership orientation activates both critical vigilance and strategic delegation simultaneously. This contextualizes all the major negative findings in the AI learning literature (Kosmyna, Bastani, Fischer): they were measuring Zone 2. The design question is not "restrict AI or allow it" — it is "Zone 2 or Zone 3." The answer is structural: learner-first sequencing, whole-category delegation, verification built into workflow, and unassisted assessment. Xu et al. (2026) confirm across 35 studies that AI-assisted performance (g=0.751) and independent performance (g=0.369) are different things — assessing without scaffolding is non-optional.
AI boosts production speed dramatically but converts only marginally to shipped output. Demirer, Musolff & Yang (2026), tracking 100,000+ GitHub developers across three AI tool generations, find a consistent productivity funnel: +740% code produced → +65% review requests → only ~20% more shipped releases. Every human-dependent review stage absorbs the gain before it converts to finished output. At the organizational level: ~95% of companies see no meaningful return from GenAI (MIT NANDA, 2025); 80% of AI projects fail — twice the rate of comparable non-AI projects (RAND, 2025). The bottleneck is not AI capability; it is human review and integration capacity. The same funnel operates in education: AI dramatically raises output volume while durable learning gains drain through reading, retrieval, and transfer. The strategic implication — "enter high in the chain" — means AI aimed at feedback, metacognitive calibration, and teacher coaching (where cognitive consolidation lives) outperforms AI aimed at content generation. The same technology, aimed at a different stage, produces the opposite result.
Oral defense is the answer to both cognitive debt and AI detection. Murgatroyd (2026) and Nurenberg (2026) independently converge on the same classroom solution: design assessment so that understanding must be demonstrated in real time, rather than trying to detect AI-generated artifacts after the fact. Murgatroyd's "Build, Analyze, Defend" scaffold — produce something with AI, interrogate it critically, then defend it orally — maps directly onto the Agentivism framework and makes AI detection irrelevant. AI detection software does not work reliably; institutions relying on it are now losing court cases filed by students falsely accused of academic misconduct.
Human judgment may be a categorically different cognitive strategy from LLM prediction, not a slower version of it. Stanford GSB research (Guilbeault, Caplan & Yang, 2026, PNAS) on "satisficing" — reaching confident conclusions from deliberately minimal data via intuitive leaps — and the Tolerance Principle offers the wiki's first mechanism-level cognitive-science account of why transferable skills and reconstructive internalization matter, distinct from labor-economics or classroom-outcome evidence. If human judgment is a different strategy suited to sparse, ambiguous conditions rather than an inferior, lower-data version of AI prediction, it does not necessarily lose value as AI's data access and compute scale up.
Only 12% of US jobs face true substitution. BCG Henderson Institute's 2026 analysis of 165 million US jobs across 1,500 occupational roles finds only 12% (the "Substituted" segment) face genuine headcount contraction — roles where AI replaces the core function and demand is bounded. The widely-cited Frey & Osborne figure of 47% at risk ignored demand expandability: when AI lowers the cost of a service, demand may expand to absorb the productivity gain (Jevons Paradox). Software engineering is the canonical case — coding AI doubled output per engineer, software got cheaper, and headcount rose after ChatGPT.
The disruption is hitting educated white-collar workers first, not CTE-track workers. Anthropic's observed-exposure data shows that the most AI-exposed workers are on average 47% higher-paid and four times more likely to hold graduate degrees than zero-exposure workers. BCG's "Limited-Exposure" segment (34% of US jobs) — defined by physical presence, sustained human interaction, and contextual judgment in variable settings — maps almost exactly onto CTE pipelines: healthcare, skilled trades, early childhood education, construction, personal services.
The entry-level hiring decline is real — but its attribution to AI is now contested. Lambert & Schindler (Warwick/LSE, May 2026) find that work-from-home and GenAI exposure are correlated at 0.77 (Spearman rank) across occupations — the same workers (software developers, management consultants) top both rankings. In joint-treatment difference-in-differences specifications using 243 million new hires across four countries, the WFH coefficient holds stable (−1.42 to −1.57pp) while GenAI attenuates to near-zero and is often statistically insignificant. The 8–11 percentage-point decline in junior-hire share below 2019 baselines may be primarily an organizational friction from remote work — surmountable through better management practice — rather than irreversible task automation. The diagnosis changes the remedy.
Automation displaces; augmentation does not — the mechanism-level finding the debate was missing. Stanford DEL's Canaries Dashboard (Brynjolfsson et al., June 2026) tracks 4.6 million workers in live ADP payroll data. Using the Anthropic Economic Index to decompose AI usage type: occupations with higher automation usage show clear negative correlation with early-career employment; occupations with higher augmentation usage show no relationship. Early-career workers in the most AI-exposed occupations are contracting at −3.8%/year since ChatGPT while the least-exposed are growing at +2.0%/year. TFP growth is neutral — no decisive macroeconomic takeoff yet — but the occupation-level early-career signal is real and compounding.
The academic paper behind that dashboard adds a direct challenge to the WFH-attribution reading. Brynjolfsson, Chandar & Chen's underlying working paper (Stanford/NBER, November 2025) finds a 16% relative employment decline for 22-25 year-olds in the most AI-exposed occupations, robust to firm-time fixed effects and a battery of alternative explanations — critically including a robustness check restricted to non-teleworkable occupations (bank tellers, travel agents, tax preparers), where the decline still holds. This complicates Lambert & Schindler's WFH-attribution argument above: if remote-work friction were the primary driver, the effect should not appear in jobs where remote work was never possible. The paper also finds employment for young health aides — a relational-sector occupation — growing faster than for older workers, a direct real-data confirmation of the relational economy thesis discussed below, and proposes a labor-economics mechanism (AI substitutes for codified "book-learning" knowledge more readily than for tacit, experience-accumulated judgment) that complements the satisficing/Tolerance Principle cognitive-science account above.
AI diffusion is constrained by institutions, not technology — and adoption is still well below half. Brad Smith (Microsoft President, June 2026) cites Narayanan & Kapoor (Knight Columbia): "diffusion is limited by the speed of human, organizational, and institutional change." Microsoft's own AI Diffusion Report (Q1 2026) puts current usage at 17.8% globally and 31.3% in the US — less than one in three American workers. The 2026 graduating class has internalized AI's impact differently than any prior cohort: they are simultaneously the highest AI adopters (college-town counties lead US adoption) and the most vocally resistant (booing AI at commencements; Princeton students wore jackets labeled "100% cotton — 100% human"). Smith's explanation for the junior-hiring decline: not just AI automation, but a "perfect storm" — AI task automation + corporate headcount reductions to fund AI capital expenditures + COVID-era hiring overhang + geopolitical uncertainty. The multi-causal diagnosis matters because different drivers require different policy remedies.
The relational economy thesis has structural backing. Alex Imas (UChicago, 2026) argues — with econometric support — that as AI commodifies cognitive production, rising real incomes shift spending toward goods and services where human involvement is inseparable from the value: care, hospitality, artisanal production, coaching, education. Historical structural change data (Comin et al., Econometrica 2021) shows income effects account for over 75% of employment reallocation patterns. The sectors CTE prepares are not merely "safe from automation" — they are the sectors that structurally expand as cognitive commodities get cheap.
Economists themselves cannot agree on AI's current labor-market direction — and enterprise AI cost is now a real adoption constraint. The New York Times (Casselman, July 2026) surveys why AI's real-time economic effect is so hard to measure: exposure-index methodologies can flip the sign of AI's estimated employment effect depending on which measure is used, federal occupational data lags by over a year, and private data sources actively conflict (ADP data shows entry-level declines in AI-exposed sectors; Ramp/Revelio data shows heavy AI adopters hiring faster). Separately, leaked internal communications reported by 404 Media (Cox, July 2026) show Atlassian, Citi, Adobe, and Amazon actively rationing AI access as costs spiral — Atlassian's AI spend tripled from $5M to $15M/month in nine months — complicating any assumption that AI adoption proceeds frictionlessly once an organization decides to adopt.
Big Tech is explicitly investing in CTE trades training. In June 2026, Zuckerberg announced Meta's $115 million "America's Workforce Academy" — free data center construction training with NCCER credentials — stating: "America is going to need hundreds of thousands of skilled tradespeople to build the infrastructure for our country to lead in AI." The predecessor program (Level-Up) received 35,000 applications in its first week. Meta Texas data center: 1,800 peak construction workers → ~100 permanent operational roles. This is the highest-profile industry confirmation in the wiki that the AI infrastructure → CTE demand claim is real, named, and funded.
Students in the AI capital are already choosing trades as an AI hedge. Bay Area high school seniors in the class of 2026 (NYT, June 2026) are explicitly avoiding tech careers because of AI and pivoting to construction and trades — articulating the physical-presence insulation argument independently: "A.I. is not going to build a home. A.I. isn't going to weld anything either." This behavioral shift is occurring in the place where AI's disruption of professional work is most visible and the evidence for it is most accessible to students. The prediction that CTE-aligned sectors are protected is showing up in enrollment decisions.
CTE programs create the choice set — they are discovery mechanisms, not pipelines. Over 95% of students in California's construction programs had never used tools before the class. One student: "I had no idea this class existed. I didn't even really consider construction seriously until I took this class." The obstacle to scaling the trade pivot is not student interest — it's exposure. State investment in CTE programs is the structural precondition; family cultural pressure (especially in immigrant families, who associate four-year degrees with social mobility) is the binding constraint once interest exists.
CTE-track occupations are doubly insulated from both post-pandemic shocks. Electricians, construction workers, and other CTE-aligned workers sit at the bottom of both WFH and GenAI exposure rankings. Whether the junior-hiring decline ultimately proves to be AI-driven or WFH-driven — the attribution is now contested — CTE-track workers face neither pressure. The protected-sector thesis holds regardless of which mechanism wins the empirical debate.
AI exposure inside CTE now has occupation-level data, not just general impressions. A four-report series from the Education Research & Opportunity Center, Advance CTE, and ACTE (March–August 2025) applies the AI Occupational Exposure Index to four of the six modernized Career Cluster Groupings. Building & Moving (38.9M workers, 26% of the US workforce) is the least AI-exposed; Caring for Communities (education, healthcare, public safety) is the most exposed; Cultivating Resources and Creating & Experiencing sit near the national average. The same pattern holds inside every single grouping regardless of its average: bachelor's-degree occupations score well above the national mean, high-school-or-less occupations score well below it — the labor market polarization thesis reproduced at the occupation level, inside CTE specifically. One report includes a sharply dissenting voice: Jesse Anglen (CEO, Rapid Innovations) predicts AI will displace "close to a hundred percent of knowledge workers" within five years, directly contradicting the wiki's dominant, more moderate labor-market consensus — noted here per the wiki's rule to flag contradictions explicitly rather than resolve them.
The occupation-level exposure map now covers all six Career Cluster Groupings — but two of them only qualitatively. Advance CTE's Responsive by Design (August 2026) extends illustrative AI-use examples to Financial Services, Management & Entrepreneurship, Marketing & Sales, and Digital Technology — the wiki's first content on any of these — plus a four-part state implementation strategy (statewide AI guidance, educator AI-literacy PD, industry partnerships, cross-sector conversations). Unlike the four-report exposure-index series above, these two groupings have no AIOE scores yet, leaving a real gap between measured and merely illustrated clusters. The report also surfaces an unresolved inconsistency inside the wiki's own source base: it groups Digital Technology as an ordinary cluster within a "Connecting & Supporting Success" grouping and splits Financial Services into its own "Investing in the Future" grouping, while a February 2026 CTE Futures webinar described Digital Technology as a crosscutting cluster within a "Managing & Connecting" grouping instead — flagged explicitly rather than smoothed over, since it's unclear whether the national framework changed again or the two sources are just describing it at different levels of detail.
The Applied Co-Intelligence model now has its first real-world classroom case studies — and a natural contrast with the Khanmigo RCT above. CTE Futures (August 2026) documents three Tennessee programs implementing the ACI model: Walters State Community College (agriculture/soil-science) reports AI-literacy gains without sacrificing subject-matter learning; TCAP Upper Cumberland (welding) reports AI-driven cross-departmental collaboration as a side effect of adoption; Jackson State Community College documents an institution-wide rollout culminating in a public AI Summit for 200+ regional stakeholders. None uses a comparison-group design, so these are implementation documentation rather than causal evidence — but they share a structural feature the Khanmigo RCT's null result lacked: AI engagement built directly into the assessed occupational task, rather than offered as an optional resource requiring student-initiated help-seeking.
The entry-level AI-exposure penalty is now internationally replicated, and the college wage-security premium has statistically inverted. Goldman Sachs (August 2026) finds the same automation-heavy, entry-level-concentrated hiring headwind across the US, Canada, Germany, France, Australia, and the UK — call centers are 27-39% below trend across these markets, and entry-level workers bear 2-6x the AI-exposure hiring drag of the broader workforce — closing part of the external-validity gap in the wiki's previously US-only ADP/Anthropic evidence. Separately, the Federal Reserve Bank of New York's Q2 2026 data (via Forbes) shows recent-graduate unemployment at 5.6% against a 4.2% overall workforce rate — the first time the college degree has been a liability rather than an asset for job security in the wiki's evidence base. And Pew Research's multi-year tracking confirms the underlying anxiety is real and specifically age-concentrated: adults under 30 are the only age group whose AI-related concern is still climbing (31% to 55% between 2021 and 2026), independently corroborating the trade-school pivot and major-switching behavior documented elsewhere in the wiki.
The K-12 AI policy gap is structural, not accidental. Stanford's AI Index 2026 finds that 80% of university students use AI for schoolwork — doubled from 40% in 2023 — but only 6% of teachers report their school's AI policies are clear. No US state has established teacher training standards for AI education. AP CS excludes AI content. State guidance is largely nonbinding. Students are integrating AI into their work at scale, without any institutional literacy framework around it.
Congress has now heard, and largely confirmed, the wiki's core evidence gap — while surfacing concrete state-level policy models. At a June 2026 Senate HELP subcommittee hearing, InnovateEDU CEO Erin Mote testified there are "currently no high-quality causal studies on the long-term effects of A.I. on student learning, equity, or social emotional development" — independently corroborating Stanford SCALE's systematic-review finding via congressional testimony rather than academic synthesis. The hearing also surfaced two concrete, operating state models: Delaware's AI Assurance Lab (independent pre-adoption testing of AI tools rather than trusting vendor claims) and Alabama's Data Scholars program (100+ paid high-school AI/data internships, curriculum refresh cycles redesigned for AI's pace of change rather than the traditional five-to-ten-year textbook cycle). The hearing also disclosed that the PowerSchool breach — 62 million students' and 9.5 million teachers' data exposed — is the largest breach of children's education data in US history.
The key design variable is whether AI provides answers or builds thinking. The evidence base consistently shows that AI tutors designed to give hints and ask guiding questions produce near-equivalent outcomes to traditional instruction; AI that provides complete answers harms independent learning and widens achievement gaps for low-prior-knowledge students. The same technology, with different pedagogical design, produces opposite effects.
The transformation starts with teachers, not tools. Microsoft Elevate's president Justin Spelhaug — who leads a $4B AI education program — cites One Laptop Per Child as the warning: many laptops "became doorstops within six months." His verdict: "Providing a tool to a student does nothing." Microsoft credentialed 2 million teachers in its first year through Elevate for Educators. The largest employer-led AI training programs converge on a teacher-first stance that the systematic research evidence supports: the most consistent positive causal effects in AI education are on the educator side, not the student side. Spelhaug names "navigate yourself" — metacognitive adaptability, not any specific tool — as the durable skill worth building. Gallup/Lumina 2026 (3,801 US college students) finds 47% have seriously considered changing their major because of AI's career impact; 16% already have.
AI equity runs deeper than an infrastructure gap — it is also a cultural representation problem. AI systems trained predominantly on Western, English-language text produce outputs that systematically misrepresent or erase indigenous and non-Western knowledge systems. In Canada (among the world's wealthiest countries), 650,000 households lack internet access — primarily indigenous and northern rural communities — while Canada ranks among the five most expensive jurisdictions globally for broadband. The answer to both problems points in the same direction: locally deployed AI on locally controlled data (data sovereignty), not scaled access to centralized Western cloud systems.
Business education is independently reaching the same conclusions as K-12 research. CEMS — a global alliance of 33 business schools and 70+ corporate partners — surveyed leadership and AI-education experts in 2026 and found, from the employer/graduate-education side, the same core findings the wiki has documented from the K-12 side: cognitive offloading erodes critical thinking, AI's fluent confident output style suppresses the evaluative scrutiny needed to catch bad answers, and "copilot not autopilot" is now the dominant framing for responsible AI use across sectors. Three separate institutions in the report — plus two existing wiki sources — independently converged on oral or process-based assessment as the practical response to AI-assisted work, without any apparent coordination between them.
raw/ ← Source documents (immutable — never modified)
wiki/ ← LLM-maintained markdown pages
index.md ← Full table of contents
log.md ← Append-only record of all ingestions and changes
CLAUDE.md ← LLM instructions: page format, ingest workflow, citation rules
Source documents (raw/) are not included in this repository. They contain copyrighted papers, articles, and reports that the wiki pages summarize and synthesize.
- Andrej Karpathy's LLM Wiki gist — the original pattern this wiki is based on
- What Is Andrej Karpathy's LLM Wiki? (MindStudio) — walkthrough of the pattern with Claude Code