Education & Learning — Biweekly Brief
The headline: Education has finished adopting AI without establishing that it helps. This month the US Department of Education's own evidence review concluded that general-purpose AI use is linked to worse learning.
The Picture
Almost every school, university and corporate training function now has AI somewhere in the teaching workflow. A survey of 1,200-plus US school principals in August put adoption at 90% — and satisfaction at 30%. Teachers report saving 5.9 hours a week, but only 18% work under any formal district guidance. A small group of organizations is getting real learning gains, and they share one trait: their systems are built to withhold answers and make the learner work, with a teacher actively driving use. Everyone else has bought distribution and is discovering that availability is not adoption, and adoption is not learning. The window to get the design right is closing, because regulators and courts are starting to answer the question for you.
This Fortnight
- The US Department of Education's research arm delivered a verdict. Its rapid review of twenty rigorous causal studies found teacher-mediated AI tutoring roughly matches conventional instruction, student-facing tools produce mixed results, and general-purpose AI use hinders learning. If you are funding AI in teaching or training, this is now the citable baseline your board and your auditors will read, so make sure your business case rests on something more specific than "we deployed it."
- New South Wales ended unsupervised take-home assessment. Australia's largest state made roughly half of every senior school qualification grade unenforceable and moved it into supervised rooms, citing the failure of AI detection tools for the 95,000 students in the 2027 cohort. Any organization that grants credentials on unsupervised written work should assume the same reasoning applies to it and start costing the redesign now.
- Independent testing showed AI tutors default to giving answers away. A benchmark of seven leading models against 462 real math tutoring transcripts found all of them over-help when simply told to tutor; a separate test of six models found that without tutoring-specific instructions they handed over direct answers 97% of the time. Buy on how a vendor constrains the model, not on which model is inside it.
- Detection tools moved from unreliable to litigable. A New York court annulled a cheating finding based solely on a 100% AI score after independent detectors scored the same essay at zero, and a Yale MBA student has filed a thirteen-count federal suit over a false positive tied to $208,500 in tuition. Independent testing of one leading detector found a 13.8% false-positive rate on verified human writing against a vendor claim of 1%. If your integrity policy still treats a detector score as evidence, it is a liability.
- The biggest platforms pushed AI to everyone at once. Google began rolling Gemini out to 150 million school users and OpenAI shipped dedicated educator and student products. Meanwhile a Chalkbeat investigation found AI-generated lesson materials containing factual errors and missing alphabet letters selling at scale on a marketplace used by 85% of US teachers, which is what unmanaged distribution looks like from the classroom floor.
Coming Up
- EU AI Act obligations are staggered, and the easy part is already live. Transparency duties applied from August 2026; the high-risk rules covering automated grading now bite on 2 December 2027. Only around a quarter of institutions have adequate AI governance policies today, so treat the delay as budget time, not relief.
- Accessibility deadlines are the harder near-term compliance test. US rules require WCAG 2.1 AA conformance by April 2027 for large public entities and 2028 for smaller ones, and only 14% of school districts met the earlier April 2026 milestone. AI captioning gets you a draft, not compliance — real-world error rates run 12–15% against vendor lab claims of 98% — so budget for human correction.
- Admissions and enrollment automation is running well ahead of its governance. Use jumped from 29% to 68% of major universities in two years, yet a public records survey found zero of 24 major public universities had any written policy or staff training for AI in undergraduate admissions. One Mexican university's AI-proctored entrance exam ended in 58,000 mandated retakes. Write the policy before the incident, not after.
What's Hard About This
- The savings are immediate; the damage shows up two exams later. A Chinese grading deployment cut teacher marking from forty minutes to ten while exam scores in the tracked student cohort fell 20% over six months. A trial with 193 teachers found AI teaching assistants reduced student motivation and achievement, worst where teachers used the output without editing it.
- Getting value requires orchestration you have to supply yourself. The deployments that work pair constrained AI with active human prompting: a trial in Sierra Leone hit 69% engagement and meaningful math gains, while two US trials found nearly half of students never logged in and the rest used the tool two to five minutes a week against the thirty needed to move the needle.
- Fairness problems are structural, and the federal remedy just narrowed. Speech recognition still errs on Black speakers at nearly double the rate for white speakers, and assessment systems scored identical resumes carrying disability-related honors lower 75% of the time. In July the US Department of Education removed the disparate-impact route for Title VI investigations, so the burden of catching these failures now sits squarely with the institution deploying them.
Go deeper: the full Education & Learning briefing — the longer analytical write-up, plus every practice we track in this domain with its maturity rating, the tools to consider, and the evidence behind our assessment.