The State of Play

Adaptive tutoring that personalises pacing and difficulty is operationally mature and widely deployed as a supporting practice—but the field faces an entrenched implementation divide between well-designed systems and unsupervised approaches. June 2026 evidence exposes the hard reality: institutional adoption scales while pedagogical impact stalls. A Guinness-certified RCT (1,662 5th-6th graders, Squirrel AI) shows concrete outcome signals—AI groups scored 8.78–13.84 points higher than traditional teaching, with 29-percentage-point gains in top-performer rates. Yet Wharton RCT evidence (770 students, 5 months) demonstrates the implementation boundary precisely: adaptive problem sequencing yields 0.15 SD gains (6–9 months equivalent) when pedagogical guardrails remain intact, while unrestricted AI access produces learning loss. Independent academic deployments confirm capability: Iowa State's AI tutor (40% voluntary adoption) yielded +4.6pp final grades (+9.1pp for heavy users). However, critical barriers dominate the landscape: Gallup/Walton survey of 2,000+ K-12 teachers found 71% received no formal guidance on AI-powered tutoring or feedback—explaining why teacher skepticism persists despite evidence. Khanmigo's documented "quiet collapse" (founder Khan: "was a non-event" for many students) illustrates why adoption announcements mask technical limitations (session amnesia, absent learner memory). A McKinsey study across 10,000+ organizations revealed the scaling paradox: 88% report deploying AI, but 86% lack operational readiness—implementation infrastructure, not technology, determines viability. The practice has converged on a hybrid human-AI model—adaptive pacing as blended-learning support under high-fidelity conditions with teacher oversight, not autonomous tutoring. Market evidence shows AI personalized learning at $6.1B (2025) heading toward $9.4B by 2028, yet systematic failures (EdTech practitioners document that "algorithmic pacing with limited human judgment" falls short of personalization claims) and organizational adoption barriers remain binding constraints on scaled pedagogical impact.

McGraw-Hill's ALEKS dominates the institutional market with 7M+ global users and 2.7x proficiency likelihood for students reaching mastery thresholds, backed by 25 years of research. Institutional vendor data shows Pearson AI-powered adaptive practice platforms delivering 90% improvement in proficiency achievement versus static content, while independent deployments confirm viability: Siyavula (160,000 students) demonstrated that UI design interventions—written prompts (+2% persistence) and visual nudges (+9%)—measurably improve student persistence through failure recovery; Efekta serves 4M students; EduAdapt serves 1.8M+ students across 250+ universities; Compass serves 50K children with 500M+ learning interactions. Large-scale university deployments show promise: a knowledge-grounded LLM feedback system deployed at scale (>1,000 students) yielded 80% improvement in student performance versus prior semesters. Market valuation places AI personalized learning at $6.1B (2025) growing toward $9.4B by 2028, with K-12 adoption surveys showing accelerating institutional technology deployment.

However, June 2026 evidence documents a critical split between capability and deployment. A meta-analysis of 36 empirical studies (2023–2025) confirms generative AI feedback effectiveness (effect size 0.61) but reveals implementation context determines outcome: collaborative/self-directed learning achieves ES 0.68–0.71, yet direct instruction yields ES≈0—documenting cognitive offloading as a binding constraint. Critical institutional barriers dominate: Gallup/Walton Foundation study of 2,000+ K-12 teachers found only 29% received guidance on AI-powered tutoring or feedback coaching (69% received no guidance on one-on-one instruction specifically); fewer than 1 in 10 received written official guidance on any AI activity. An analysis of organizational adoption patterns (McKinsey, n=10,018 organizations) revealed the scaling paradox: 88% declare AI deployment but 86% are not operationally ready to adapt AI into daily operations—implementation infrastructure (decision authority, role definition, measurable outcomes, delegated autonomy, ownership clarity) determines viability, not technical capability. Practitioner research documents why high-visibility deployments fail: Khanmigo adoption remained minimal (Khan Academy founder's public admission that AI tutoring was "a non-event" for many students), with technical failures (session amnesia, absence of persistent learner memory) preventing adaptive systems from maintaining context across interactions. Systematic scope limitations persist: research shows AI tutoring addresses only ~16% of human learning development, with adaptive systems showing reduced metacognitive activity and no transfer without human support. Professional development gaps, cost barriers (70%+ of under-resourced institutions face implementation challenges), algorithm bias concerns, and documented equity gaps remain binding constraints on scaled impact.

• 2018: Adaptive learning platforms (ALEKS, Knewton Alta, CogBooks) transitioned from pilot to institutional production deployment. ASU achieved 20-point success rate improvement in college algebra; Knewton Alta adoption reached 250 institutions with 87% mastery rates in vendor data. Peer-reviewed research documented persistent barriers to adoption.

• 2019: Evidence of both capability and limitations emerged. Third-party validation (Johns Hopkins) confirmed efficacy; Squirrel AI demonstrated large-scale commercial viability (2M+ students, 2000+ centers in China). Practitioner case studies revealed critical failure modes (ChalkTalk's 8% success rate) and evidence of cognitive biases in learner self-evaluation within ALTs. Expert analysis (RAND) narrowed appropriate scope to structured domains and teacher-supporting roles. Market correction evident as Knewton collapsed (raised $180M, sold for ~$10M fire sale).

• 2020: Platforms consolidated around incumbent publishers and established scale in specific markets. ALEKS deployed across international institutions (Philippines case study with 82% confidence gains); Squirrel AI reached 2.6k learning centers serving 2M students in China. Systematic reviews confirmed research maturity. Efficacy evidence (Korbit study: 2-2.5x learning gains) competed with evidence of persistent limitations (human-in-the-loop models required, effectiveness limited to narrow domains). Market consolidation accelerated (McGraw-Hill with ALEKS, Wiley acquiring Knewton assets).

• 2021: Research and deployment evidence continued to accumulate despite practitioner skepticism. Peer-reviewed studies confirmed adaptive learning efficacy across multiple modalities (speech and typing interfaces); Taiwan research documented achievement gains in online courses. ALEKS expanded institutional footprint with multi-campus deployments; Knewton's scaled metrics (15B+ personalized recommendations) contrasted with its business collapse (acquisition at steep discount), reinforcing evidence that technical capability does not guarantee market viability. Critical practitioner voices challenged hype and questioned whether adaptive personalization genuinely replaced human interaction. Stakeholder research (23-participant study) documented persistent concerns about social boundary violations despite recognition of personalization benefits.

• 2022-H1: Evidence consolidated around realistic capabilities and limitations. ALEKS showed measurable K-12 gains (5.1% increase in highly proficient scores on state assessments), with stronger effects for disadvantaged students; Knewton Alta received industry recognition (second SIIA CODiE award). Simultaneously, peer-reviewed research documented clear constraints: teacher-led instruction outperformed ALEKS for underachieving students; integration exceeded teachers' digital competence in primary classrooms. Engineering advances (neural network assessment improvements) demonstrated platform maturation, while growing research on implementation infrastructure (teacher dashboards, professional development) indicated the practice required substantial supporting ecosystem to work effectively. Consensus solidified that adaptive pacing is a powerful supporting tool for specific subjects and contexts, not primary instruction.

• 2022-H2: Academic engagement and practitioner perspectives dominated evidence. Peer-reviewed research topic in Frontiers on AI personalization techniques signaled sustained scholarly attention; practitioner surveys from Finland revealed skepticism about teacher replacement but acknowledged self-directed learning benefits. Deployment-level evidence strengthened: Every Learner Everywhere report showed 26,400 students across 12 institutions with equity gap narrowing by third term of adaptive courseware use. Implementation research (Tutti dashboard study) documented real-world integration of adaptive systems requiring teacher oversight and real-time monitoring, reinforcing human-in-the-loop necessity. News coverage reflected broader AI tutor adoption with persistent efficacy doubts, especially for language learning. By year-end 2022, adaptive pacing had shifted from growth narrative to established practice with well-understood strengths (structured domains, disadvantaged student support, measurable gains) and clear deployment prerequisites (teacher training, dashboard support, subject-matter fit).

• 2023-H1: Platform capability advancement continued alongside broader AI adoption narratives. McGraw-Hill deployed neural network enhancements to ALEKS in April, achieving 20% reduction in assessment time and 9% material mastery increase, signaling ongoing algorithmic refinement in major platforms. Industry adoption metrics showed acceleration: 25% of global education organizations reported successful AI implementation (up from 14% in 2019), with intelligent adaptive learning identified as a disruptive technology. Implementation practice data from lighthouse institutions revealed gaps in equity-focused adoption: only 25% of faculty used student performance data for weekly instructional modifications and 38% integrated adaptive learning with inclusive teaching practices, indicating implementation infrastructure remained limiting factor. Consistent with prior windows, evidence affirmed both capability (measurable gains in structured domains) and constraints (teacher-led instruction remained superior for underachieving students; full automation ineffective without pedagogical oversight).

• 2023-H2: Research validation strengthened alongside deployment evidence, establishing clearer boundaries for effective AI-tutoring implementation. Carnegie Mellon's three-school quasi-experimental study demonstrated that hybrid human-AI tutoring models work at scale (585 students, $700 annual cost) with lower-achieving students showing greater gains, validating the teacher-augmentation hypothesis. However, critical limitations became more visible: University of Würzburg research showed large language models tested as standalone tutors achieved only 82% accuracy on thermodynamics questions—well short of the 95% threshold needed for reliable unsupervised tutoring, indicating AI-only approaches remain immature. Industry metrics from McGraw-Hill reinforced platform consolidation, with ALEKS showing 14% year-to-date activation growth and digital revenue at 61% of total business, reflecting institutional commitment despite maturing competitive landscape. Academic perspectives (ECER 2023) questioned whether adaptive learning was driven by genuine pedagogical evidence or primarily by commercial and policy promotional narratives. By year-end 2023, the field had solidified around a hybrid human-AI model with well-documented scope boundaries (structured domains, teacher oversight required), clear cost economics ($700/student), and specific deployment prerequisites (professional development, monitoring infrastructure). Full automation had effectively failed; personalized adaptive pacing had become a supporting tool within blended learning ecosystems rather than a replacement for instruction.

• 2024-Q1: Research validation broadened beyond mathematics and science. A meta-analysis of 27 studies confirmed adaptive learning's positive effect on reading literacy (g=0.29), extending evidence of efficacy to new domains. Real-world deployment data showed EIDU's adaptive learning tool serving 225,000 students across 4,000 schools in Kenya, demonstrating viability in low-resource settings with algorithmic pacing and feedback. Market consolidation continued with McGraw-Hill ALEKS maintaining 14% activation growth and 61% digital revenue share. However, critical expert assessments reinforced maturity boundaries: technologist Satya Nitta (IBM Watson) documented a five-year failed attempt to build generalized AI tutors, citing hallucination and fundamental limitations requiring human oversight. New market entrants (Eureka Labs, launched by OpenAI researcher Andrej Karpathy; PeTai; OIAI) announced personalized tutoring platforms, though without independent deployment validation. The consensus remained stable: adaptive pacing is a proven supporting tool for structured subjects within blended learning, effective for lower-achieving learners, but dependent on teacher integration and operational infrastructure; AI-only tutoring approaches had not matured beyond proof-of-concept.

• 2024-Q2: Evidence consolidation confirmed mature operational status with emerging implementation barriers. McGraw-Hill's FY2024 results documented 7M+ ALEKS users and 12% growth, sustaining scale. New research expanded efficacy evidence to adult learners (Apprentice Tutors) and higher education (ALS-LMS integration showing statistically significant performance gains); peer-reviewed adoption research from rural China identified key enablers (system quality, teacher support, computer self-efficacy). However, practitioner research surfaces critical limitations: foundry10's study of 25 classroom teachers found that while adaptive tools improved efficiency and engagement, teachers faced persistent barriers (training gaps, implementation support deficits, data reliability issues). NSF-funded ASEE study across 330 students showed small negative effects on knowledge assessment but positive classroom environment, signaling context-dependence of effectiveness. Product development continued (ALEKS Adventure for K-3 launch), and platform algorithmic improvements (neural networks for assessment) demonstrated ongoing maturity. Consensus shifted firmly toward pragmatism: adaptive pacing works at scale within defined contexts and requires substantial implementation infrastructure; autonomous AI tutoring had been largely abandoned in favor of human-AI hybrid models.

• 2024-Q3: Research and practitioner perspectives continued to validate both capability and persistent limitations. A comprehensive systematic review (2010-2025) found consistent evidence of 20% performance improvements while highlighting the complex mixed landscape requiring greater research rigor, reinforcing field maturity. McGraw-Hill Q1 FY2025 financial results confirmed sustained growth with 14% activation increase and 8% unique user growth across ALEKS and Connect platforms, pushing K-12 digital adoption to 50% mix. Critical expert voices balanced optimistic vendor metrics: learning technologist Clark Quinn cautioned against generative AI for adaptive sequencing without rich pedagogical models, while Cognitive Resonance CEO Benjamin Riley warned against AI-only tutoring, citing Wharton research showing ChatGPT math students learned less than peers. Real-world deployments continued: SkillDict's adaptive e-learning platform at Paks Nuclear Power Plant demonstrated measurable gains in knowledge acquisition and confidence through personalized pacing in rigorous safety training. By September 2024, adaptive learning systems remained operationally established at scale with well-documented efficacy in structured domains and clear limitations for autonomous approaches, reinforcing the sector consensus that personalized pacing works best as teacher-augmenting support within blended learning rather than as primary instruction.

• 2024-Q4: Evidence matured into final-quarter affirmations of both proven efficacy and well-defined limitations. Khan Academy published large-scale efficacy results (December 2024) demonstrating ~20% greater learning gains across ~350K students with 30+ minutes weekly mastery-based practice, providing vendor-scale deployment validation. McGraw-Hill's FY2025 mid-year results confirmed 10% growth in ALEKS unique users, sustained institutional adoption, and rollout of generative AI features. Independent evaluation from Digital Promise documented mixed outcomes from Gates Foundation-funded pilots (Carnegie Learning, Amplify, Discovery Education), emphasizing caution and identifying key risks of hallucinations and data inaccuracy. Peer-reviewed research continued accumulation: systematic review of ITS literature identified personalized/adaptive learning as central evolution area with persistent implementation challenges and ethical concerns; Korean study of 79 college students confirmed ALEKS-driven performance gains with variation by initial knowledge level; NSF AI Institute expert analysis balanced recognition of ITS effectiveness for well-defined tasks against critical limitations (poor performance on open-ended problems, collaboration, math accuracy still below reliable thresholds). By year-end 2024, adaptive learning systems had solidified as operationally mature, evidence-validated supporting tools for structured domains within blended learning—with clear consensus that full automation remained unfeasible and teacher-AI hybrid models represent the viable deployed form.

• 2025-Q1: Institutional adoption continued at scale with evidence of both capability maturation and implementation constraints. McGraw-Hill announced ALEKS Adventure study (2025-2026) targeting K-3 adaptive math supplementation, signaling continued product expansion through rigorous evaluation. Large-scale deployment metrics confirmed global viability: Ensar Solutions' EduAdapt platform served 1.8M students across 250+ universities in 45 countries, demonstrating multinational scale with 79% course completion rates and 9% dropout rates. Systematic reviews of Intelligent Tutoring Systems (arXiv March 2025) affirmed AI advancements in adaptability and engagement while documenting persistent challenges in ethical considerations and cognitive adaptability. Market forecasts (February 2025) projected 22.2% CAGR growth for adaptive learning through 2034 across K-12, higher education, and enterprise, identifying adoption drivers (AI solutions, personalized learning) and barriers (development costs, privacy compliance, educator resistance, interoperability). Critical perspective from industry leadership remained cautious: McGraw-Hill's chief AI officer publicly warned against full generative AI tutoring replacement of human teachers, emphasizing risks of inadequate challenge and social skill loss. Implementation infrastructure continued as binding constraint for effectiveness. By quarter-end, consensus held that adaptive pacing is operationally proven for blended learning support in structured subjects, requires substantial teacher professional development and monitoring infrastructure, and fundamentally depends on human-AI integration rather than full automation.

• 2025-Q2: Vendor integration and ecosystem maturity indicators strengthened alongside critical evidence of adoption-outcome gaps. McGraw-Hill integrated Pearson's PRoPL interim assessment into K-12 curriculum solutions (June rollout to California, expanding nationwide), signaling ecosystem consolidation toward sophisticated data-driven personalization infrastructure. McGraw-Hill's Q2 financial results documented sustained scale with 7M ALEKS global users and 18% digital revenue growth to $972M (54% of total revenue), with K-12 digital billings surging 24%, confirming institutional commitment and business viability. Market research (GII, April 2025) projected market expansion from $4.03B (2024) to $14.12B by 2030 (23.19% CAGR). However, critical limitations surfaced in independent research: a cross-sectional survey of 300 higher education students and lecturers found moderate-favorable attitudes toward AI tutoring but less-significant positive correlation with learning outcomes, suggesting adoption enthusiasm may exceed validated pedagogical impact. Adoption barrier analysis identified persistent challenges: high development/maintenance costs, data privacy concerns, algorithm bias, and limited human interaction, reinforcing earlier findings that implementation complexity and cost remain primary constraints. By quarter-end, a clear tension had emerged between scaling vendor infrastructure (ecosystem integrations, user base expansion, revenue growth) and stagnating evidence of learning impact when adoption enthusiasm is measured against outcome validation.

• 2025-Q3: Portfolio expansion and infrastructure consolidation continued alongside critical reliability evaluations. McGraw-Hill expanded ALEKS offerings to calculus (September 2025), extending AI-driven personalized learning across full mathematics curriculum, signaling continued product maturity and confidence in core platform architecture. McGraw-Hill's post-IPO strategic pivot accelerated institutional adoption through ecosystem integration: digital revenue reached 61% of total sales with 95% adoption of digital-first model, reflecting deep commitment to AI-powered personalization infrastructure. Educator and student adoption research documented rapid integration: Michigan Virtual's 2025 snapshot showed widespread AI tool uptake in K-12 and higher education alongside persistent concerns about reliability and algorithm bias. Critical limitations remained center-stage: Common Sense Media evaluation of four major AI tutoring tools (Google Gemini, Khanmigo, Curipod, MagicSchool) identified biased outputs, hallucinations, and failure to detect missing student comprehension—confirming that production-deployed tutoring tools remain below thresholds for reliable unsupervised use. Capability validation continued: Carnegie Learning's adaptive tutoring program received competitive Evidence for Impact grant funding (160+ applications), signaling rigorous evaluation commitment. By quarter-end, the sector faced growing scrutiny on adoption-outcome alignment: institutional deployments at scale (7M+ ALEKS users, McGraw-Hill's 95% digital adoption, global platforms serving 1.8M+ students) contrasted with flat learning-outcome validation, persistent implementation barriers (algorithm bias, cost, educator resistance), and accumulating evidence that AI-only approaches remain immature despite infrastructure scaling.

• 2025-Q4: Academic and practitioner evidence crystallized fundamental practice boundaries. Peer-reviewed systematic literature review (Pelánek, International Journal of AI in Education, December 2025) documented persistent modeling challenges in adaptive learning (accuracy, cognitive load, bias prevention). Meta-analysis of 48 peer-reviewed studies confirmed STEM efficacy alongside critical limitations: over-reliance on AI reduces critical thinking and independent problem-solving, with many implementations yielding only modest gains versus traditional instruction. Founder critique of production AI tutoring tools identified specific failures: ChatGPT math accuracy only 50%, tools give answers instead of teaching, widespread student cheating (89% use ChatGPT for homework). McGraw-Hill sustained operational momentum (ALEKS serving 7M+ global users; EduAdapt platform at 1.8M+ students across 45 countries), while business consolidation (McGraw-Hill IPO, ecosystem integrations) signaled infrastructure maturity. By year-end, sectoral consensus had narrowed: adaptive pacing remains a validated supporting tool for bounded structured domains within blended learning, but efficacy is conditional on human oversight, professional development, careful subject selection, and operational infrastructure. Implementation barriers (cost, bias, educator resistance) persist as binding constraints on scaled impact. The practice had moved decisively from infrastructure scaling questions to addressing the harder problem of translating technical capability into reliable classroom outcomes within realistic deployment constraints.

• 2026-Jan: Institutional deployment momentum sustained despite persistent efficacy questions. McGraw-Hill's ALEKS remained the dominant K-12 and higher education platform (2.7x proficiency improvement for mastery-based learners, 25+ years research validation), with digital revenue mixing to 53% at McGraw-Hill and 27% K-12 order processing efficiency gains. Market remained growth-oriented: adaptive learning valued at $4.8B with 19% CAGR through 2030s, PowerSchool reaching ~50M K-12 students. Critical evaluation confirmed earlier findings: industry analysis documented K-12 implementation barriers (fidelity, professional development, device access) as outcome determinants; practitioner research identified core limitation that self-paced adaptive-only models show 80-90% dropout rates. Vendor consolidation accelerated with ecosystem integrations (Pearson assessment into McGraw-Hill K-12 solutions) signaling platform maturity. Analysis of technology deployment patterns showed adaptive pacing worked effectively for mastery-based, structured subjects under high-fidelity conditions (documented 8-percentile gains for RAND algebra deployments) but required cohort structures and human accountability layers for completion. By month-end, adaptive learning systems had become institutionalized infrastructure supporting blended learning—operationally proven yet fundamentally limited by implementation complexity and persistent efficacy questions at scale.

• 2026-Feb: International deployment evidence and critical scope analysis confirmed practice maturity boundaries. World Bank pilot in Côte d'Ivoire demonstrated adaptive learning efficacy in developing-country contexts (2,000 TVET students showed 11.2 months learning gain in math, 5.8 months in French, with greatest benefit for initially struggling students), validating transferability beyond developed markets. Ecosystem adoption signals strengthened: Coursera survey of 4,200+ faculty and students showed 95% AI tool usage with 47% citing personalized learning as key benefit, indicating broad integration. However, critical scope limitations emerged: UCL professor Rose Luckin documented that AI tutoring addresses only ~16% of human learning development (content delivery and procedure drilling), citing research showing AI-assisted learners exhibit reduced self-monitoring and metacognitive laziness with no transfer without AI support. Research analysis (SIAI) cited 2025 Nature study showing AI tutor outperformed active learning but documented hallucination risks, inequality concerns, and training gaps (76% UK, 69% US teachers lack formal AI training). By month-end, adaptive learning systems remained operationally entrenched with proven efficacy for bounded structured subjects, but critical voices and independent evidence reinforced understanding that human-AI hybrid models are required and that pedagogical scope remains fundamentally limited.

• 2026-Mar: Controlled research delivered mixed but informative signals: an AEI synthesis of empirical studies found RL-based adaptive sequencing with dynamic difficulty adjustment produced 0.15 SD learning gains (equivalent to 6–9 months progress) among 700+ high school students, while a quasi-experimental study (n=120) confirmed statistically significant performance improvements with adaptive support as metacognitive scaffold. A K-12 EdTech platform in India demonstrated the model's viability at 50,000+ students with 22% test score improvement through adaptive difficulty adjustment. Critical counterevidence reinforced scope limits: synthesis of tutoring research reaffirmed that generic AI without adaptive guardrails (ChatGPT unguided) produces 17% worse exam outcomes, sharpening the distinction between well-designed adaptive systems and unstructured AI access.

• 2026-May: May 2026 evidence crystallized why implementation design, not AI sophistication, determines efficacy. Meta-analysis of 72 AI-teaching studies (Frontiers Psychology) confirmed positive effects (g=0.586) with heterogeneity driven by "AI type and implementation context." Massive field trial on Siyavula (160,000 students, 17M practice problems; Carnegie Mellon, CHI 2026 Honorable Mention) showed UI design interventions measurably improve persistence: written prompts +2%, visual nudges +9%, combined +11%. ML optimization research (50,700 learners) showed adaptive spacing rules produced 69% longer retention and 50% higher re-engagement. Critical practitioner research surfaced why AI tutors fail at scale: Khanmigo's "quiet collapse" documents engagement barriers and knowledge-transfer gaps; 75% day-30 failure rate analysis identified five-pillar architecture requirement. Later-May evidence reinforced the implementation divide: Wharton RCT (770 students, 5 months) confirmed adaptive problem sequencing yields 0.15 SD gains (6–9 months equivalent) driven by task persistence, not content quality; RAND Cognitive Tutor Algebra I (54 high schools, 68 middle schools) documented 0.2 effect size at scale; hybrid human-AI model (635 students, grades 5–8) achieved 36% skill proficiency and 61% MAP growth. Stanford SCALE Initiative synthesis of 20 K-12 RCTs found AI improves immediate performance but gains disappear on independent assessment (EEG shows reduced deep learning activity), confirming cognitive debt as a binding constraint. UK DfE confirmed commitment to deliver AI-powered tutoring to 450,000 disadvantaged pupils by 2027, backed by £23M EdTech Testbed and 15 RCTs. Field consensus held: adaptive pacing is pedagogically effective under high-fidelity implementation with human oversight, but cognitive debt, engagement collapse, and implementation infrastructure remain binding constraints on scaled impact.

• 2026-Apr (April 10–24): Implementation divide crystallized in new evidence. Rigorous dual-RCT comparison published April 14 showed Harvard's pedagogically-designed adaptive tutoring achieved 2x learning gains while Wharton's unrestricted GPT access produced 17% exam decline—confirming implementation design as critical success factor. Meta-analysis of 81 studies (April 15) provided strong controlled evidence: effect size 1.01 for K-12 science with mastery-based adaptive scaffolding. However, high-profile deployment failures surfaced: Khan Academy founder's candid admission (April 10) that Khanmigo AI tutoring failed to achieve adoption and impact targets despite visibility; critical analyst assessment documented Khanmigo and Alpha School real-world deployment failures showing promised benefits not materializing. K-12 trend analysis mapped adoption acceleration (teacher use doubled 25% to 53% in single year) alongside equity barriers (34-percentage-point rural-urban adoption gap). Systematic review (50 studies, 2018–2025) confirmed engagement gains 18–25% and dropout reduction 28%, while identifying that 70% of under-resourced institutions face implementation barriers. Deployment evidence at scale included Efekta serving 4M students (largest AI learning deployment), Compass 50K students with 500M+ interactions, Medly 74% GCSE improvement. By late April, consensus reinforced: adaptive pacing is operationally proven for structured domains within blended learning under high-fidelity conditions, but autonomous AI tutoring remains ineffective and implementation infrastructure (teacher training, monitoring, subject selection) is the binding constraint on pedagogical impact.

• 2026-Jun: Independent RCT evidence sharpened the capability-deployment divide. Iowa State's two-year deployment (160–180 students, voluntary AI tutor use) yielded +4.6pp final grades overall and +9.1pp for heavy users — confirming adaptive tutoring efficacy under naturalistic conditions. A Guinness-certified Squirrel AI RCT (1,662 students, 5th–6th grade) recorded 8.78–13.84 point gains over traditional teaching with 29pp higher top-performer rates, and a knowledge-grounded LLM adaptive feedback system (>1,000 students) showed 80% performance improvement versus prior semesters. Large-scale field deployments added new evidence of reach: NITI Aayog's case study of Filo Edtech's Sampurna Shiksha Kavach reached 285,000+ students across Indian states with science pass rates +13.8pp and female students +18pp; CNM community college demonstrated 22–47% math course pass rate increases through ALEKS adaptive placement; and Evelyn Learning's community college deployment showed measurable retention gains for non-traditional students. McGraw-Hill's Q4 2026 earnings confirmed 7.5M users across eight AI adaptive learning tools with Rowan College reporting 47% higher exam pass rates. ETH Zurich's TutorRL research demonstrated that RL-trained open-source LLMs guiding through questioning rather than answers achieve top MathTutorBench rankings, validating pedagogically-designed adaptive approaches at the model level. However, Stanford SCALE's usage study revealed the binding constraint on engagement: despite structured weekly time, only 53–61% of students accessed scheduled AI tutoring and median weekly use was 2–5 minutes — confirming tool availability alone is insufficient without engagement infrastructure. World Bank analysis of 26,000-student Chinese longitudinal data reinforced the critical design distinction: unstructured AI use for homework produced a 1.4 SD exam performance decline (4× larger than adaptive tutoring benefits), sharpening the contrast between designed adaptive systems and uncontrolled AI access. Gallup/Walton survey of 2,000+ K-12 teachers found 71% received no guidance on AI-powered tutoring or coaching, and Brookings analysis of 67 years of EdTech research identified three critical conditions for AI tutoring success — confirming that institutional readiness, not technology, remains the binding constraint. Khan Academy's public admission that Khanmigo adoption was "a non-event" for many students — attributed to session amnesia and absent persistent learner memory — reinforced why announcement-level adoption and real pedagogical impact remain decoupled.