Education
Last reviewed
May 9, 2026
Sources
26 citations
Review status
Source-backed
Revision
v2 ยท 3,987 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 9, 2026
Sources
26 citations
Review status
Source-backed
Revision
v2 ยท 3,987 words
Add missing citations, update stale details, or suggest a clearer explanation.
Education is one of the sectors most directly reshaped by the rise of generative AI. The launch of ChatGPT in November 2022 changed almost overnight what students could do with a laptop and a writing prompt, and forced teachers, school districts, and universities to rethink assignments, grading, and even what classroom learning is for. By early 2026, AI tools are embedded throughout education: K-5 students get math help from Socratic AI tutors, high schoolers debate whether Khanmigo counts as cheating, college freshmen run their lecture notes through NotebookLM for podcast-style summaries, and graduate students paraphrase, cite, and detect with a stack of overlapping tools. The 2025 RAND survey found 54% of K-12 students used AI for school, up more than 15 points in two years; the 2025 Higher Education Policy Institute survey put student use of generative AI for assessments at 88%.
This page is a gateway to the wider topic. The long-form article on AI in education covers the history (PLATO, intelligent tutoring systems, the LLM turning point) and the research base in detail. Below, the focus is on the products, policies, and arguments that define the current landscape.
For decades, classroom technology meant smart boards, learning management systems, and the slow grind of edtech pilots that mostly failed to move test scores. Adaptive math platforms like ALEKS and Carnegie Learning's MATHia had already shown that software could personalize instruction at the skill level, but they were narrow, expensive, and limited to specific subjects. ChatGPT collapsed all of that into a single chat box that could explain photosynthesis, write a sonnet, debug Python, and draft a college admissions essay.
Within two months of launch, ChatGPT had reached 100 million users. Schools panicked. New York City Public Schools blocked the tool on school networks in January 2023, citing concerns about "negative impacts on student learning, and concerns regarding the safety and accuracy of content." The Los Angeles Unified School District followed with its own preemptive block. Both reversed course within months. By May 2023 New York had lifted its ban and announced an AI policy lab; the conversation moved from "how do we keep this out" to "how do we use it without ruining the assignment."
The more important shift happened on the supply side. OpenAI, Google, Anthropic, and a wave of startups began building education-specific products. Khan Academy partnered with OpenAI on Khanmigo. Duolingo built Max on top of GPT-4. Synthesis Tutor came out of Elon Musk's Ad Astra school. Coursera launched Coach. Within roughly eighteen months, education went from one of the most ChatGPT-resistant sectors to one of the most heavily targeted.
One-on-one AI tutors are the most visible product category. They aim at something education researchers have wanted for decades: the "two-sigma problem" identified by Benjamin Bloom in 1984, which found that one-on-one tutoring produces a roughly two standard deviation improvement over standard classroom instruction. Software has been chasing that effect ever since.
| Tutor | Built on | Pricing (2026) | Notes |
|---|---|---|---|
| Khanmigo | GPT-4 / GPT-5 | $4/month or $44/year for individuals; free for teachers | Khan Academy's tutor; designed to ask Socratic questions rather than give answers. Grew from 40,000 to 700,000 K-12 students between 2024 and 2025. |
| Duolingo Max | GPT-4 | $29.99/month | Adds Roleplay (simulated conversations) and Explain My Answer to the standard Duolingo course. Video Call with Lily added 2024. |
| ChatGPT Edu | GPT-5 | Institutional license | Built on ChatGPT Enterprise; first deployed at scale by Arizona State University after the January 2024 partnership. ASU made it free to all 180,000+ students, faculty, and staff starting October 2025. |
| Synthesis Tutor | Custom GPT-4 stack | $29/month | K-5 math tutor descended from the Ad Astra school at SpaceX. Founder Chrisman Frank previously worked at ClassDojo. |
| Socratic by Google | Gemini | Free | Mobile app that lets students photograph a problem and get a step-by-step explanation. |
Khanmigo has been the most-watched experiment because Sal Khan made specific, falsifiable claims about it. In his 2024 book Brave New Words, Khan argued AI tutors could give every child a Bloom-style two-sigma boost. The actual rollout has been more uneven. A 2025 study found that classes using Khan Academy alongside teacher instruction made slightly faster learning gains, but lower-performing students saw few improvements. Khan Academy responded by redesigning the product so Khanmigo appears directly inside practice problems instead of waiting for students to seek it out, after internal data showed students were not asking for help as often as the team had hoped.
Synthesis Tutor takes a different approach. Rather than guide students through homework, it runs a Socratic dialogue covering the K-5 math curriculum and pushes harder questions when answers are correct. It bills itself as the "world's first superhuman math tutor," which is the kind of claim education writers have been hearing for forty years; the open question is whether the underlying model is finally good enough to back it up.
The older generation of adaptive platforms used hand-built knowledge graphs and Bayesian models rather than LLMs. They are still in heavy use, partly because they have years of validation data and partly because they integrate with existing curricula in ways general chatbots do not.
| Platform | Owner | Subjects | Approach |
|---|---|---|---|
| Squirrel AI | Yixue (China) | K-12 math, English, Chinese | "Large adaptive model" trained on data from 24 million students and 10 billion learning behaviors. Operates 3,000+ physical learning centers. Named to TIME's Best Inventions of 2025. |
| Carnegie Learning MATHia | Carnegie Learning | Grades 6-12 math | Cognitive Tutor descendant; based on John Anderson's ACT-R architecture from Carnegie Mellon. 2025 CODiE Award winner. RAND randomized controlled trial validated improved Algebra I outcomes. |
| ALEKS | McGraw-Hill | Grades 3-12 math, college math, chemistry | Knowledge-space theory; uses adaptive questioning to map a student's "knowledge state" and pick the next problem accordingly. |
| DreamBox Learning | Discovery Education | K-8 math | Pioneered "intelligent adaptive learning" branding. Originally founded 2006. |
| Knewton (defunct) | Wiley | Adaptive engine | Raised $180M; partnered with Pearson; sold to Wiley in 2019 at a fraction of peak valuation. The cautionary tale of pre-LLM adaptive learning. |
Squirrel AI is interesting because it inverts the usual model. Where most US adaptive products are sold to schools as software, Squirrel runs its own brick-and-mortar tutoring centers staffed by a mix of AI and human teachers. The company claims this lets it gather richer behavioral data than competitors can. Its expansion into the US market in 2024-2025 will be a real test of whether the approach travels.
This is where the most students actually spend their AI time. The everyday workflow looks like: brainstorm with ChatGPT, research with Perplexity, summarize PDFs with NotebookLM, polish with Grammarly, paraphrase with QuillBot, and pray that none of it trips a detector.
Perplexity has carved out the research niche by combining LLM responses with live web search and inline citations. Students like it because the answers come with footnotes, which makes the resulting paper easier to defend than a raw ChatGPT export. Perplexity offers a 50% student discount at $10/month, one of the few major AI products that does.
NotebookLM from Google launched in 2023 and exploded in 2024 once it could turn uploaded sources into a podcast-style "Audio Overview." Two synthetic hosts read through your lecture notes or research papers and have a conversation about them. Students treat it as a study aid for commuting; researchers use it to triage a stack of papers before deciding which to read in full. The September 2025 update added Brief, Critique, and Debate formats and pushed audio coverage past 80 languages.
QuillBot and Grammarly sit one step earlier, at the editing layer. QuillBot paraphrases. Grammarly checks grammar, tone, and clarity, and now includes a generative writing assistant. The trick of running ChatGPT output through QuillBot to defeat detection is widely documented and, as of late 2025, no longer reliable. Turnitin's update specifically targets AI-paraphrased text.
ChatGPT itself remains the default first stop for most students. OpenAI launched ChatGPT Edu in May 2024 as a university-licensed version with admin controls and stronger data privacy commitments, after Oxford, Wharton, UT Austin, ASU, and Columbia had built up usage on ChatGPT Enterprise. ASU's October 2025 decision to give every student GPT-5 access at no individual cost, paid for through a system license, is the model many other large universities are studying.
Replit is the most interesting case here. Founded as an in-browser coding environment, it pivoted hard into AI with Replit Agent in 2024, which lets users describe an app in plain English and have it built and deployed. The user base went from 30 million to 40 million between Agent's launch and September 2025, with much of that growth coming from students and hobbyists who had previously bounced off learning to code. CEO Amjad Masad's stated long-term thesis is that the students who learn to build with Replit now will be Replit users for life, the same way Microsoft captured a generation of office workers through school computer labs.
Codecademy (now part of Skillsoft) has integrated GPT-4 into its courses for personalized hints and code review. Wolfram Research's Wolfram|Alpha is heavily used in physics, chemistry, and college math, and now ships a ChatGPT plugin so the LLM can hand off symbolic math to Wolfram's engine and avoid hallucinating answers.
For coding more broadly, GitHub Copilot, Claude Code, and Cursor have crossed over from professional use into student workflows. CS departments are split on what to do about this. Some require students to disable Copilot during assignments; others assume it will be used and grade for understanding via oral exams or code walkthroughs.
Language apps were among the first to find a real product-market fit for LLMs. The case for AI here is concrete: practicing speaking with a patient, infinitely available conversation partner is exactly the missing piece in self-study language learning, and it is something a video call with an AI character can plausibly deliver.
Duolingo's Max tier (powered by GPT-4) added Roleplay and Explain My Answer in March 2023, then Video Call with Lily in 2024. Internal Duolingo data shows 34% better grammar retention with Explain My Answer, and 78% of Roleplay users report increased speaking confidence within four weeks. By 2025 Duolingo had used AI-assisted creation (with human review) to launch 148 new language courses, accelerating content production by an order of magnitude over what its human team could produce alone.
Babbel rolled out an AI conversation partner of its own. Memrise added MemBot, a chat-based tutor. Smaller specialists like Speak (a US startup focused on conversational English for non-native speakers) have built businesses entirely around AI voice practice. The category is crowded, but the user benefit is unusually clear, which is rare in edtech.
The MOOC platforms have taken to AI faster than universities themselves. Coursera launched Coursera Coach in 2024, an AI assistant that breaks down concepts, runs interactive practice before graded assessments, and offers Socratic dialogue inside courses. By late 2025, Coach had supported over a million learners, with Coursera reporting a 9.5% higher quiz pass rate on the first attempt and 11.6% more lessons completed per hour. The University of Michigan was an early institutional adopter, using Coach to bring "in-classroom teaching methods into their online courses."
edX has rolled out an AI tutor inside its courses, and LinkedIn Learning has added AI-driven "Coaching" features that let users ask questions about course content. The pattern across all of them is similar: AI as a study buddy and concept explainer, embedded next to the existing video lectures.
For corporate L&D and professional certification (project management, data analytics, cloud certifications), AI is increasingly the way employers expect employees to learn. The shift from cohort-based courses to on-demand AI-tutored learning is one of the quieter but more consequential changes in the adult education market.
The AI detection industry exists almost entirely because of the academic integrity panic that followed ChatGPT. The big three are GPTZero, Turnitin's AI detector (built into the platform schools already use for plagiarism), and Originality.ai (focused on professional and SEO content but used in academia too).
| Tool | Stated accuracy (2026) | False positive rate | Notes |
|---|---|---|---|
| GPTZero | ~84% | 2-4% on native English | Founded 2023 by Princeton student Edward Tian. Has published ESL de-biasing work claiming 1% false positive rate on non-native English speaker text. |
| Turnitin AI | 85-90% | 1-3% on native English | Deliberately calibrated to allow ~15% of AI content through to keep false positives low. Late 2025 update specifically targets paraphrased text. |
| Originality.ai | 85% (RAID benchmark) | 5-9% on professional content | Highest accuracy on paraphrased content (96.7%); accepts higher false positive rate by design. |
The central problem with all three is the false positive rate on non-native English speakers. A 2023 Stanford-affiliated study published in the journal Patterns found that seven AI detectors flagged 61% of essays by non-native English speakers as AI-generated, with unanimous incorrect verdicts on roughly 20% of papers. The pattern is consistent: students who write more formulaic or simpler English (often because they are still learning) trip the same statistical patterns the detectors associate with AI generation.
The consequences are concrete. Multiple universities, including Princeton and MIT, have advised against relying on AI detection alone. Some institutions disabled Turnitin's AI detection module specifically over ESL false positive concerns. The University of Reading ran a blind study in which 94% of AI-generated submissions went undetected, and 68% of teachers report using detection tools while only 54% feel effective at identifying AI content. The detection arms race is not one schools are obviously winning.
A related concern is hallucinated citations. ChatGPT-3.5 fabricates 39.6% to 55% of citations in literature reviews; GPT-4 still produces 18% to 28.6%. The University of Mississippi published a 2024 study analyzing student-submitted AI-generated sources and found rampant fake DOIs, made-up authors, and journal names that did not exist. A May 2026 Lancet analysis found a steep rise in fraudulent citations across the academic literature, blamed on AI use during literature review. Students who submit AI-drafted papers without verifying citations are not only committing academic dishonesty in many institutional definitions; they are also producing demonstrably false work that breaks the moment a grader follows a footnote.
K-12 is where the policy fights are most visible, partly because it is the most regulated sector and partly because parents are paying close attention. As of March 2025, 28 US states have published AI guidance for K-12 education, up from fewer than ten in early 2024.
| State / Body | Approach |
|---|---|
| Connecticut | Launched an AI Pilot Program in spring 2025 across seven districts, introducing students in grades 7-12 to state-approved AI tools. |
| California | Comprehensive AI guidance covering responsible use, data privacy, and equity. |
| Virginia | Framework for AI literacy and ethical use in K-12 classrooms. |
| North Carolina | Integration guidelines emphasizing teacher training and digital literacy. |
| US Department of Education | 2025 federal guidance on AI use in schools, including recommendations for equity, privacy, and age-appropriate deployment. |
The single biggest gap is teacher readiness. While 63% of US teens report using AI tools like ChatGPT for schoolwork, only 30% of teachers report feeling confident using the same tools. Roughly 74% of districts said they planned to provide teacher AI training by fall 2025, but "planned to" and "actually delivered" are different numbers. Surveys of teachers who do use AI weekly find an average of 5.9 hours saved per week, which translates to roughly six extra weeks of reclaimed time across a school year, mostly going into lesson planning, differentiated materials, and parent communications.
Internationally, the OECD has been running its Artificial Intelligence and the Future of Skills project for years, comparing AI performance against the PISA assessment used to benchmark 15-year-old students across countries. GPT-3.5 solved 73% of the PISA reading test and 66% of the science questions; GPT-4 hit 85% and 84%. PISA 2025 added a Foreign Language Assessment and a "Learning in the Digital World" innovative domain, and the OECD has announced a PISA 2029 Media and AI Literacy assessment. The OECD and European Commission jointly published a draft AI Literacy Framework for Primary and Secondary Education in 2025, the first major international attempt to standardize what AI literacy means in compulsory schooling.
Universities have mostly devolved policy to the course or instructor level, on the theory that an introductory writing seminar and a graduate computer science course need different rules.
Stanford's published guidance is representative: "instructors are free to set their own policies regulating the use of generative AI tools in their courses, including allowing or disallowing some or all uses of such tools." Using AI to substantially complete an assignment is not permitted unless explicitly allowed; students should disclose any non-incidental use; when in doubt, default to disclosure. UPenn takes a similar instructor-by-instructor approach with the same emphasis on transparency and disclosure as a violation of the academic integrity policy when omitted.
MIT has leaned into integration, with departments like the Sloan School publishing course-specific guidelines that often allow AI use for brainstorming and editing while prohibiting it for graded analysis. Princeton declined to ban ChatGPT outright in early 2023 and issued faculty guidance instead. The University of Oxford, Wharton, UT Austin, ASU, and Columbia were all among the first to deploy ChatGPT Enterprise at scale, and ASU's October 2025 expansion to give every student GPT-5 access free of charge is the most visible institutional bet so far.
The deeper question universities are grappling with is what kinds of assessment still work. Take-home essays are increasingly hard to grade fairly when half the class is using AI and half is not. Schools that have redesigned assessments to be more AI-resilient (oral exams, in-class writing, process portfolios, project-based evaluation) report 40% fewer AI-related integrity issues than schools that keep the old format and rely on detection. The cost is that these formats are more labor-intensive for instructors, which is itself a constraint.
The optimistic case for AI in education is straightforward: a personal tutor for every student, finally. Bloom's two-sigma effect, available at $4 a month. The pessimistic case is more layered, and worth taking seriously.
Learning loss and over-reliance. If a student can get an LLM to draft every essay, they may never learn to organize their thoughts or revise their own prose. The same concern applies to math (correct answers without understanding the steps) and to coding (suggestions accepted without comprehension). The risk is that AI becomes a crutch rather than a scaffold. There is some early evidence here. A 2024 MIT Media Lab study tracked brain activity (EEG) of students writing essays with and without ChatGPT and found measurable reductions in cognitive engagement among the AI-assisted group, with weaker recall of their own essays a week later. The study is small and the methodology is debated, but it points at a real worry.
Equity. Paid tiers of AI tools are out of reach for many families. Duolingo Max at $29.99/month, ChatGPT Plus at $20/month, and the various premium tutoring products are not free, and the schools that can afford institutional licenses are not the schools serving the lowest-income students. Khan Academy's free tier and OpenAI's nonprofit pricing for Khanmigo partially address this, but the gap between a student with a $200/year stack of paid AI subscriptions and one without is real, and it is widest in exactly the places where the gap was already widest.
Hallucination. Students who do not understand that LLMs make up plausible-sounding facts and citations submit work that turns out to be factually wrong. This has cost students grades, scholarships, and in some cases enrollment. The hallucination rate has dropped with each model generation, but it has not gone to zero, and most students are not trained to verify outputs.
Detection bias. As covered above, AI detectors disproportionately flag non-native English speakers. The students most likely to be falsely accused of cheating are often the students with the least institutional support to push back.
Data privacy. AI educational tools collect performance metrics, learning patterns, interaction logs, and (in proctoring) biometric data, on minor students. FERPA and COPPA in the US and GDPR in the EU all apply, but enforcement is uneven and most schools are not equipped to audit a vendor's data practices in detail.
Teacher displacement vs. teacher empowerment. This is the longest-running argument in edtech, and AI has rekindled it. Teachers' unions in several countries have raised concerns about AI being used to justify larger class sizes or reduced staffing. Vendors counter that AI saves teachers six weeks a year on routine tasks and frees them to do more direct instruction. Both can be true at once, depending on whether districts pocket the savings or reinvest them.
Validation gap. A widely cited statistic from 2025 is that 86% of educational organizations use generative AI, the highest of any US industry. Yet most AI educational tools have not undergone independent validation, and very few have been tested through randomized controlled trials. The gap between adoption and evidence is wider than in almost any prior edtech wave, and that is saying something given the field's history.