AI in Education: Personalized Tutors and Automated Curriculum Design

The three top real uses in 2026

AI-driven personalized tutoring and adaptive learning, AI-supported assessment and feedback, and AI as an assistant offloading teachers' routine administrative workload are three distinct, complementary applications — not one generic "AI in the classroom" story. (etcjournal.com)

Scale: adoption is real and fast

Khanmigo grew from roughly 68,000 student and teacher users across partner districts in the 2023-24 school year to more than 700,000 in 2024-25, expanding from 45 to more than 380 district partners — and by 2026 reached 1.4 million users across 70+ countries, with nearly half of the two million total signups being teachers who joined once it became free. (valueaddvc.com, minssam.com) Khan Academy and Duolingo have independently deployed autonomous AI tutoring agents that together serve over 50 million active learners — Duolingo Max's AI agents now drive more than 30% of Duolingo's revenue in 2026. (callsphere.ai)

Measured learning outcomes: real, but mixed

A 2025 randomized controlled trial tracking students using Khanmigo for math support three times weekly found a 0.34-standard-deviation gain in algebra performance over a single semester. (etcjournal.com) Separately, pilot districts reported a 1.4 grade-level improvement, and some math deployments show learning gains up to 23%. (valueaddvc.com) These are genuine, controlled-study results, not marketing claims.

But a separate, important line of research complicates the picture on what kind of learning AI tutoring produces. One study found AI-assisted students answered 48% more problems correctly, but their score on a concept-understanding test was 17% lower — AI improved procedural performance without producing deeper conceptual learning. (rand.org) The emerging consensus is that the level of human tutor involvement is a key factor in whether AI tutoring produces real learning or just faster answers — a well-designed AI tutor built on pedagogical best practices measurably outperformed in-class active learning in one RCT, while poorly scaffolded AI use produces answer-getting without understanding. (rand.org)

The critical-thinking concern is now backed by survey data, not just anecdote

RAND's American Youth Panel found that more US students from middle school through college used AI for homework in 2025, and 60% of them worry that AI use will harm their own critical thinking skills. Teachers share the concern at a higher rate: 70% worry AI weakens critical thinking and research skills, and over half of students say using AI in class makes them feel less connected to their teachers. (rand.org)

The practical fix experts point to is not less AI access but more structure around it: teacher training and student AI-literacy instruction are cited as the most effective way to manage the risk, but most teachers and students currently use these tools without any formal training or institutional guidance. (rand.org)

The teacher-workload use case has the least controversial ROI

Beyond tutoring, Khanmigo helps teachers with lesson planning, rubric generation, quiz creation, and explanation drafting — one Northern California school district estimated it saved teachers about five hours per week. (valueaddvc.com) This administrative-offload use case doesn't carry the same critical-thinking tradeoff debate that direct student tutoring does, since it's the teacher's workflow being automated, not the student's reasoning process.

Special education is where AI adoption is growing fastest — and where the stakes are highest

The most underreported adoption story in 2026 education AI isn't general classroom tutoring — it's special education, where overworked and understaffed teachers are turning to AI specifically to manage the enormous paperwork burden of Individualized Education Programs (IEPs). A CDT (Center for Democracy and Technology) survey found 57% of special education teachers nationwide used AI to help develop individualized plans for their students in the 2024–25 school year, up sharply from 39% the year before — an 18-point jump in a single year, among the fastest adoption curves in any education AI use case measured. (npr.org)

The CDT data also breaks down what specifically teachers use AI for within IEP development, and the pattern shows a clear preference for AI-assisted judgment over AI-generated output: about a third of special educators used AI to identify trends in student progress and help set goals, and to summarize the content of existing IEP and 504 plans. Fewer used it for more consequential, judgment-heavy tasks — 21% used AI to write the narrative portion of a plan, and only 15% used it to fully write an IEP or 504 plan outright. (npr.org) That gradient — high usage for pattern-recognition and summarization, low usage for full document generation — mirrors the same "AI assists, human decides" structure that shows up in the legal and customer-service domains, and it's a more disciplined adoption pattern than the raw 57% headline number suggests on its own.

Stakeholder sentiment is notably more positive here than the general critical-thinking concerns raised earlier in mainstream classroom AI use: 64% of parents of students with an IEP or Section 504 plan, and 63% of students with either plan, said this use of AI is a good idea. (npr.org) That's a meaningfully different reception than the 60–70% critical-thinking-erosion worry reported among general-education students and teachers — the difference likely comes down to what's being automated: IEP paperwork and progress-summarization are administrative burdens most families are glad to see reduced, whereas AI doing a student's actual reasoning work triggers a different, justified concern.

But heightened use in special education also carries elevated legal and privacy risk that general classroom AI tools don't. IEPs and 504 plans contain protected health and disability information governed by IDEA (Individuals with Disabilities Education Act) and Section 504 of the Rehabilitation Act, and education-law researchers have flagged that AI tools generating or summarizing this content raise open questions about data handling, vendor access to protected student records, and whether an AI-assisted accommodation decision could be challenged as not reflecting genuine individualized judgment — a core legal requirement of an IEP under federal law. (k12dive.com) Districts adopting AI for IEP support in 2026 are being advised to treat it as a documentation and drafting aid layered under a human case manager's legally required individualized decision-making, not a system that outputs the plan itself — echoing, in a special-education-specific and higher-stakes form, the same "AI drafts, human decides and is accountable" principle threaded through every other domain in this piece.

Higher education tells a different adoption story than K-12

Everything above concerns K-12 tutoring and IEP support. College is a distinct market with its own usage patterns and a notably more positive self-reported outcome picture. More than half — 57% — of US college students report using AI in their coursework at least weekly, including roughly one in five using it daily, and students in business, technology, and engineering programs use it most frequently of any field of study. (news.gallup.com) The Coursera AI in Higher Education Report, released February 2026, found four in five students report AI has improved their academic performance, and — notably, given the "AI does the thinking for you" critique raised earlier — 63% say they use AI for less than half of their academic tasks, suggesting most students in this cohort are treating it as a supplement rather than a replacement for their own work. (news.gallup.com)

The controlled-study evidence at the college level is also stronger than the K-12 RCT data cited earlier. A peer-reviewed randomized controlled trial published in Scientific Reports in mid-2025 found a well-designed AI tutor outperformed traditional in-class active learning with an effect size of 0.73–1.3 standard deviations — more than double the 0.34 SD gain Khanmigo produced in the K-12 algebra RCT — and students using the AI tutor reached higher post-test scores in less time than their in-class peers. (news.gallup.com) Separately, a Microsoft-backed study of Indiana students given AI homework support found a 10% increase in grades alongside a 40% reduction in time spent completing assignments. (news.gallup.com)

That gap between the K-12 RCT effect size (0.34 SD) and the higher-ed RCT effect size (0.73–1.3 SD) is worth sitting with rather than treating as a simple "AI works better in college" conclusion — it more likely reflects differences in tutor design quality, subject matter (algebra versus the broader coursework tested in the higher-ed study), and the fact that college students self-select into AI use rather than being assigned it uniformly the way K-12 pilot cohorts often are.

The measurable gap AI tutoring hasn't closed: emotional read

One concrete, measured limitation cuts across every age group and explains why "human tutor involvement" keeps surfacing as the deciding factor in outcome quality. Human tutors can read a student's emotional state — frustration, disengagement, confusion that isn't being voiced — with about 92% accuracy. The most advanced AI tutoring systems currently manage roughly 68% accuracy on the same task. (news.gallup.com) That 24-point gap is a plausible mechanical explanation for the RAND finding cited earlier — that AI-assisted students got more answers right while showing weaker conceptual understanding: an AI tutor that can't reliably detect that a student is confused rather than simply slow is structurally more likely to let a student proceed with a superficial or memorized approach instead of intervening the way a human tutor would when they sense something isn't landing.

Sources: (news.gallup.com)

This is the strongest evidence yet for why the field keeps converging on "AI tutor plus human oversight" rather than either extreme. The emotional-read gap doesn't disappear as models get more capable at answering questions correctly — it's a distinct capability that requires either better multimodal signal (tone, hesitation patterns, response latency) or, more practically in the near term, a human in the loop who catches what the AI tutor's 68% accuracy misses roughly a third of the time.

A real, near-term regulatory deadline

The EU AI Act classifies AI systems that determine access to education, evaluate learning outcomes, or monitor students during exams as high-risk. Obligations become enforceable August 2, 2026, requiring a risk-management system, data governance, and mandatory human oversight for any tool in that category. (etcjournal.com) Notably, this classification targets assessment and access-decision AI specifically — supplemental tutoring tools like Khanmigo fall outside the high-risk category as long as they don't determine grades, admissions, or exam outcomes.

The practical takeaway

The strongest current evidence supports supplemental tutoring support with human oversight (Khanmigo's math use case) and teacher administrative offload — not full curriculum replacement or unsupervised AI-driven assessment. Given that 60-70% of both students and teachers already worry about critical-thinking erosion, and that procedural gains have been shown to come at the cost of conceptual understanding in at least one study, education AI deployments should pair tool access with explicit AI-literacy training rather than treating access alone as the intervention. Any tool touching assessment, access decisions, or exam monitoring needs to take the EU's high-risk classification and its August 2026 obligations seriously if operating in that market.

Sources: ETC Journal — Three Best Uses of AI in Education in 2026, RAND — American Youth Panel, AI Homework and Critical Thinking, ValueAddVC — AI Tutors Are Here 2026, Minssam — Khanmigo/Duolingo EdTech Revolution 2026, CallSphere — AI Agents in Education, NPR — Special Education Teachers Turn to AI for IEP Help, K-12 Dive — Heightened AI Use in Special Education Brings Elevated Risks, Gallup — AI Is Routine for College Students, Despite Campus Limits