Ignited Labs
Research Report · As of September 28, 2026

AI and Schools Around the World

A global survey of AI-based schools, independent research evidence, and national programs. Bottom line: AI is gradually breaking down school roles, but not replacing the institution itself - and there's a wide gap between vendor claims and independent evidence.

20 of 800+high-quality causal studies Stanford found on AI in K-12
$2.4Bglobal edtech VC funding, 2024 - the lowest in a decade
70,000teachers in AI training in Israel
1.2 trillion wonKorea's spend on AI textbooks, scrapped after one semester
86%of children aged 9-17 use generative AI
01 — Bottom Line

AI breaks down roles, it doesn't replace the school

As of late 2026, nothing has shown that AI can replace the school as an institution. The strongest independent evidence says AI helps when there's a structure that gets students to use it properly. When students use it on their own, or use general chatbots that just give answers, the gains are small, short-lived, or negative.

What AI does do is take apart the school's roles one by one. Content delivery, practice and feedback, and lesson prep by teachers are being automated fastest. Motivation, supervision, socialization and certification still belong to the school. — The bottom line of this report
02 — Executive Summary

Three things worth understanding before you go on

AI-based schools exist - but their evidence is weak

It comes mostly from the vendors themselves. Alpha School's claim of "2.6x growth" has been taken apart by independent analysts. Its non-selective public school, Unbound Academy in Arizona, reached 28% proficiency in English and 10% in math in its first year, after promising the state 65% and 60%. The Sabrewing program in London started with about seven students and hasn't published results.

The serious research: tool design and student engagement are decisive

A Stanford review found only 20 high-quality causal studies out of more than 800 papers. A two-year trial of Khanmigo found small gains of 0.06 to 0.08 standard deviations - no better than Khan Academy without AI - because students barely used the tutor. Unrestricted access to GPT-4 lowered test scores by 17% after access was taken away. The biggest gains come from structured or human-supervised setups: Nigeria with 0.31 standard deviations, Tutor CoPilot with 4 percentage points, and LearnLM with 5.5 percentage points.

Governments are pushing AI literacy, but retreating from AI that replaces teaching

The South Korean government spent more than 1.2 trillion won (about $850 million) on AI textbooks, and after one semester reclassified them as "supplementary material." China bans elementary students from using open generative AI on their own. The UAE, Estonia, India and Israel are expanding structured programs. The gaps in the startup market are in motivation, orchestrating peer learning, certification, and audio-based retrieval practice - not in yet another chat tutor or teacher assistant.

03 — Key Findings

The numbers behind the story

The gap between claims and evidence

2.6xAlpha School's growth claim on MAP tests - not independently verifiedAlpha School announcement
28% / 10%Actual proficiency in English / math at Unbound Academy, year one - vs. a promised 65% / 60%Kelsey Piper reporting, The Argument
20 of 800+high-quality causal studies on AI in K-12, out of the entire reviewed databaseStanford SCALE, March 2026

Controlled trials and funding

0.06-0.08 SDKhanmigo's per-year gains in a two-year cluster trial - identical to Khan Academy without AINBER 35620
0.31 SDLearning gains in Nigeria, with facilitated after-school sessions - 1.5 to 2 years of schoolingWorld Bank, Edo State
$2.4BGlobal edtech VC funding, 2024 - 89% below the 2021 peak, the lowest since 2014HolonIQ

Six key findings emerge from the research:

  1. The gap between claims and evidence is large. The most famous model, Alpha School and 2 Hour Learning, has not released raw data to independent reviewers. Its only non-selective public test, Unbound Academy, missed its own proficiency targets by a wide margin.
  2. The bottleneck is engagement, not model capability. In the largest and longest LLM tutor trial (Oreopoulos and Low, NBER 35620), 96% of students tried Khanmigo. But the median student messaged it on only a third of practice days, and in only 17% of practice sessions where they made a mistake. Access doesn't turn into learning.
  3. Guardrails matter. Tutors that give hints and scaffolding preserve or improve learning. Chatbots that give answers raise practice scores but can hurt learning once the tool is taken away.
  4. AI that serves teachers has the lowest risk and the best evidence for time savings, but the savings are smaller than reported, and the link to learning hasn't been shown. In a Gallup and Walton survey (March-April 2025, 2,232 respondents), teachers who use AI weekly (32%) estimated they save about 5.9 hours a week, but that's self-reported. In a UK EEF controlled trial (259 teachers, 68 schools), the time saved on lesson planning was 25.3 minutes a week (31%), with no meaningful difference in quality. AI-generated lesson plans tend toward lower-order thinking (90% of activities in a study of 311 plans). The highest value likely comes from tools that improve the teaching itself: Tutor CoPilot improved student mastery by 4 percentage points at a cost of about $20 per teacher per year, and automated feedback (M-Powering Teachers) improved teachers' engagement with student ideas by 13%.
  5. National programs are converging on the same pattern. Most teach AI literacy, restrict unsupervised use by young children, keep the teacher accountable, and roll out in stages. Blanket mandates for AI-delivered instruction, like Korea's, have failed.
  6. Capital is scarce and concentrated. According to HolonIQ, global edtech VC funding stood at $2.4 billion in 2024, 89% below the 2021 peak and the lowest level since 2014. AI money goes mostly to teacher assistants, tutors, and language learning. Few funded players remain in peer learning, certification, generative worlds, and audio-based learning.
04 — AI-Based Schools That Made a Dramatic Change

No model has shown independently verified results at scale

Alpha School / 2 Hour Learning / TimeBack (US)

What they did. A private network of K-12 classrooms founded in Austin in 2014 by MacKenzie Price and Brian Holtz. Billionaire Joe Liemandt (Trilogy/ESW Capital) is its director and financial backer. Students spend about two hours each morning on adaptive academic software, now packaged in Alpha's platform called TimeBack, including apps like Alpha Read and Alpha Write. Afternoons are devoted to workshops, life skills, and projects. The adults are called "guides," not teachers. Liemandt says students spend an average of 121 minutes a day on the software and finish a grade level's subject in 20 to 30 hours, citing Bloom's "two sigma" research on one-on-one tutoring. Much of the "AI" is adaptive learning software built on statistical models that predate the LLM wave, with new AI tutor layers on top. It isn't primarily a chatbot.

Scale and cost. As of April 2026, Wikipedia counted 13 campuses, including Austin, Brownsville, Miami, Palm Beach, New York, San Francisco, Santa Barbara, Lake Forest, and Scottsdale. More campuses are planned for 2026-27. Tuition ranges from $10,000 in Brownsville to $75,000 in San Francisco and Palo Alto. There are sub-brands: Texas Sports Academy, NextGen Academy for gamers, Founders School, and Montessorium. The online options are Alpha Anywhere, marketed as "learning 2x faster," and GT Anywhere, marketed as "learning 3 to 7x faster." US Education Secretary Linda McMahon has publicly supported the model.

Selection. The student population is largely self-selected, and at most campuses it's affluent. Analysts note that when most of the difference between schools comes from family self-selection, a $75,000 school would outperform even with ordinary teaching.

The claims (from the vendor)

  • Students "grow 2.6x faster than their peers" on national MAP tests
  • "Most" students are in the 99th percentile
  • The best reach "up to 6.5x growth"

Independent criticism

  • Kelsey Piper (The Argument, August 25, 2026): "2x growth" divided by an expected median gain that in high school is near zero, so random noise registers as "9x growth." An Alpha parent, Peter Naimoli, calls the 2.6 figure "a miscalculated and meaningless number." The gifted-student figure was reportedly based on five kids with two tests each
  • Alpha gives large incentives for test performance, up to prizes like a Nintendo Switch or an iPad - critics say this "teaches to the MAP test" and inflates the appearance of real learning
  • Piper examined the TimeBack apps and found most of them unimpressive (for instance, AI-generated Harry Potter-style fan fiction); middle schoolers mostly don't get books to read. Her conclusion: the effect comes from the incentives and the environment, not the software
  • Robert Pondiscio, Jared Cooney Horvath, and Dylan Kane: the data, methods, and selection have never been independently examined
  • Alpha refused to give data to WIRED and the Boston Globe; raw data promised for May 2026 still hadn't arrived by August 2026

Additional controversies are reported, from secondary sources: a WIRED investigation into students who left years behind, IXL closing Alpha's account in July 2025, families leaving in Brownsville, and concerns about surveillance and student wellbeing. Wikipedia also notes scrutiny of the ties between the school and for-profit vendors.

Charter rejections and the voucher track. Unbound Academic Institute, tied to Alpha's founders, filed applications for online charters in Pennsylvania, Arizona, North Carolina, Arkansas, and Utah. Only Arizona approved one. Pennsylvania rejected its application in 2025, calling the model "untested" and unproven to meet state standards. The application there proposed a staff of 17 for 500 students with no physical building, and payment to 2 Hour Learning of $5,500 per student, versus $2,000 to $2,500 in applications in other states. In Arizona, the sister private school Novatio can be fully funded through the state's education savings account voucher.

Current status (September 2026). Alpha keeps growing despite heavy negative press. Parents often say they know the numbers are inflated and still value the school anyway, thanks to the peer group, the pace, and the afternoon activities. Commentators tied to the Clayton Christensen Institute describe it as a disruptive niche model: "If you don't like Alpha, it's probably not for you."

Two more cases
Unbound Academy (Arizona)
The closest thing to a natural experiment
  • The model: a free, online public charter school for grades 4-8, and the first public school in Arizona marketed as having "AI teachers." Opened fall 2025, running the 2 Hour Learning platform. Admission requires only Arizona residency, "no grades, transcripts, or essays." Together with Novatio it served about 250 students
  • Year-one results (August 2026): promised 65% proficiency in English and 60% in math (versus national averages of 42% and 34%); actual results were 28% and 10%. "1.8x growth" on MAP. Piper: "a failure by Alpha's own terms"
  • The framing shifted: from chatbot characters ("Phoebe and Philip") to a website emphasizing "certified teachers live," and from "1.8x growth" to "2.8x" in marketing - a quiet retreat from the "AI replaces teachers" message
  • The lesson: without an incentive environment and a selected peer group, the software alone looks like ordinary edtech
David Game College, Sabrewing Program (London)
Operating but small, with no evidence either way
  • The model: launched September 2024, described as the UK's first GCSE class "without teachers." Students attend in person; mornings are devoted to adaptive AI platforms (VR has also been reported), afternoons to life skills. Groups of up to 20 students with three "learning coaches"
  • Actual uptake: AFP reported in January 2025 that the pilot had seven students, far fewer than the planned 20
  • The development: in July 2025 it expanded "to all GCSE and A Level students" with a personal AI tutor named "Violet." Tuition is £27,000 (UK students) up to £35,000 (international); admissions cycle for September 2026
  • Evidence: no outcome data has been published, only promotional quotes like "students are thriving"
Others and the broader pattern
The author's assessment

Searches for AI-based schools in Asia, Africa, and Latin America turned up mostly national curriculum programs, not school redesigns. The notable exceptions are systemic efforts (see Part 3). The most significant change in a public system is Israel's shift to AI-personalized English learning in grades 7-8, in response to a teacher shortage.

In 2026, "AI school" mostly means a morning of adaptive software and an afternoon run by people, usually in an expensive private or micro-school format.

05 — What the Research Actually Shows

The best results come when a human and structure are in the loop

No rigorous study has yet shown AI replacing a full year of schooling with durable gains.

Stanford SCALE: Evidence Review on AI in K-12, 2026

March 2026; Pessler, Martinez-Kleiss, Agnew, Loeb. More than 800 papers in the AI Hub database were reviewed, as of October 2025 (the database has since passed 1,100). Only 20 were high-quality causal studies, and none of them examined student use of AI in US K-12 classrooms. Findings: performance mostly improves with the tool, but results are "mixed" once it's taken away; tools with pedagogical guardrails outperform general chatbots; teacher-facing tools cut prep time without hurting quality; most studies are short-term, and few examine equity, wellbeing, or social development.

Stanford SCALE
Brookings: "A New Course for Students in an AI World"

January 14, 2026. A year-long "pre-mortem" covering more than 50 countries, more than 500 interviews and focus groups, more than 400 studies, and a Delphi panel. Conclusion: "At the current stage, the risks of using AI in children's education outweigh the benefits" (56% of inputs addressed harms, 44% benefits). The central risk: a "doom loop" of offloading thinking to the tool. Caveat: an NEPC review by Prof. William Penuel (University of Colorado Boulder) found the report good at anticipating harms but weak as an action plan.

Brookings Institution

Key trials

StudyContext & durationResultCaveats
Kestin et al., Harvard physics
Scientific Reports, 2025
Undergraduate physics, a research-based pedagogy AI tutor vs. active learning in classThe tutor beat active learning; learning gains of more than 2xCollege students, short units, one course; not K-12
De Simone et al., World Bank, Nigeria
Edo State
6 weeks, June-July 2024; 12 after-school sessions of 90 minutes; GPT-4 via Microsoft Copilot; pairs with facilitating teachers; 9 schools, first year of senior secondary (around age 15)0.31 SD overall; 0.23-0.24 in English; 0.21 on the year-end exam; "1.5-2 years" of schooling; 3.2 years of schooling per $100Uneven attrition (64% in the treatment group vs. 50% in the control); dropping one school lowers the English effect to 0.156 (significant only at 10%); the control group got no alternative activity
Oreopoulos and Low, Khanmigo
NBER 35620, 2026
Two-year cluster trial, 18 middle schools in Tennessee, remedial math0.06-0.08 SD per year; 0.14 implied for a full year of participation; identical to Khan Academy without AI96% tried it, but the median student messaged it on only a third of practice days, and in only 17% of sessions where they made a mistake
Bastani et al., PNAS
June 2025, Turkey
About 1,000 high school math studentsWith access: +48% (plain GPT), +127% (GPT Tutor). After access was removed: plain GPT -17%; GPT Tutor no harm but no gain eitherA "crutch" effect; guardrails prevent harm but don't create learning
Tutor CoPilot
Wang, Loeb, Demszky et al., Stanford
Real-time AI suggestions for human teachers; about 900 teachers and 1,800 students (version 1; version 2: 700+/1,000+), March-May 2024+4 percentage points in mastery; +9 points for students of lower-rated teachers; about $20 per teacher per yearMastery measured within the platform; sample differs between versions of the paper
Google DeepMind LearnLM and Eedi
arXiv, December 2025, UK
Pilot trial, 165 students, five schools, ages 13-15, summer 2025; teachers supervised LearnLM drafts5.5 percentage points more solved new problems (66.2% vs. 60.7%); 76.4% of AI drafts were approved with no edits or minimal editsSmall and exploratory; compares supervised AI to humans alone, not AI alone

What this means:

  1. The best results come from a human in the loop plus structure: facilitated sessions in Nigeria, supervised drafts in LearnLM, and a teacher assistant in CoPilot
  2. When students choose whether to use it, as with Khanmigo, usage is sparse and the effects look like ordinary practice software, which typically yields 0.05 to 0.20 SD
  3. AI that gives answers can hurt learning without the tool, and this is the most important finding
  4. No rigorous study yet shows AI replacing a full year of schooling with durable gains
06 — Countries and National Programs

Top-down mandates fail; gradual training keeps expanding

Countries that imposed AI-delivered instruction from the top down failed; countries that start with AI literacy and teacher training keep expanding.

CountryWhat happenedNumbersStatus & lesson
South KoreaAI-based digital textbooks (math, English, computer science) launched in March 2025 as a flagship of the Yoon administrationThe government spent 1.2+ trillion won (about $850 million); publishers spent about 800 billion won (about $567 million). Adoption dropped from 37% to 19% of schools; 2,095 schools remained. 98.5% of 2,626 teachers said the training wasn't enoughIn August 2025 the National Assembly stripped their legal status after one semester and reclassified them as "supplementary material." Complaints: factual errors, privacy, screen time, and workload. Publishers sued. The clearest failure of a top-down mandate
EstoniaAI Leap (TI-Hüpe), a public-private fund led by the president's office and the ministry, in partnership with OpenAI and AnthropicPhase 1 (September 2025): about 20,000 students in grades 10-11, 3,000-4,700 teachers (sources differ). An Estonian-language ChatGPT app launched in January 2026; by March 2026, 7,700 students had activated an account, 47% of them weekly users, and 60%+ of teachers use ChatGPT/Gemini weekly. Expansion targets conflict between sources: Eurydice cites 58,000 students/5,000 teachers by 2026, Euronews cites 48,000/6,700 within two yearsExpanding in 2026/27 to all upper-secondary and vocational schools. The leading model of "teacher-led Socratic AI"
ChinaThe Ministry of Education published two directives in May 2025: general AI education, and use of generative AI in schoolsBeijing: mandatory AI education, 8+ hours a year from fall 2025, starting at age 6Elementary students aren't allowed to use generative AI on their own; teachers aren't allowed to use it as a substitute for core instruction, to answer students, or to grade them. Submitting AI-generated homework is banned. AI literacy is rising; AI replacing teachers is explicitly banned
UAEAI became a subject in all government schools, from mandatory kindergarten through grade 12, starting 2025-261,000+ teachers trained; renamed "Artificial Intelligence and Technology" for 2026-27. On September 2, 2026, a program was approved for all public and private schools. KHDA/MIT RAISE: 80,500 students and 3,600 teachers by February 2030Start date for private schools not yet announced. The most comprehensive mandatory AI literacy program
IndiaAI and computational thinking will be integrated across all grades starting the 2026-27 school year (NEP 2020, NCF-SE 2023)CBSE and NCERT are building frameworks; training via NISHTHA; the SOAR initiativeStill in the strategy and rollout phase; the scope and quality of implementation are not yet known
IsraelA national AI program was announced in February 2025; "Project 720" for personalized learning70,000 teachers to be trained; 3,000 mentors from 400+ companies (Google, Microsoft, Apple, Nvidia); 5 tools including a Minecraft-based interface; about 30 services evaluated; a Gemini-based chatbot called "QBot." AI-based English was piloted in 28 schools, expanding in 2026-27 to all 180 middle schoolsAn explicit response to an "acute teacher shortage." The 2026-27 school year (about 2.6 million students) emphasizes AI in learning and exams. A national case of AI filling a workforce gap
USNo federal mandate; state-level guidance is proliferatingA Virginia law took effect July 1, 2026; bills in 30+ states; New York and Los Angeles banned AI at certain timesFederal leadership openly supports AI-based models, including McMahon's support for Alpha
Not verified in this round Singapore, and the UK's Oak National Academy lesson tool Aila and Department for Education guidance, were within the research scope but weren't verified in this round. Check any figure about them against another source.

Adoption data

Teachers
Weekly AI use (Gallup-Walton, 2025)
32%
AI use at work, Israel (TALIS 2024)
44%
Students and children
Students who used AI for schoolwork, 2025 (RAND)
54%
AI use for homework, May 2025 (RAND)
48%
AI use for homework, December 2025 (RAND)
62%
Children 9-17 who use generative AI (Common Sense, 2026)
86%

Students and teachers are using AI far faster than institutions are setting rules. Unsupervised use is the norm, and that's exactly the pattern the research links to offloading thinking onto the tool.

07 — The Ten Categories

Maturity, evidence, players, and crowding

Tutors and teacher assistants are saturated and offered for free by the tech giants; peer learning, certification, generative worlds, and audio are nearly empty content areas. The crowding assessment is the author's own, based on funding patterns, Big Tech's entry, and the number of products.

CategoryWhat's happeningMaturityEvidenceNotable players & movesCrowding
AI TutorsSocratic tutors, and "study modes" inside general assistantsHigh (product); Low (impact)Khanmigo small; Nigeria and LearnLM positive with structure; Bastani shows harm without guardrailsKhanmigo (from 40,000 students in 2023 to about a million with access, usage stalled); ChatGPT Study Mode; Gemini Guided Learning; Claude Learning Mode; Eedi and LearnLMVery crowded. Tech giants are giving it away free
Teacher AssistantsLesson planning, rubrics, personalizationHighAbout 25 min/week in planning (31%); 5.9 hours is self-reported onlyMagicSchool (Series B ~$45M, Jan 2025); Brisk ($15M); SchoolAI (Series A $25M); ChatGPT for Teachers (free through 2028); Claude for Teachers; Gemini for Education; Microsoft CopilotSaturated, has become a commodity
Agentic LearningAI that plans, organizes, reminds, and acts across toolsEarlyAlmost none in K-12Alpha's TimeBack is the closest thing to an "orchestrator"; Big Tech agentsMissing / early
Voice InterfacesSpeaking practice, read-aloud, oral tutoringMedium (languages); Low (other)Strong for language speaking (product data); few controlled trialsSpeak ($78M, OpenAI-backed); Praktika, SpeakX, Loora, Stimuler; AI-based English in IsraelCrowded in languages, open elsewhere
AI CompanionsEmotional and social chatbotsHigh use, low legitimacyMostly risk evidence; Brookings flags socio-emotional harmConsumer platforms (Character.AI); mostly excluded from schoolsHigh regulatory and ethical risk
Adaptive CurriculaMastery-based sequencing (pre-LLM), now with LLM-generated contentHigh (mature software)0.05-0.20 SD when used as intendedIXL, DreamBox, Khan, Magma Math; TimeBack; Korea's AI textbooks (failed)Crowded with veteran players
Simulation LearningVirtual labs, role-playMediumOld evidence for labs; little in the LLM eraMinecraft Education, virtual lab vendorsModerately underserved
Generative WorldsAI-generated 3D worlds and gamesVery earlyNoneMostly experiments; Minecraft-based interfacesMissing. Cost and safety are the barriers
Automated AssessmentFeedback on essays, short-answer checkingMedium57% improvement in grading/feedback (Gallup); China bans AI grading studentsBrisk, MagicSchool, Eedi, KhanMedium for feedback; missing in valid, cheating-resistant assessment
Peer-Learning OrchestrationAI that forms groups, facilitates discussion and peer feedbackEarlyAlmost noneFew startups; the workshops at Alpha and Unbound are run by peopleThe biggest gap, and the role schools will give up last
Credentials Outside SchoolMastery records, skill badges, portfoliosLow for K-12None for K-12Adult/employment players (Multiverse, Preply) dominate fundingMissing in K-12. The barrier is university/employer recognition
Where the money goes HolonIQ reports about $2.4 billion in edtech VC funding in 2024, the lowest in a decade and 89% below the 2021 peak. For 2025: about $2.6 billion in total investment or $2.4 billion in VC, depending on the definition. In Q1 2026: $512 million across 63 deals; in H1 2026: about $1 billion, down 26% year over year. Workforce training takes most of the money. In K-12, "AI-based personalization" (Subject, Gizmo) stood out. One secondary tracker (New Market Pitch) counts only about $229 million across 26 rounds of pure AI-in-education companies from Q1 2025 through Q1 2026, with tutors and teacher assistants taking about 85%. What this means: AI-based K-12 education is a small, concentrated VC market. Startups compete less with each other and more with the free offerings from OpenAI, Google, Anthropic, and Microsoft.

Cautionary tales

General information about these cases; not re-verified in this research round.

08 — Synthesis

Which school roles are breaking down?

Content delivery and practice are under the most pressure; motivation, supervision, and socialization are barely affected.

School roleBreakdown pressure (5-10 years)Why
Content deliveryHighFree tutors and study modes from Big Tech are everywhere; 86% of children already use generative AI
Practice and feedbackHighAdaptive practice with cheap LLM feedback; decent evidence when the practice actually happens
Teacher prep and adminHigh (within the school)Proven but modest time savings; strengthens the school rather than replacing it
AssessmentMedium, in the opposite directionAI breaks take-home assessment, so schools are shifting to supervised, in-class, and oral assessment. This actually increases the value of the physical school as a place of verification
CertificationLowUniversities and states still control credentials; no alternative K-12 credential has market recognition
Motivation and structureLow (the central bottleneck)Khanmigo and Unbound show that software without structure fails; Alpha's apparent impact comes from incentives and environment
Supervision and socializationVery lowWorking parents need a structure for their kids; even Unbound runs live-video lunches and quarterly in-person events

Realistic scenarios, 2031-2036

The probabilities are the author's own judgment, not forecasts from any source.

~60%Most likely: the "augmented school." Schools remain, and AI takes over prep, personalization, practice, and feedback. The school day shifts toward supervised practice, discussion, projects, and face-to-face assessment. RAND explicitly recommends flipped models with AI-free classroom time
~25%A growing niche: "micro-schools with compressed academics." Alpha-style models spread through vouchers and education savings accounts in the US and through private markets elsewhere. The evidence stays contested, and results depend on selection and incentives
~15%Systemic replacement in shortage areas. In places with teacher shortages (English in Israel, rural regions, 250 million children without reliable access to content per Brookings), AI delivers instruction with facilitators instead of certified teachers. The only scenario where AI replaces teachers to a meaningful degree; Nigeria's facilitated sessions are the best evidence for it

Equity. The pattern cuts both ways. Affluent families get AI together with heavy human structure, like Alpha at $40,000 to $75,000. Students from lower-income families are more likely to get AI instead of human structure: free, less reliable tools used alone, which is exactly the pattern the research links to harm. Unbound's results suggest that the non-selective, online version of an AI-based model doesn't close gaps on its own. Nigeria suggests the facilitated version can.

09 — For the Investor, the Systems Changer, and the Mentor

What this means

The meaningful change won't come from AI that explains better, but from whoever solves motivation, assessment, and certification at scale. This is the author's own assessment based on the findings above, not a finding from any particular source.

What could really change the school, and what probably won't

  1. Solving engagement at scale. The problem nobody has solved: Khanmigo reached about a million students with access, and usage stalled. Whoever shows consistent use by unselected students over a full year will change the picture more than any model improvement
  2. Reliable assessment in the AI era. Homework no longer proves anything. Whoever builds oral, process-based, or supervised assessment that teachers and systems trust will also shape what gets taught, because assessment dictates curriculum
  3. Alternative credentials that get recognized. As long as universities, employers, and admissions bodies recognize only school credentials, classroom time keeps its monopoly. Recognition of mastery records or portfolios is what could break the school open from the outside
  4. AI that fills teacher shortages. This is the only place where real replacement of instruction is plausible in the near term: middle-school English in Israel, facilitated sessions in Nigeria. Here, the best evidence comes from combining AI with a facilitating adult, not from AI alone
  5. Changing the teacher's role. Tutor CoPilot shows that a tool improving weaker teachers is cheap and effective. The shift from a teacher who delivers content to a teacher who coaches and supervises is a deeper change than replacing the textbook

Less likely to change anything fundamental: yet another general chat tutor (offered free by Big Tech), top-down mandated AI textbooks (Korea), and consumer AI companions (high use, high risk, no proven educational value).

Three perspectives
Wearing the investor's hat
  • Where to be careful: tutors and teacher assistants have become commodities against free offerings from OpenAI, Google, Anthropic, and Microsoft. The first question for any company: what happens when this feature arrives free in a tool the teacher already has installed?
  • Where there's room: engagement and habit loops, integrity and assessment, peer-learning orchestration, credentialing infrastructure, and "AI with a facilitator" solutions for teacher-shortage markets. Selling to governments rolling out national programs (Estonia, UAE, Israel, India) is a viable path, with clear regulatory risk
  • Due-diligence questions: is there usage data from unselected students? Was learning measured without the tool? Who distributes, the teacher or the system? How dependent is the product on a single regulatory decision? Korean publishers lost about 800 billion won on one such decision
  • Market context: edtech capital is at a decade low, so it's reasonable to demand more evidence and real traction before investing
Wearing the systems-changer's hat
  • The lesson from Korea and Estonia: top-down mandates without teacher training fail. Staged rollouts, training before student access, a clear legal basis, and a tool that guides thinking rather than giving answers, hold up
  • Israel is now in the middle of its rollout: 70,000 teachers in training and AI-based English in all 180 middle schools. This is the moment to demand structured evaluation: comparison groups, measuring learning without the tool, and separate testing for weaker students
  • Push for assessment reform: the shift to in-class and oral assessment is a precondition for AI to help learning rather than just task completion
  • Protect equity: the danger is that strong students get AI together with human structure, while weaker ones get AI instead of human structure. That is exactly the pattern the research links to harm
Wearing the mentor's hat at MindCET

Six questions worth asking every founder:

  • What gets the student to use it, and why would they still be using it in week three?
  • What does the student know without the tool, a month later?
  • What happens to the product when its core feature arrives free in ChatGPT or Gemini?
  • Who's the paying customer, and what does the actual procurement path look like?
  • Does the product strengthen the teacher or bypass them?
  • What's the evidence plan: a pilot with a control group, or just usage metrics?

It's also worth pushing founders to design with guardrails: hints and questions instead of answers. The Bastani research shows this is the difference between a tool that helps and one that hurts.

Signals worth watching in 2027-2028

10 — Caveats

Limitations of this research

Note
  • A large share of the documentation on Alpha and Unbound comes from journalism and blogs, including Piper, Astral Codex Ten, WIRED, and parent reviews, because Alpha hasn't released data. Unbound's proficiency figures come from Piper's reporting on the school's own presentation. Arizona's official letter grades were not checked here
  • Kestin et al.'s effect sizes are summarized as framed in the published paper and were not re-extracted
  • The Nigeria results are sensitive to one school and to uneven attrition. LearnLM is a small pilot study. Tutor CoPilot's sample differs between versions of the paper
  • Funding totals vary by tracker and definition ($2.6 billion in investment vs. $2.4 billion in VC for 2025 per HolonIQ). Startup user numbers are company claims
  • The cautionary tales and some category details (companions, simulation, generative worlds, credentials) are based on general knowledge, not new verification in this round. Details on Singapore and on Oak and the UK Department for Education were not verified
  • Everything in the scenarios section is a forecast, not an observed fact
  • The 5.9-hour figure is self-reported from the 2025 survey. In an earlier version of this report it was mistakenly attributed to the 2026 survey