AI in Education: Where Things Actually Stand in 2026
Prepared by Richstorm.co
KEY TAKEAWAYS
Sam Altman's education predictions describe a self-motivated learner — closer to his own childhood than most students' — and offer no actual implementation plan.
Venture capital in AI-education overwhelmingly favors habit-forming practice apps like Duolingo over tools that verify whether real learning happened.
Computer science and liberal arts face different problems: CS is rewriting what it teaches, liberal arts is rewriting how it verifies learning.
The US, EU, and China are running three genuinely different bets — fragmented experimentation, risk-based regulation, and a centrally mandated 2030 target — and none of them is finished.
China's national AI education rollout is a multi-year plan targeting 2030; Squirrel AI, one of its best-known adaptive-learning companies, shows how much the market itself has already had to adapt along the way.
The Prediction That Started This
OpenAI's Sam Altman has said, across several talks and essays, that education hasn't kept pace with AI and needs to change — essays as homework are “never going to be quite the same again,” and “the way we teach people is going to have to change and the way we evaluate students is going to have to change.” His clearest structural claim is that children will eventually have virtual tutors offering personalized instruction at whatever pace they need.
What he hasn't offered is a plan. No proposal for how assessment should actually work, how teachers should be retrained, or how funding and accreditation adapt. It's a founder's directional prediction, not a policy — and it's worth checking against what's actually happening, rather than treating it as a roadmap.
It's also worth naming what the prediction assumes: a learner who pushes themselves to go faster and deeper once given the freedom — which describes Altman's own account of his childhood, and describes a self-motivated student, not most kids. The real constraint for most learners was never pace; it's motivation and structure offered by school, which a personalized AI tutor doesn't automatically supply.
Where the Money Is Actually Going — and Where It Isn't Tracked at All
The best-funded category in AI-education venture capital is narrow, self-directed practice tools — language learning alone has pulled over $400 million across a handful of companies. Teacher-facing productivity tools (lesson plans, rubrics, grading assistance) have added roughly $90 million and now serve millions of teachers through platforms like MagicSchool AI. Assessment and grading tools — the category aimed at verifying whether learning actually happened — raised a small fraction of that, which one research firm tracking the sector called “arguably underfunded relative to its classroom importance.”
Duolingo is the clearest public example of what the best-funded category actually optimizes for — though its success predates generative AI by years. The company launched in 2011 on crowdsourced translation and a basic spaced-repetition flashcard method; in 2016 it published its own machine-learning model to predict when an individual user is about to forget a specific word. Generative AI features (conversational practice, roleplay) were only added starting in 2023–2024. The habit loop that actually drives its business — streaks, leaderboards, points — is applied behavioral psychology, not an AI capability. By late 2025, Duolingo had passed 50 million daily active users and $1 billion in annual bookings, but its stock has also fallen roughly 80% from a May 2025 peak amid questions about whether that AI-driven growth is plateauing.
There's a gap worth naming in all of this funding data: dedicated sector trackers like this one count venture-backed education startups, not general-purpose AI usage — and that usage turns out to be enormous. OpenAI's own August 2026 usage report found US homework- and classwork-related ChatGPT messages peak above 460 million per week during the school year, staying above 180 million even through summer. College Board separately found 84% of US high school students reported using generative AI for schoolwork in 2025, and a UK university survey found assessment-related use jumped from 53% to 88% in a single year. None of that is captured in any education-startup funding report, because a student using ChatGPT directly isn't using an ‘education startup’ at all.
Two Majors, Two Different Problems
Computer science has to rewrite its content, not just its tests. Several universities are adding AI-specific requirements rather than folding AI into existing coursework: the University of Iowa's Fall 2026 curriculum adds a new required course, “Computing, Ethics, and Society,” which the department chair has said explicitly is part of the AI-driven overhaul, not a pre-existing requirement being relabeled. A global ACM survey found more than two-thirds of computing educators have already changed their assessment approach, though “no single model has emerged.” On the hiring side, a 2026 industry survey (CoderPad's State of Tech Hiring) found 46% of hiring leaders now allow AI in technical interviews, with another 20% deciding case by case — and when AI is allowed, catching and fixing the AI's mistakes is one of the strongest signals hiring teams look for, not raw output. Separately, one company's analysis of over 19,000 AI-monitored interviews found the share of candidates flagged for undisclosed AI use during live coding rounds rose from 9% to 48% in six months.
Liberal arts has a narrower problem: the content — reading, interpreting, constructing an argument — hasn't changed. Only the proof mechanism has broken. This is where oral exams have re-emerged: one voice-AI pilot ran 36 undergraduate oral exams and found 8% of students couldn't discuss their own submitted work at all. But a static oral exam just adds a memorization step on top of unchanged homework — what actually resists gaming is adaptive follow-up questioning a rehearsed answer can't survive.
On assessment more broadly: a proctored, in-person paper test already verifies whether someone can produce material themselves — cheaply, and without any AI involved. For remote learners, that gap is already commoditized: AI-based exam monitoring (webcam and behavior tracking during online tests) is now widely adopted — one market research firm put institutional adoption around 70%, though estimates of the market's actual size vary a lot between research vendors (figures for this specific category range from roughly $3 billion to $9 billion by the early 2030s), which itself is a sign of a young, loosely defined market rather than a single settled number.
Three Countries, Three Bets
No consensus exists yet on the right institutional response, and the clearest evidence of that is how differently the three largest education systems in the world are approaching the same problem, at the same moment.
China: Policy Timeline and a Market Case Study
China's plan has real substance behind it, built over nearly a decade: a 2018 higher-ed AI research initiative, rural “smart hardware” rollouts under the 14th Five-Year Plan, 184 schools designated as national AI education pilot bases in 2024, and K-12-wide AI curriculum standards published in May 2025. The April 2026 “AI+Education Action Plan” converts that groundwork into a nationwide mandate, targeting full implementation by 2030 — a multi-year timeline, one year into which the current pilot-and-standards stage is roughly what would be expected.
Squirrel AI, one of the country's best-known AI adaptive-learning companies, is a useful case study in how this market has actually evolved. Its original business supplied an AI tutoring system to off-campus tutoring centers nationwide. China's 2021 “double reduction” policy, which restricts for-profit off-campus academic tutoring for compulsory-education students, required the company to restructure that model — its founder has described a period of significant debt and a 2025 court-approved restructuring. The company has since shifted its business toward AI-equipped learning-machine hardware sold through franchise retail, and software subscriptions sold directly to public schools, which fits the current rules restricting this category of product to institutional buyers. By 2026, the company reports having scaled substantially under this model, including an entry into the US market — figures that come from the company's own disclosures.
The broader pattern across the plan's rollout: real, multi-year policy investment and genuine pilot-program results, with full nationwide classroom adoption still ahead of the 2030 target — consistent with where a plan at this stage would be expected to be.
Where This Leaves Things
A concrete pattern shows up in all three: the clearest evidence of success in each country comes from a specially resourced flagship — California State University's deal with OpenAI, the EU's formal high-risk rules, China's 184 designated pilot schools — while nobody has verified what an ordinary, average institution is actually doing day to day. Flagship examples get the press coverage; typical adoption is the thing nobody has really measured yet, in any of the three.
Inside that uncertainty, a few things are worth watching more closely than the rest. General-purpose generative AI tools — OpenAI's ChatGPT, Anthropic's Claude, Microsoft Copilot, Google's Gemini — are already being used informally for learning, even though that usage doesn't show up in any education-specific funding data; OpenAI and Anthropic are both also moving toward public markets. Gamified, habit-forming learning tools in the Duolingo mold are worth watching in subjects beyond language, where the same atomic, right-or-wrong content structure could transfer. Teacher-facilitation tools, by contrast, are a workflow-efficiency category, not a learning-outcomes one — useful, but a different kind of bet.
Sources
OpenAI (“The Intelligence Age,” essay, 2024; public remarks, Tokyo 2023 and July 2025; usage report, Aug. 2026); College Board (survey on generative AI use among US high school students, 2025); HEPI and Kortext (Student Generative AI Survey 2025); ACM Education Advisory Committee (Task Force on Generative AI and Programming Assessment, global survey, May–Oct. 2025, published July 2026); EDUCAUSE (assessment and generative AI analysis, 2026); New Market Pitch (“AI in Education Startup Funding 2025–2026,” June 2026); University of Iowa Department of Computer Science (Fall 2026 curriculum announcement) and Business Record (“Computer science programs in Iowa transform in the era of AI,” June 2026, quoting CS chair Alberto Segre); CoderPad (State of Tech Hiring 2026 survey); Apollo Technical, citing Fabric's analysis of AI-monitored coding interviews (2026); arXiv (“Scalable and Personalized Oral Assessments Using Voice AI,” Mar. 2026); Dataintelo (“AI Exam Proctoring Market Research Report 2033”) and other market-research vendors with differing size estimates for the same category; Duolingo, Inc. (investor updates 2025–2026; VentureBeat and Springer Nature coverage of Duolingo's AI and spaced-repetition history); California State University and OpenAI (agreement announcement, Jan. 2025); European Commission, Directorate-General CONNECT (EU AI Act Annex III and Digital Omnibus on AI, Regulation (EU) 2026/1744); China Ministry of Education and four co-issuing ministries (“AI+Education Action Plan,” Apr. 2026, via Center for Security and Emerging Technology translation and WTL Governance); 21st Century Business Herald and China Entrepreneur, interviews with Squirrel AI founder Li Haoyang on the company's post-“double reduction” restructuring, 2021–2024; Guanyan Report, industry analysis of China's “AI self-study room” sector, 2025.
