The State of Play

A living index of AI adoption across industries — where established practice meets the bleeding edge
UPDATED DAILY
← All domains

🎓 Education & Learning

AI for teaching, tutoring, assessing, and managing learning experiences. Mostly leading-edge: adaptive tutoring and automated grading are approaching good practice, but institutional adoption is slow due to academic integrity concerns and uneven infrastructure. Three practices are bleeding-edge, including AI-generated curricula and autonomous classroom agents. Most trajectories are stalled — policy and pedagogy lag behind the technology.

15 practices: 2 good practice, 11 leading edge, 2 bleeding edge

The Headline

AI that saves staff time works when a person checks it. AI that judges students, or does their thinking for them, is being pulled back, and Google is now turning pupil access off by default.

15
practices tracked
13
at the frontier
2
moving this fortnight
3,153
evidence items

The Picture

Most schools and universities now have AI tools almost everywhere, but use them lightly: only 15% of students with Khanmigo access use it regularly, and teachers in a large UK trial used Oak National Academy's lesson assistant for just 7% of lessons. The institutions getting value use AI for staff work that a person reviews, such as lesson preparation, enquiry handling and first-pass marking of structured answers. In a randomized trial of 464 teachers, Oak's assistant cut preparation time to 76% of the comparison group's with no loss of quality. Where AI acts directly on students, peers are retreating. New York City has restricted student-facing AI through 8th grade, research universities in the US, UK, Singapore and South Africa are switching off AI-writing detection, and the evidence on tools that judge students keeps getting worse. Being slow with student-facing AI is not the risk. Moving early without oversight is.

This Fortnight

  • Google will switch off Gemini, NotebookLM and Gemini in Classroom by default for K-12 student accounts, starting in the US, until administrators turn them on. The move follows questions from Senators Blumenthal and Hawley and planned restrictions in New York and Los Angeles. Whether pupils get AI is now an explicit decision for district leaders, so expect parents and boards to ask who made it and on what evidence.

  • An independent benchmark found that none of 22 AI models matched human agreement when grading essays, and most could be gamed by a note planted in the answer. The same models matched human graders on short answers, and essay length alone outscored most of them. Machine marking of structured work, with review, is defensible. Essay grading is not, and any vendor selling it should be asked to show how it resists manipulation.

  • An independent evaluation of the AI writing-feedback tool EnlightenAI across nine Uncommon Schools high schools found no detectable effect on writing or AP US History scores. Students submitted 4,846 assignments, yet almost none revised their work using the feedback. The quality of the feedback is no longer the bottleneck. Getting students to act on it is, and no tool fixes that without teachers building it into their lessons.

  • An audit of 311 deployed AI-written exam questions found that the AI's own difficulty labels bore almost no relation to how students actually performed. A separate trial with 140 urology residents found supervisor-selected AI exam items significantly easier than human-written ones. AI-generated questions are fine as drafts for practice quizzes, but not for exams with consequences unless they are piloted and reviewed first.

  • UNC Chapel Hill dropped AI essay scoring in admissions for 2026-27 because AI-written essays were pushing its ratings up. Detection is no answer either: a review of 90 studies calculated that even a 1% false-positive rate would wrongly flag more than 223,000 first-year US college essays a year. Any process that uses AI to judge applicants or authorship needs a person making the final call, plus a written policy, which most institutions still lack.

Coming Up

  • US public entities serving 50,000 or more people must meet the federal web accessibility standard by April 26, 2027, and smaller ones by April 26, 2028. In a controlled test, deaf sign-language users gained 18.7% in comprehension from Zoom's automatic captions, against 41.7% with a live interpreter. Budget for human review of captions and keep interpreters, rather than treating automatic captions as a cheap substitute.

  • EU AI Act obligations for high-risk education AI, including automated essay scoring, apply from December 2, 2027, and they require documented human oversight and transparency. List every tool that grades, screens applicants or monitors students, and record who reviews its output. Make sure that review is real: Greek teachers corrected harsh grades 22% less often when told the grades came from AI.

  • Turnitin has postponed its contested data-use terms to September 2027, while Newcastle, Westminster and Southampton are ending their contracts. Newcastle cited the possible use of student and staff data for AI purposes, and Westminster called the core plagiarism checker largely ineffective against generative AI. Check what your detection contract permits on student data, and plan for assessment redesign rather than a renewal.

What's Hard About This

  • The same tool that safely saves teachers time can harm learning when students use it instead of thinking. In a University of Maryland randomized trial of 2,379 undergraduates, access to an optional AI course tutor cut final grades, and first-generation students lost the most. Products rarely distinguish the two uses, so who controls the tool matters more than which tool you buy.

  • Buying the tools is easy, but getting students to use them is not. In Stanford's trials, 40 to 47% of elementary students never logged in to the AI tutoring platform they were given. Gains in trials depend on adults scheduling, checking and following up, which costs staff time that procurement budgets rarely include.

  • Every credible deployment depends on a person checking the output, and that check eats into the savings while proving weaker than assumed. Curators in Karnataka had to change 94.4% of AI-translated Kannada lesson content, against 3.9% of the English. Meanwhile Mesquite ISD cut drafting of individualized education programs (special-education plans) from up to three hours to about 45 minutes per student without telling families, which shows how quickly oversight and disclosure slip when the time savings are real.

2012View this domain's timeline →Today

Practices in this Domain (15)

PRACTICETIERTREND
Accessibility & accommodation support in educationLEADING EDGE— Steady
AI tutoring — conversational & guided discoveryLEADING EDGE— Steady
AI tutoring — personalised pacing & adaptive difficultyGOOD PRACTICE— Steady
Automated grading & assessmentLEADING EDGE↘ Slowing
Coding education & interactive exercisesLEADING EDGE— Steady
Curriculum design & content generationLEADING EDGE— Steady
Education administration & admissions automationLEADING EDGE— Steady
Educational content adaptation & summarisationBLEEDING EDGE— Steady
Formative feedback generationLEADING EDGE— Steady
Language learning with conversational AILEADING EDGE— Steady
Learning analytics & student risk identificationLEADING EDGE— Steady
Plagiarism & AI-content detectionBLEEDING EDGE↓ Declining
Question & exam generationLEADING EDGE— Steady
Simulated practice environmentsLEADING EDGE— Steady
Skills assessment & competency mappingGOOD PRACTICE— Steady
Read the full technical briefing (1,771 words) →

Where AI Stands in Education & Learning

Education is the domain where AI's most obvious benefit can be its most damaging side effect. In most fields, removing effort is the whole point. In learning, the effort is often the thing being produced. The evidence now splits cleanly along that line. When AI takes work off teachers and administrators, the gains are measured and repeatable. Oak National Academy's Aila cut Key Stage 2 preparation time to 76% of the comparison group's in a school-randomised trial of 464 teachers, with no loss of resource quality. Gradescope supports more than 13,000 instructors across 500+ universities. NYU's admissions chatbot handled 55,000 messages a year and resolved 96% without escalation, although nobody measured whether its answers were accurate. When AI takes work off learners, results depend on design and often go negative. In a 405-pupil trial, pupils preferred LLM-only summaries, yet those summaries produced the worst comprehension and memory of any condition. In a University of Maryland randomised trial of 2,379 undergraduates, access to an optional GPT-4o course tutor cut final grades by 0.37 SD, and first-generation students lost the most.

Momentum sits with mature adaptive systems and with tutors that are deliberately constrained. McGraw-Hill's ALEKS has more than 7 million users. Independent trials of adaptive maths platforms cluster between 0.04 and 0.37 SD, which is modest but real. Guarded Socratic designs have produced gains in pre-registered trials: Gemini's Guided Learning added 0.258 SD in maths across 1,763 Sierra Leonean students. The binding constraint is not model capability. It is whether students use the tools at all. Khan Academy says only 15% of students with Khanmigo access engage regularly. In Stanford's SCALE trials, 40–47% of elementary students never logged in. Teachers used Aila for 7% of lessons, and use fell over the ten-week trial. The tools are almost everywhere and lightly used.

Progress stalls wherever AI makes a consequential judgement about a student: summative exam items, essay grades, misconduct findings, admissions decisions and dropout flags. AI-writing detection is in open retreat across research universities in the US, UK, Singapore and South Africa. Prompted LLMs still fall short of human agreement on essays. Early-warning models look accurate overall but miss most of the students they exist to find. Governance trails practice. Student Defense found that none of the 20 institutions answering its records requests had a policy specifically governing AI in undergraduate admission decisions. Gallup and Walton found that 69% of K-12 teachers had received no guidance on AI-powered tutoring. Education also differs from its neighbouring domains in who carries the risk: minors, and students with little power to contest a decision. That is why political pressure is now reversing defaults. New York City has restricted student-facing AI through 8th grade, and Google is withdrawing default-on access for school pupils. The settled pattern is AI as a reviewed drafting aid for staff. AI that acts directly on students is being pulled back.

What's New, 2026-10-23 to 2026-11-06

The biggest move of the fortnight was a retreat from defaults. Google will switch off Gemini, NotebookLM and Gemini in Classroom by default for K-12 student accounts, starting in the US, until administrators turn them on. The decision follows questions from Senators Blumenthal and Hawley and restrictions planned by New York and Los Angeles. Evidence on tools that judge students also hardened against them. A benchmark of 22 prompted LLMs found 18 scoring short answers at 102–119% of human inter-rater agreement. On essays, none reached the published human agreement of 0.745, and essay length alone beat 16 of the 22 models. Fourteen models raised grades by more than 5% of the scale when a note was planted in the answer, and seven gave credit for blank or meaningless answers. Accelerate published an independent evaluation of EnlightenAI across nine Uncommon Schools high schools. Students submitted 4,846 assignments, almost none revised, and the tool had no detectable effect on writing or AP US History scores. In question generation, An and Wang's audit of 311 deployed items found AI difficulty labels correlated at ρ = 0.06 with how students actually performed. A trial with 140 urology residents found supervisor-selected AI exam items significantly easier than human-written ones. On detection, an analysis of 55 policy documents found that 16 of the 23 taking a position reject detectors. A 90-study review concluded that false-positive rates cannot be pooled, and calculated that a 1% rate would wrongly flag more than 223,000 first-year US essays a year. In admissions, UNC Chapel Hill dropped Slate's Reader AI essay scoring for 2026-27 because AI-written essays pushed its ratings up.

On conversational tutoring the picture held steady rather than worsening. New reporting on the Maryland trial puts the grade loss at 0.11 points on a four-point scale and adds that students went to office hours less. Against that, a randomised trial of 1,328 secondary students in Benin City, Nigeria, found a teacher-guided package built on Microsoft Copilot raised assessment scores by 0.310 SD, although the AI's own contribution cannot be separated from the curriculum and pair work around it. The accessibility evidence was sobering. With four deaf ASL users, Zoom's automatic captions produced a comprehension gain of 18.7%, against 41.7% with a live interpreter. Clova Note and Plaud reached word error rates of 42.78% and 54.33% on a Korean maths lecture, against about 3–4% on an English talk. Mesquite ISD's IEP tool, built with Panorama, cut drafting from up to three hours to about 45 minutes per student, and the district did not tell families. In simulation, an npj Digital Medicine pilot trial with 58 students found no OSCE advantage for LLM standardised patients over human ones. A year-long UK trial of SimPatient found no significant between-group gain either. A Burkina Faso early-warning model built on records for 11,365 pupils saw its dropout F1 score fall from 0.50 in the first term to 0.26 in the second and at most 0.03 in the third. CodeSignal made Agentic Skills Intelligence generally available, citing more than 21 million evaluations and no independent accuracy data. The DWP's AI Work Assistant offers skills mapping to several thousand private-beta users without requiring human review of each interaction.

Key Tensions

  • Teacher time against learner effort. The same capability that safely saves staff time harms learning when students use it in place of thinking. Aila and ChatGPT lesson planning cut preparation time with no loss of quality, but unguarded chatbot access left Turkish students 17% lower on a later closed-book exam, and optional tutor access cost Maryland undergraduates 0.37 SD. Products rarely distinguish the two uses, so who controls the tool matters more than the tool itself.

  • Good output, ignored by students. Feedback and tutoring quality are no longer the bottleneck. Whether students take them up is. In the EnlightenAI study no student revised more than one assignment using the feedback, Tennessee students used Khanmigo in only 17% of error moments, and only 6% of Deer Valley students used the Goblins tutor for half an hour a week. More elaborate prompts raised the share of feedback comments offering solutions from 6.5% to 45.5% without changing how often students acted on them.

  • Agreement is not validity. Headline accuracy keeps hiding what matters. Prompted LLMs match human graders on short answers but are fooled by a planted note, AI difficulty labels describe the prompt rather than the item, and deep knowledge tracing models can raise AUC while their predictions become less coherent. Early-warning systems show the same gap: seven models on 453 Estonian students reached 74–78% accuracy but caught only 8–32% of low performers.

  • Human review: required, costly and weak. Every credible deployment depends on a person checking the output, and that check erodes the savings while proving less reliable than assumed. Karnataka curators had to change 94.4% of Kannada Shiksha Copilot content blocks against 3.9% of English ones, and Greek teachers corrected harsh grades 22% less often when told they came from AI. Meanwhile the DWP's skills assistant runs without mandatory per-interaction review, and Mesquite ISD did not tell families about its IEP tool.

  • Vendors expand while institutions withdraw. Features keep multiplying: Claude for Teachers, Canvas's IgniteAI module agent, and Pearson's agreement, disclosed in its trading update for the nine months to September, to acquire Workera. Institutions are switching defaults off. Newcastle and Westminster are leaving Turnitin, NYC has restricted student AI through 8th grade, and Google is ending default-on pupil access. Products are reaching schools faster than schools are agreeing to let them reach students.

Top 10 Evidence Items

  1. Google to switch off default AI access for school students (news-coverage) — This is the clearest instance of the summary's claim that political pressure is reversing default-on AI access for minors. https://theprint.in/feature/google-to-switch-off-default-ai-access-for-school-students/3091449/
  2. Beyond agreement: a benchmark of how language models grade student work, and how they fail (research-paper) — Shows the agreement-is-not-validity tension directly: near-human scores on short answers collapse into gameable, unreliable essay grading. https://edtechdev.github.io/aied/articles/porcaro-ai-grading-benchmark-2026/
  3. EnlightenAI Evaluation: AI Writing Feedback Tool | Accelerate (industry-report) — A rare independent, large-sample test of AI feedback finding zero effect because students simply don't revise — the uptake bottleneck in concrete numbers. https://accelerate.us/research/enlightenai-ai-writing-feedback-tool-evaluation/
  4. aied/content/ko/faqs/check-ai-generated-question-quality.md at main · edtechdev/aied (tutorial) — Quantifies how little AI-generated difficulty labels track real student performance, undercutting trust in AI judgement of assessment quality. https://github.com/edtechdev/aied/blob/main/content/ko/faqs/check-ai-generated-question-quality.md
  5. AI content detection in student text assignments, false positive rates, and academic integrity (research-paper) — Puts a hard number on the cost of detector false positives, explaining why institutions are abandoning AI-text detection en masse. https://edtechdev.github.io/aied/articles/maghami-ai-detection-false-positives-2026/
  6. Dartmouth admissions office has no plans to screen applicant essays with third-party AI detectors (news-coverage) — A concrete admissions reversal showing institutions pulling back from AI judging students once it visibly distorted outcomes. https://www.thedartmouth.com/article/2026/11/holty-ai-in-admissions
  7. Adaptive Generative Tutoring in Resource-Limited Secondary Education: Effects on Participation, Achievement, and Digital Equity in Nigeria (research-paper) — Illustrates how positive tutoring results keep arriving entangled with curriculum and human scaffolding, so the AI's own contribution stays unproven. https://www.mdpi.com/2227-7102/16/11/1780
  8. From prediction to early policy action: machine learning early warning for primary school dropout in Burkina Faso's EMIS (research-paper) — Shows an early-warning system's accuracy decaying sharply as the school year progresses, undermining its use for the students it's meant to flag. https://www.progres.ujkz.gov.bf/publications/15405/details
  9. DWP: Work Hub (case-study) — A live case of a consequential tool — skills assessment feeding employment support — running without mandatory human review, matching the governance-lag tension. https://www.gov.uk/algorithmic-transparency-records/dwp-work-hub
  10. Does automatic captioning provide the same access as an interpreter? (research-paper) — A peer-reviewed, if small, result showing automated accessibility support still falls well short of human interpretation for disabled learners. https://academic.oup.com/jdsde/advance-article-abstract/doi/10.1093/jdsade/enag075/9013407