K-12Issue 154 records
K–12
Fresh K-12 signal pairs high-security adaptive-tutoring gains with binding state AI control architectures and a federal evidence caution. An EEF/NFER interleaved RCT of Maths-Whizz across 63 English primary schools (10,369 pupils) finds about one additional month of Star Maths progress overall (g=0.08; high security) and about two months for FSM pupils (g=0.14)—scarce causal evidence that purpose-built AI tutoring can move attainment with equity upside, subject to UK transferability limits. On the control side, California’s chaptered AB 1159 (Sept 10) prohibits operators from using covered student information and persistent unique identifiers to train generative AI or develop AI systems, while Maryland’s enrolled SB720 AI Ready Schools Act requires SEA guidance, local AI policies within 120 days, central-office AI coordinators, and MSDE procurement rubrics (effective June 1, 2026). An IES REL Northeast & Islands synthesis underscores why guardrails matter: WWC identified no AI student-outcome studies, only ~20 rigorous causal AI education studies exist, and general-purpose AI that replaces cognitive effort can hinder learning. Together: scale constrained tutors on unaided maths metrics; ship non-training DPAs and coordinator/rubric operating models before expanding student GenAI.
- Evidence records
- 4
- Cross-source patterns
- 2
- Evidence classes
- 3 government evaluation1 independent research
- Outcomes
- 1 effective1 cautionary1 emerging1 mixed
- Source freshness
- 2 recent1 new this fortnight1 older, newly relevant
- Research completed
- 2026-09-21
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Purpose-built adaptive tutoring shows high-security maths gains while federal synthesis warns general-purpose AI can hinder learning
EEF/NFER’s Maths-Whizz interleaved RCT finds ~+1 month maths progress overall (+2 months for FSM) with high security, whereas IES REL reports WWC has no AI student-outcome studies and flags general-purpose AI uses that replace cognitive effort as learning risks—so districts should scale constrained tutors on unaided outcomes, not open chatbots on engagement.
Operating questionAre we evaluating adaptive tutoring on interim/unaided maths gains while separately restricting general-purpose student AI that substitutes for thinking?
Supporting evidenceEducation Endowment Foundation (evaluation by National Foundation for Educational Research)Institute of Education Sciences — REL Northeast & Islands (Ask an Expert)
State AI floors are splitting into operator training bans and SEA-mandated coordinator/rubric procurement models
California’s chaptered AB 1159 bars operators from using covered student information and persistent identifiers to train generative AI, while Maryland’s enrolled SB720 requires local AI policies within 120 days of SEA guidance, central-office AI coordinators, and MSDE rubric-based tool evaluation—implying multi-state districts need both contract non-training controls and operating-model staffing for governed procurement.
Operating questionDo we have both GenAI non-training contract controls and a named AI coordinator/rubric workflow wherever state floors demand them?
Supporting evidenceCalifornia State Legislature / Office of the Governor (chaptered law)Maryland General Assembly / Department of Legislative Services
Evidence ledger
Full record · every source keeps its link and limitations
EEF/NFER Maths-Whizz RCT: +1 month maths progress overall; +2 months for FSM pupils (g=0.08/0.14)
In an interleaved RCT across 63 English primary schools with 10,369 pupils (Years 2–5), allocation to Maths-Whizz—an AI-enabled adaptive tutoring programme intended for 45–60 minutes per week alongside usual maths teaching—produced an average of one additional month of maths progress on Renaissance Star Maths (effect size 0.08, 95% CI 0.02–0.14; n=8,983; p=0.01) with a high EEF security rating. FSM-eligible pupils gained about two additional months (effect size 0.14, 95% CI 0.06–0.22; n=2,301; p=0.01), though that subgroup estimate carries lower security due to smaller n. Average delivery cost was about £21.51 per pupil per year for two year groups. Authors highlight personalised scaffolding and gap-filling as plausible mechanisms for stronger FSM effects.
Why it matters, evidence and limitations
- Why it matters
- Provides scarce high-security causal evidence that a purpose-built adaptive AI maths tutor can raise primary attainment—with larger gains for disadvantaged pupils—informing U.S. district tutoring procurement and equity claims while requiring UK-to-U.S. transferability caveats.
- Evidence and measured results
- Opened the EEF CloudFront evaluation PDF end-to-end via curl and pdftotext. Verified Table 1 key conclusions, Table 2 effect sizes/security/cost ratings, sample (63 schools; 10,369 baseline pupils), dosage guidance (45–60 minutes/week), and high-security language for the primary finding.
- Limitations and uncertainty
- UK primary efficacy trial (not U.S. K-12); vendor programme with school co-pay; FSM subgroup lower security; implementation fidelity and home access vary. Why now: first unarchived large independent AI-tutor RCT with equity-positive primary maths results, contemporaneous with U.S. homework-GenAI learning-cost and state AI-control debates.
California AB 1159 chaptered: operators barred from using student data/identifiers to train generative AI
Chaptered AB 1159 (Chapter 182), approved September 10, 2026, amends California student-privacy operator rules to prohibit using covered information—including persistent unique identifiers created or gathered by an operator’s site, service, or application—to train a generative AI system/service or develop an AI system. The act also enacts the Higher Education Student Information Protection Act (HESIPA) extending parallel protections to higher-education students, with higher-ed operative timing described in contemporaneous summaries as July 1, 2027. Obligations fall on covered operators rather than schools as primary duty-bearers, reshaping vendor DPAs and training-data clauses for K-12 and postsecondary edtech.
Why it matters, evidence and limitations
- Why it matters
- Creates a binding California commercial floor against training GenAI on student data/identifiers—material for multi-state districts and vendors already adjusting to Florida AI internet-safety floors and union/vendor privacy standards.
- Evidence and measured results
- Opened California LegInfo bill text page (chaptered version) via curl; verified Governor approval/filing date September 10, 2026, Chapter 182 designation, and Legislative Counsel digest language barring use of covered information/persistent unique identifiers to train generative AI or develop AI systems; cross-checked with PRC and CalMatters reporting for operative framing without substituting them as the primary citation.
- Limitations and uncertainty
- Operator-facing statute (schools not primary duty holders); some higher-ed timing is staggered; LegInfo HTML is official text presentation rather than a stamped PDF. Why now: unarchived chaptered primary signed Sept 10, missed by prior editions, now binding alongside other state AI floors.
Maryland SB720 AI Ready Schools Act: SEA guidance, local AI policies in 120 days, coordinators, and procurement rubrics
Enrolled SB 720 (Artificial Intelligence Ready Schools Act) requires MSDE to publish AI guidance for systems, educators, parents, and students; local school systems must adopt an aligned AI policy within 120 days of guidance release, designate a central-office AI coordinator, and procure AI tools consistent with DoIT requirements. MSDE must publish a rubric/evaluative tools for local AI-tool selection, incorporate AI literacy into workforce-preparation and K–12 computer science standards by June 1, 2027, provide educator/leader PD, and staff a Maryland AI Education Collaborative. The bill takes effect June 1, 2026 (collaborative provisions terminate May 31, 2028). DLS projects GF expenditure increases beginning FY2027 (~$610k) and notes local boards may face significant staffing costs for AI-experienced personnel.
Why it matters, evidence and limitations
- Why it matters
- Shows a second U.S. state model—SEA guidance plus mandatory local coordinators, 120-day policy adoption, and procurement rubrics—distinct from Florida’s internet-safety AI rule and California’s operator training ban, useful for multi-state operating comparisons.
- Evidence and measured results
- Opened the official Maryland General Assembly fiscal/policy note PDF via curl and pdftotext; verified enrolled-revised status, effective date June 1, 2026, AI coordinator and 120-day policy requirements, MSDE rubric duties, PD/collaborative provisions, and fiscal summary figures.
- Limitations and uncertainty
- Fiscal note summarizes enrolled bill requirements; local implementation depends on MSDE guidance publication timing; not an impact evaluation of learning outcomes. Why now: unarchived enrolled Maryland AI operating model, newly relevant beside California training bans and Florida floors as states diverge on control architectures.
IES REL synthesis: WWC finds no AI student-outcome studies; only ~20 rigorous causal AI studies; general-purpose AI can hinder learning
Responding to REL Northeast & Islands Governing Board urgency, the Ask an Expert team reports that IES’s What Works Clearinghouse identified no research studies on AI and student outcomes, while a 2026 comprehensive review found only about 20 rigorous education studies with causal evidence on AI impacts—mostly outside elementary settings and often postsecondary. Emerging patterns: teacher-mediated/AI-augmented tutoring may promote learning; student-facing AI tools show mixed effects; general-purpose AI use may hinder learning when it replaces cognitive effort (lower exam performance, shallower processes, weaker recall despite perceived helpfulness). Guardrails emphasize purposeful design, human-in-the-loop use, and not treating access alone as an outcome.
Why it matters, evidence and limitations
- Why it matters
- Gives SEA/LEA leaders a federal WWC-framed caution that causal AI evidence remains thin—useful counterweight when vendors cite engagement metrics or when states mandate rapid AI policy buildout.
- Evidence and measured results
- Opened the IES/REL HTML page via curl; verified August 10, 2026 byline, WWC empty-result claim, “only 20 rigorous” causal studies language, and the three emerging patterns including general-purpose AI learning harms.
- Limitations and uncertainty
- Blog synthesis rather than primary meta-analysis; August 2026 publication predates last k12 run—included with why-now as unarchived federal framing amid September state AI floors. Does not replace reading the underlying 2026 review.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Government evaluation
- A public body’s measured evaluation or documented pilot.
- Independent research
- Research conducted outside the implementing organization.