Retrospective classroom notes / 100 notes
Guidance, studies, tools, policies and the older precedents behind AI in schools, each read from the classroom outward: what it costs, who it trains, who it reaches and where it fails.
Historical event dates and source dates are separate from the preparation date of this local edition. Every note is a retrospective draft prepared 16 September 2026; none was published on its historical date.
100 notes
Skinner's own 1958 account of Pressey's device shows what immediate feedback changed, and why the machine failed to scale.
Skinner's 1958 paper and its 1954 companion argued for machine pacing, leaving motivation largely undiscussed.
The university's own history dates PLATO to 1960 and ties its 1980s decline to cheaper microcomputers.
Papert's 1980 Mindstorms and the Logo Foundation's own history set out constructionism without measuring it.
RAND's 147-school study of Cognitive Tutor Algebra I found no year-one effect and a costed year-two gain.
NCTM's 1989 standards and later calculator and technology statements show the same tool-versus-skill argument aging.
Khan Academy's own pages date its 2005 start and give the funded, donor-based scale Khanmigo now sits on.
A joint Harvard-MIT report's four engagement tiers show enrolment and completion measuring different things.
A 318-school randomised evaluation of One Laptop per Child found no math or language test-score effect at 15 months.
Pooled effect sizes from 2014 and 2016 show intelligent tutors matching tutors on narrow tasks, not generative ones.
The US education department's first AI report names five recommendations and says its own guidance does not bind schools or vendors.
England's department set a non-binding school position in 2023 and later safety standards addressed mainly to AI vendors.
UNESCO's 2023 guidance urges an age limit and human oversight for classroom AI without binding any country to a figure.
A 2023 national framework named six AI principles for schools; a 2025 review found gaps in workload, market power and equity.
MEXT's 2023 provisional guideline and 2024 revision reason through examples and leave the AI age limit to vendors' own terms.
The AI Act names four education uses as high-risk and sets duties for both the system's builder and the school that deploys it.
Executive Order 14277 creates a task force and tells agencies to redirect existing funds toward AI, without new money or a school mandate.
A July 2025 department letter says AI spending is already allowable under existing grants, paired with a still-proposed new priority.
A 2025 law amendment reclassified Korea's AI digital textbooks from certified textbooks to teaching materials.
A randomised crossover trial in one Harvard physics course found an engineered AI tutor produced larger test-score gains than active learning.
A randomised field trial found unrestricted GPT-4 raised practice scores but lowered exam scores, while a hint-only version avoided the drop.
A World Bank-evaluated randomised pilot found an AI tutoring programme in Edo State raised assessment scores after six weeks, with caveats on scale.
Pew's survey of US teens found reported ChatGPT use for schoolwork rose from 13% to 26% in a year, with gaps by race and grade level.
RAND surveys of teachers and principals found AI use rising unevenly, with higher-poverty schools reporting less use and less guidance.
England's evidence body found a cluster-randomised trial where ChatGPT-using teachers saved 25 minutes a week, with no pupil outcomes measured.
Researchers found AI-text detectors flagged 61% of non-native English essays as AI-written, versus 5% of native essays, in a small pilot study.
Challenge Success surveys found self-reported cheating rates stayed flat after ChatGPT's release, though a stable rate is not a clean verdict.
An unreviewed MIT preprint found weaker brain connectivity among 54 participants using a chatbot to write essays, with a caution against overreading it.
Khan Academy's published Khanmigo results are internal engagement metrics and pilot-district accounts, not independent, peer-reviewed evaluation.
Khan Academy's own posts show a GPT-4 pilot narrowing into a paid Khanmigo product for adults 18 and over.
OpenAI's own announcement and product page detail data controls for ChatGPT Edu but state no price.
Google's own help pages set ages for personal and work accounts but leave school access to administrators.
Microsoft's documentation covers data protection and licence tiers for education but not a minimum student age.
Anthropic's own launch post describes Learning mode's Socratic prompts and three campus-wide deals.
MagicSchool's own pricing and privacy pages set out free and paid tiers and named compliance certifications.
Turnitin's own launch and false-positive posts state reach, confidence and a claimed error rate.
OpenAI's release and educator posts state a 26% detection rate and a later withdrawal for low accuracy.
Chegg's SEC-filed results state a March 2023 spike in ChatGPT interest slowed new customer growth.
Google's launch and help pages describe grounding meant to cut errors, plus an AI-voiced audio feature.
The chancellor's own account and the department's initial statement show why the ban came and went so quickly.
Government documents trace the shift from local caution to a jointly endorsed national approach in under a year.
The board's own statement and its academic integrity guidance show what citing a chatbot means in a graded essay.
Joint guidance sets malpractice rules for AI in coursework and lists checks a teacher can actually perform.
Policy pages show a stage-based rule for AP Seminar and AP Research, enforced through teacher checkpoints.
TEA's own scoring-process and Q&A documents describe the hybrid model, its cost driver and its rescore path.
The regulator's own approach paper leaves specifics to the exam boards it oversees, including JCQ.
GPTZero's marketing claims near-perfect accuracy; a peer-reviewed test scored it 54 percent, mid-pack among 14 tools.
A Massachusetts memorandum and order describes the process a district followed and why a judge found it adequate.
The FTC's 2025 COPPA amendments keep the school-consent shortcut vendors use but add a retention-and-deletion policy duty.
FERPA's school-official exception, not a vendor's privacy notice, decides whether an AI tool may hold student records without parental consent.
The children's code applies to edtech used through a school whenever a vendor shapes children's data use beyond the school's instructions.
Ed launched with board fanfare in March 2024; by June the vendor had furloughed staff and the district pulled the tool offline.
CDT's parent, student and teacher surveys find student activity monitoring is used more for discipline than safety and falls hardest on some groups.
The TAKE IT DOWN Act criminalises nonconsensual intimate depictions, including digital forgeries, and gives platforms a year to build removal.
Australian and UK guidance treat an AI-faked nude as an indecent image and set out what a school, not just a platform, should do next.
A 2024 order let E-Rate pay for student hotspots off school premises; a 2026 Federal Register order rescinds that reading of the statute.
NCES data show 97% of children have some home internet, but device type, income and disinterest complicate what access means.
UNICEF's 2021 guidance gives schools and vendors nine child-centred AI requirements, later expanded to ten in a 2025 revision.
NEA's 2024 task force report proposes principles on human judgment, evidence and equity, without a binding classroom rule.
AFT's $23m Academy, backed by Microsoft, OpenAI and Anthropic, aims to train 400,000 educators within five years.
Oak's transparency filing names its model, data and human checks, detail rarely public for a classroom AI tool.
England's £4m content store aims to train AI on curriculum data; a follow-up cites a developer-estimated time gain.
ISTE, now ISTE+ASCD, publishes an AI-ready graduate profile and paid AI courses, without one named AI standard.
TeachAI's 2023 toolkit, backed by five organisations, warns explicitly against copying its sample language verbatim.
TALIS 2018 found US teachers taught 28 of a 46-hour week; that split is the baseline any saved-time claim must beat.
EEF rates feedback high-impact from 155 studies, while a leading grading tool claims half the time with no cited basis.
A widely cited tracker, run by a firm that also sells AI training, counts 35 states with K-12 guidance in 2026.
NCES found 74% of US schools faced hiring difficulty in 2024; a staffing gap differs from what a workload tool solves.
ESSA's evidence-based definition gives districts a precise question to put to any AI vendor: which tier, and what study.
AB 2876's enacted text only requires AI literacy consideration whenever three frameworks are next revised.
Tennessee's 2024 law and Ohio's 2025 budget bill both mandate a policy's existence, with very different enforcement.
The AI Leap 2025 announcement named prospective partners; the delivered pilot and its uptake looked narrower.
The CSU's own announcement names ten vendor partners and system-wide scope but discloses no contract cost.
ASU's January 2024 announcement named intentions; its later reporting gave proposal, acceptance and licence totals.
1EdTech's own catalogue separates roster and grade interoperability certification from its privacy and AI rubrics.
The DfE's 2025 product safety standards give schools clauses to demand of vendors, with no independent verifier.
Published policies from Maryland, Philadelphia and New York show governance, product and age-based approaches.
Board circulars show a graded Class IX subject and a separate, ungraded pedagogical layer sharing one name.
A randomised EEF trial cut planning time with ChatGPT, while JCQ guidance and a detector's own limits mark where assistance ends.
A Turkish RCT found an unguarded chatbot cut later exam scores while a hint-only version did not, unlike a vendor's design claims.
Civil-rights law requires notices families can understand; department toolkits cover four languages, leaving AI tools to fill the rest unevaluated.
IDEA and England's SEND code require documented, reviewed plans; neither names AI, so the old process still applies to new tools.
WHO caps early-years screen time, AAP asks for a family plan, and UNESCO proposes an AI age floor, but none covers an AI toy for toddlers.
Internet Matters and Common Sense Media both treat homework help as legitimate AI use that still needs a parent's verification.
Code.org, MIT RAISE and Common Sense Education all teach AI literacy from around grade five, leaving younger children uncovered.
England's 2026 phone-ban guidance and a US federal survey address personal phones only, leaving AI on school devices unaddressed.
DfE sets AI boundaries for schools while Oak National Academy supplies actual prompts, and its own moderator still checks the output.
Pew, Common Sense Media and RAND surveys show broad teen AI use and a school guidance gap that tracks poverty, not use itself.
Reading three AI tools' privacy policies side by side shows what each actually promises about training and data retention.
US and UK evidence frameworks define four tiers behind the phrase evidence-based, from randomised trials to a stated rationale.
A vendor's own false-positive claim, an independent bias study and exam-board guidance together show why one score should not decide a case.
Ofqual, JCQ and the IB show supervised exams stay protected while unsupervised coursework needs new citation rules.
Product safety standards, interoperability standards and federal FERPA guidance turn vague vendor promises into checkable procurement questions.
TeachAI, Australia's national framework and DfE guidance name similar practices but differ on whether a school needs one new AI policy at all.
A funder's evidence tiers, a Harvard RCT and an EEF school trial show what a genuine trial needs that a one-term pilot usually cannot supply.
Khan Academy, MagicSchool and OpenAI publish per-student, per-teacher and no price at all for their education AI products.
A national teen survey finds most students report no teacher discussed safe AI use, and blocked tools are often moved to a personal device.
Google, Microsoft and OpenAI's own admin documentation show three different default settings for age, data training and access.
Singapore's ministry states four AI commitments and routes most classroom AI through its own Student Learning Space platform.
The testing body's own 2005 paper and current pages describe e-rater's role beside a human reader, not instead of one.
Try another category or clear your search.
