Skill acquisition & deliberate practice support
198 evidence items
AI that supports individuals in acquiring new skills through structured practice, feedback, and progression tracking. Includes practice exercise generation and personalised feedback; distinct from L&D recommendations which suggest resources rather than structuring practice.
Overview
Skill acquisition and deliberate practice support is AI that structures the practice itself, generating exercises, giving personalised feedback and tracking progression, rather than merely pointing learners at resources. It is a leading-edge practice, slowing: named deployments report real gains and major vendors ship generally available tooling, but the independent syntheses closest to analyst recognition say the evaluation base is thin and that many learners, left to choose, simply do not use the tutor. The deciding tension is design. Mastery-based workflows with human mediation produce durable learning, while unstructured assistance inflates assisted performance and erodes the skill being practised. It is worth caring about, but doing it well demands pedagogical expertise most organisations lack.
Current Landscape
Duolingo reported 58.7M daily active users for Q2 2026, up 23% year on year, with 12.7M paid subscribers. QAI Finance's analysis shows Duolingo's gross margin guided down from about 73% in Q1 2026 to about 69% in Q4 2026, which it attributes to third-party LLM inference for Duolingo Max and AI-generated course content. The same analysis puts Duolingo's monthly active user growth at about 6%, decelerated from 33%.
Engagement, not access, limits student-facing tutors. Khan Academy's Khanmigo has 700K+ users across 380+ U.S. districts, yet only 15% of Khanmigo users actively engage despite access. A two-year Tennessee randomised trial of Khanmigo across 18 schools and about 6,900 student-terms found 96% of students tried the tool at least once. Median use was one third of practice days and 17% of error-correction sessions. Gains matched Khan Academy without AI, which points to practice structure rather than AI capability as the driver of achievement.
Course-integrated tutoring can displace learning rather than add to it. A working paper on a randomised trial with 2,379 undergraduates and 30 instructors found that access to a GPT-4o-based study assistant reduced final grades by 0.37 standard deviations and learning management system participation by 0.90 standard deviations. Losses were greater among first-generation students. The authors conclude the tutor may displace engagement with existing learning activities. The paper has not been peer reviewed.
Syntheses tie gains to human mediation. Stanford's SCALE Initiative synthesis reports 40–47% non-use rates for AI-led tutoring. It finds emerging evidence that AI assisting human tutors is equally or more effective, which makes human involvement a condition of scaled deployment. Brookings research confirms gains only when educators remain present rather than leaving the AI autonomous.
Short, structured sessions can match human tutors on narrow tests. The StudentBench preprint randomised 2,383 participants and reports pooled AI and expert human tutoring gains as statistically equivalent. One AI tutor matched human gains at $0.0052 per percentage point versus $4.81 for humans. Sessions lasted one hour, and the paper has not been peer reviewed.
The research base rarely measures learning itself. A process-centred synthesis of 88 LLM-tutoring studies finds only 9 (10.2%) connect learner-state estimation, learner-conditioned intervention, an explicit policy and a downstream outcome. Only 10 (11.4%) measure human learning, and one measures transfer or delayed retention. The same synthesis flags that Khanmigo access did not translate into substantive tutoring interaction.
Reading tutors now run at state scale, with mixed reception. Axios reports that Iowa's Amira rollout has reached more than 96,000 students, with the state funding it for the first 110,000 elementary students through July 31, 2028. Students using Amira at least 20 minutes a week in 2025-26 gained half a school year's worth of progress compared with students who did not, Axios reports. A West Des Moines parent opted her daughter out after the tutor repeatedly misheard her, and Waukee declined to adopt it.
Independent evaluation lags these rollouts. TechTarget reports that San Francisco Unified School District licensed Amira for supplemental K-5 literacy instruction. It quotes Rebecca Winthrop of Brookings saying researchers have conducted relatively few external evaluations of generative AI interventions in school districts. Michael Horn of the Christensen Institute recommends outcomes-based contracts that tie vendor payment to learning milestones. Alpha School expanded from 15 to 50 campuses this year. The coverage reports no learning gains for any of the systems it describes.
Vendors have moved their products towards scaffolded learning. Google, OpenAI and Anthropic announced a Q3 2026 shift from quick-answer efficiency to Socratic questioning, step-by-step progression and guided reasoning. The Google and Khan Academy partnership released interactive Gemini diagrams for August 2026 back-to-school.
Restraint has to be designed in explicitly. The Allen Institute's TutorMoments framework shows LLMs default to over-helping rather than productive struggle. A Harvard/MIT study (N=1,280) of reinforcement-learning adaptive tutoring that calibrates help to learner state preserved skill retention (0.70 effect) while keeping accuracy gains. In an 18-student classroom study, TestCompass, a tutor that gives feedback on unit-test suites without writing tests, kept novices developing and verifying tests themselves at comparable quality. Neither group reported short-term gains in perceived testing knowledge or confidence.
Open-source practice generators are gaining developer traction. DeepTutor, an Apache-2.0 self-hosted tutor from the University of Hong Kong Data Intelligence Lab, generates practice questions and exams grounded in learners' own documents and adapted to their mistake history. It passed 40,000 GitHub stars on 20 September 2026, with over 5,000 forks and 100+ contributors by October 2026. Its release log adds Mastery Path modes, daily practice and question-bank practice, reaching v1.6.15 on 19 October 2026. The project reports no learner outcome data.
Structured practice with calibrated feedback is where gains appear. Duolingo guidance describes two to three scheduled AI-tutor sessions a week with repair drills and generalisation exercises producing skill gains, while unfocused exploration yields minimal transfer. A ten-week quasi-experiment with 88 first-year EFL students at an Ethiopian university found a ChatGPT draft, prompt, reflect and revise cycle outperformed teacher-marked instruction on writing performance, metalinguistic reflection and writing self-efficacy. Its authors note that learners with lower metacognitive awareness tend to copy ChatGPT output.
Evidence from performance domains remains associative. A Frontiers in Psychology study followed 512 conservatory musicians over 16 weeks and found AI-assisted practice engagement associated with later self-regulated learning and instructor-rated performance. The authors report these as longitudinal associations rather than causal effects, so the study cannot show that AI-assisted practice caused the improvement.
Simulated practice partners show measured gains in clinical communication. A JMIR Medical Education randomised trial assigned 51 students and clinicians either to SOPHIE, a generative-AI virtual patient giving structured feedback, or to reading. Blinded raters scored greater improvement in all three serious-illness communication skills for the SOPHIE group. The control was passive, and outcomes were measured only immediately after training.
Automated feedback quality remains unsettled. A SimConverse calibration study reports that no consolidated best practices exist for calibrating AI feedback. A survey of 29 primary studies on misconception diagnosis in programming education finds LLM systems show only fair to moderate chance-corrected agreement with experts and a false positive bias, at higher cost than rule-based methods. A systematic review of 12 studies on generative AI for EFL speaking and pronunciation practice identifies 48 challenges, with output accuracy and reliability the most persistent, and AI feedback repeatedly described as mechanical and formulaic.
Unguarded assistance still trades learning for visible performance. Population-level evidence from 26,811 students over 30 months shows homework gains of 18% reversing to exam losses of 20%. Medical training research documents a never-skilling risk when AI is introduced before foundational mastery. The authors of the university trial describe a learning penalty seen with general-purpose AI in other studies, which they suggest purpose-built tutoring safeguards may mitigate.
Gains from AI assistance can vanish when the tool does. A Nature Human Behaviour randomised trial gave 487 management consultants ChatGPT for data science tasks outside their skill sets. Treated workers outscored controls on three tasks but did not perform as well as expert data scientists. Without ChatGPT afterwards, they showed no evidence of being better at technical questions. The authors describe AI as an exoskeleton whose gains depend on continued use. Several authors are BCG employees and one works for OpenAI.
Deskilling can stay invisible in output metrics. A theoretical model by Diener and colleagues shows novices who accept a competent AI's answers staying trapped at low skill, while experts lose skill as output accuracy holds steady. It proposes learning incentives that make independent work self-sustaining. Its examples use illustrative values and do not establish how often the trap occurs.
Workplace skill formation faces its own pressures. In the workplace, 85% of enterprise learners cannot apply AI training to their work contexts. A Chief Learning Officer analysis argues agentic AI moves employees from performer to approver, removing the repeated work through which judgement develops. It cites Singapore's Model AI Governance Framework for Agentic AI, which warns that agents taking over entry-level tasks cause skill degradation. A Utah State University pilot with Cooperative Extension professionals found a custom GPT tutor and an online course each raised evaluation knowledge, but combining them produced the smallest self-efficacy gain.
Demand for skills is rising alongside teacher training. An American University survey documents employers requesting AI skills rising from 11.6% to 42.6% between 2024 and 2026, with 80%+ student adoption within six months. Pakistan's Sindh Department trained 3,503 teachers in Khanmigo with a 91% pass rate, and AI lesson-planning adoption there rose from 0% to 79%. Ease of assessment design rose from 51% to 92%.
Teacher augmentation is outpacing student tutoring. Teacher-facing tools such as MagicSchool (6M) and Brisk (1M) gained 2024-25 market share from student tutors as K-12 districts prioritised teacher augmentation. Instructure research found 68% of K-12 educators using AI at least occasionally while 45% had received no formal AI training. LATAM higher education reports 92% student and 79% faculty AI engagement.
Institutional design, not technology, blocks broader adoption. U.S. K-12 uptake remains constrained by integration friction and by the absence of rigorous independent evaluation in most rollouts. Infrastructure is a harder limit elsewhere, with 40% of primary schools globally lacking internet. Most organisations lack the pedagogical expertise to build skill-preservation mechanisms, and teacher training adds a further burden. Practice contexts often fail to match the contexts where skills are used. Outcomes depend on a named human being accountable for engagement, which platform access alone does not supply.
Tier History
Evidence (198)
— Theoretical model: accepting AI answers traps novices at low skill and lets experts deskill invisibly while output accuracy holds. It proposes learning incentives and has no human data.
— Negative signal: a PRISMA review of 12 studies finds 48 challenges in GenAI speaking practice. Accuracy is the most persistent, and the feedback is called mechanical and formulaic.
— Release log shows DeepTutor shipping structured-practice features: Mastery Path modes, daily practice, question-bank practice and grouped assessments (v1.6.15, 19 Oct 2026). No outcome data.
— Peer-reviewed RCT of 487 consultants: ChatGPT lifted scores on unfamiliar data science tasks, but there was no retained technical knowledge once it was removed. Several authors are from BCG and one from OpenAI.
— Feedback-only tutor (TestCompass) kept 18 novices doing hands-on test work with comparable quality, but neither group reported short-term gains in testing knowledge or confidence.
193 more · latest 2026-10-23 →
— Survey of 29 studies: LLM misconception diagnosis shows only fair-to-moderate chance-corrected expert agreement and a false-positive bias, at higher cost than rule-based feedback methods.
— Registered RCT (n=51): SOPHIE AI virtual patient with structured feedback beat a reading control on all three blinded-rated communication skills, though only measured immediately after training.
— Synthesis of 88 LLM-tutoring studies: only 9 close the learner-state-to-outcome loop, 10 measure human learning and one measures transfer or delayed retention, so the evidence base is thin.
— Randomised 2×2 workplace pilot: custom GPT tutor and online course each raised Extension professionals' knowledge and self-efficacy, but combining them gave the smallest self-efficacy gain.
— SFUSD licenses Amira and Alpha School grows from 15 to 50 campuses, while Brookings and Christensen Institute voices note few external evaluations and urge outcomes-based contracts.
— Negative signal: argues agentic AI removes the repeated real work that builds judgement, shifting staff from performer to approver, citing Singapore's agentic AI governance framework.
— Methods study calibrating AI rubric scoring and feedback in the SimConverse simulation platform against a blinded expert; notes no consolidated best practice exists for calibrating AI feedback.
— Third-party analysis of Duolingo filings: gross margin guided from ~73% to ~69% on LLM inference costs for Duolingo Max and AI content, with MAU growth slowing to ~6%; Q1 figures predate held Q2 data.
— Iowa's statewide Amira reading-tutor rollout reaches 96,000+ students with a reported half-year gain for regular users, alongside speech-recognition complaints, an opt-out and one district declining.
— Independent RCT (2,379 undergraduates, 30 instructors): GPT-4o course tutor cut final grades by 0.37 SD and LMS participation by 0.90 SD, with larger losses for first-generation students.
— Preprint randomising 2,383 participants: pooled AI and expert human GRE tutoring gains statistically equivalent, one AI tutor at $0.0052 vs $4.81 per percentage point; one-hour sessions, unreviewed.
— Brandon Hall Group analysis finds only 17% of organisations reach deliberate-practice maturity; 39% lack foundational infrastructure (personalised pathways, gap analytics, scaled access)—shows capability and structure, not technology, are the binding constraints at scale.
— Workera survey of 1,000 U.S. professionals quantifies adoption–practice inversion: 67.8% use AI tools regularly but 56.4% receive no work time for upskilling—evidence of 'adoption-without-policy crisis' that the practice's landscape already identifies as systemic.
— Scoping review of 153 GenAI medical education reports shows encouraging short-term signals for history-taking and clinical reasoning, but only 13 tracked retention and zero documented patient-level outcomes—confirming the practice's transfer and durability challenge.
— Narrative review of 20 studies identifies AI maths tutoring benefits only when valid learner modelling combines with adaptive tasks, calibrated scaffolding and teacher mediation—confirms increased platform interaction does not guarantee conceptual understanding or transfer.
— Mixed-methods study of 302 Turkish primary maths teachers shows moderate positive perceptions masked by serious concerns: data privacy, reliability and 'students' possible over-reliance on ready-made responses'—evidence of cognitive offloading mechanism identified in the practice.
— Multiple surveys (Talogy 207 HR leaders, PwC, Gartner 12,004 employees) document erosion of entry-level roles as deliberate-practice ground: 78% concerned about skill loss, junior roles most AI-exposed are seven times more likely to demand senior capabilities—validates practice's deskilling risk.
— PRISMA systematic review of 73 EFL studies (2015–Feb 2026) identifies AI chatbots supporting willingness to communicate via private practice spaces and personalised feedback—rare positive evidence on deliberate-practice efficacy in language learning.
— Wharton analysis citing Microsoft Research (319 knowledge workers), BetterUp (1,150 desk workers) and MIT Media Lab documents how heavy AI use erodes independent thinking—mechanism explaining why deliberate-practice support can actively undermine learning if poorly designed.
— Three-wave, 16-week study of 512 conservatory musicians links AI-assisted practice engagement to later self-regulated learning and instructor-rated performance; observational, not causal.
— Independent Instruction Partners evaluation of 20 AI learning tools (16 deep dives) found general-purpose chatbots reduce effortful thinking and leave learners 'isolating'; only purpose-built tutors showed promise—confirms tool design determines whether AI supports or undermines skill acquisition.
— Secondary analysis of RCT with 2,899 student sessions showing 64–67% of students with high AI-assisted practice scores later scored below 50% on unaided exams—critical evidence of assisted-performance illusion in AI-supported deliberate practice.
— NBER RCT with 6,000+ students across 18 Tennessee schools shows AI tutoring effectiveness depends entirely on design, guardrails, and student willingness; without compulsion, students largely avoid the tutor despite availability.
— Compilation of adoption metrics from primary sources (RAND, UNESCO, Gallup, HolonIQ) showing 86% global student AI use, 53% K-12 teacher use, +42% market CAGR, and <10% institutional guidance—signals mainstream adoption outpacing governance.
— Case study of El Salvador's national AI tutoring expansion with World Bank backing: 1,198 volunteer students at 171 schools scored above national average on PISA for Schools; expansion planned to 5,000+ schools and 1M+ students over 18 months.
— Washington University empirical study of competitive programming distinguishing AI substitute (harmful) vs. complement (beneficial) modes, with proctored evaluation gates separating skill preservation from scaffolding dependence in real-world elite skill contexts.
— Major vendor (Google) general availability of 30+ integrated AI tutoring tools across Classroom, Gemini, Search, and Chromebook—signals ecosystem maturity and enables broad institutional deployment with existing district infrastructure.
— Critical assessment of New Mexico's AI tutoring mandate documenting data-privacy concerns, scoring inaccuracy, and acknowledged lack of efficacy evidence—important negative signal on adoption barriers and implementation risks in governance contexts.
— Documents the 'deskilling problem': 39% of workers report AI weakened their skill set; junior engineers with AI access show measurably weaker code comprehension and debugging despite working faster—revealing how AI substitution harms skill development.
— Duolingo Q2 2026: DAU +23% to 58.7M, paid subscribers +17% to 12.7M, with CEO guidance for sustained >20% DAU growth—confirming platform adoption breadth and sustained engagement with AI-powered language practice mechanics at consumer scale.
— Google and Khan Academy launched interactive Gemini-powered diagrams for Khanmigo (August 2026): students manipulate visual elements, receive contextual feedback; early signals show increased engagement and improved answer correctness—representing product evolution in response to identified barriers.
— Pakistan's Sindh Department trained 3,503 teachers (91% pass rate) in Khanmigo; AI lesson-planning adoption rose 0% to 79%, assessment design ease 51% to 92%, with teachers saving 5+ hours/week—demonstrating professional skill acquisition deployment at scale in Global South context.
— Stanford SCALE Initiative authoritative synthesis: AI-led tutoring shows 40–47% non-use rates and insufficient evidence base; AI-assisted human tutors shows emerging promise; high-impact tutoring requires human relationships and structured engagement.
— SEC Form 8-K disclosed Duolingo's 27.4% YoY DAU growth on Aug 17, 2026, confirming sustained user engagement with AI-powered language practice despite industry disruption narratives; primary regulatory source at scale.
— Quasi-experimental study (13,037 students) comparing AI feedback workflows: 'Enacted Feedback' (structured student dialogue) achieved 26.2% uptake vs 14.1% directed and 0.1% self-directed—demonstrating workflow design as critical to feedback utility in deliberate practice.
— Two-year RCT across 18 Tennessee schools (N~6,900 student-terms) found Khanmigo gains (0.06–0.08 SD/year) matched Khan Academy alone; 96% trial rate but median engagement only 1/3 of days—identifying engagement as binding constraint to skill acquisition deployment.
— Large RCT (6,145 students, Hamilton County Schools) shows AI tutoring value depends on mastery-based workflow structure: AI improved post-error accuracy (+8.5pp) and recovery time, with 3.2pp delayed-test gains when embedded in workflow—design determines outcome.
— Peer-reviewed research documents never-skilling risk in medical training; proposes three-stage approach (establish competency without AI, teach critical evaluation, supervised integration) to preserve deliberate practice in clinical skill acquisition.
— Implementation guide for structuring deliberate practice routines with Duolingo Max: 2–3 Roleplay sessions/week, 5–15 minute sessions, turning wrong answers into repair drills via reconstruction and generalization. Identifies adoption barrier: gap between exploration and disciplined repetition determines skill gains.
— Google, OpenAI, and Anthropic pivoting learning products from quick answers to scaffolded Socratic approaches: asking questions, providing hints, breaking content into steps. Represents strategic shift by major AI vendors away from efficiency gains toward active learning and skill preservation.
— Physicians Bajaj and Sakran propose pedagogical interventions grounded in Bjork's desirable difficulties: require independent reasoning before AI consultation, periodic no-AI cases (parallel to FAA pilot currency), and sequential ordering (unaided diagnosis first) to preserve deliberate practice in clinical training.
— Allen Institute TutorMoments framework (replay-based evaluation on math tutoring transcripts): LLMs when instructed to 'tutor well' over-help (excessive scaffolding, rarely push for reasoning), diverging from pedagogically sound tutoring; identifies calibration gap as critical design challenge for AI-supported skill development.
— American University survey (N=483, tracked 2024–2026): employers requesting AI skills rose 11.6% to 42.6% (285% increase); student adoption 80%+ within 6 months; shows rapid shift from supplementary to baseline skill requirement, with growing demand for coding/technical instruction.
— Harvard/MIT reinforcement learning study (N=1,280) shows adaptive tutoring that calibrates help-level by learner state and confidence preserves skill retention while improving accuracy, validating mechanism to balance immediate performance with long-term learning.
— AWS Developer Advocate documents deliberate practice methodology using AI assistants: goal-based learning, spec-driven development, strategic collaboration. Applies Pearson & Gallagher's 'I do, we do, you do' (gradual release of responsibility) with zone-of-proximal-development scaffolding.
— Q1 2026 deployment metrics: 56.5M DAU (+21% YoY), 20,500 AI-generated course units per quarter (10× acceleration), new speaking-focused features (spoken tokens, speaking adventures); demonstrates sustained deliberate practice investment.
— AEFP policy synthesis of 20 high-quality causal studies: tools prompting reasoning and active engagement reinforce cognitive work for deeper learning; systems generating complete answers reduce cognitive work, directly undermining skill acquisition.
— Meta-analysis of 35 experimental studies: knowledge acquisition improved with AI assistance, but higher-order thinking and critical analysis showed smaller/reversed gains; demonstrates cognitive domain determines efficacy, not tool capability alone.
— 454-student semester-long RCT: accurate, grounded, popular RAG chatbot produced zero learning gains despite high student satisfaction, separating perceived helpfulness from efficacy in skill acquisition.
— RCT with 90 learners: personalized feedback yielded highest long-term mastery but temporarily reduced next-attempt performance due to cognitive load, demonstrating temporal trade-off requiring dynamic assessment of learner state.
— Deep behavioral analysis documenting Duolingo's deliberate practice engagement mechanics: 34.7% DAU/MAU ratio, 10M+ users with 1+ year streaks via loss-aversion streaks and variable rewards, driving 41% YoY revenue growth at scale.
— Rigorous analysis distinguishing implementation factors from platform effects, comparing Harvard RCT (custom-built, tightly scaffolded) vs general platforms; meta-analysis finding: effects larger on tutor-aligned assessments than transfer tasks.
— RCT with 275 programming students: both AI conditions improved task completion and reduced frustration, but neither produced knowledge gains, demonstrating the 'comfort trap' where easier work masks absent learning transfer.
— Technical playbook identifies five architectural pillars required for effective AI tutors (curriculum RAG, mastery model, pedagogy, conversation, engagement); cites ~70% day-30 collapse when pillars missing; Khanmigo 2025 RCT +0.34 SD.
— Metro Nashville DEC deployment: 19.50% faster time-to-competence, 10.95% higher mastery, 95.45% trainer alignment. Real named-org deployment scaling skill training in high-stakes domain.
— University of Pennsylvania adaptive sequencing study with 800 Taiwanese high school students: personalized difficulty achieved 6-9 month equivalent improvement over fixed-difficulty control; demonstrates zone-of-proximal-development value.
— Empirical study of 16,851 real tutoring interactions identifying concrete elaboration (analogies, examples) significantly aids understanding; response length inversely associated with learning. Informs deliberate practice design.
— VC market analysis: Khanmigo grew 17x to 700K users but only 15% use it regularly (269K weekday interactions). Key finding: 'AI tutoring in 2026 has an adoption problem, not a technology problem.'
— North Carolina cut Khan Academy contract 95% (from $10M to $500K) citing Sal Khan's admission Khanmigo 'has been a non-event for students'; reflects institutional skepticism despite infrastructure investment.
— UC San Diego ASPIRE system across 12 courses (2,200+ students): Foundations of Precalculus failure rate dropped 35.3% to 11.4% through embedded AI tutoring preserving productive struggle. Demonstrates scale with named institution.
— Course restructuring scrambled learner progress, leaving users confused; 46% say annoying but continuing, 9% quitting. Reveals platform maturity gap when redesigns erase user investment.
— 30-month longitudinal study of 26,811 Chinese secondary students: AI assistance raised homework scores 18% but reduced exam performance 20% (1.4 SD effect) within 6 months, establishing cognitive offloading as population-scale mechanism undermining durable skill development.
— RCT (N=90 students): Hybrid AI+teacher co-design outperformed AI-only and teacher-only conditions on motor-skill gains (d=1.42 vs 1.01 and 0.78), lesson quality, and efficiency (67% planning-time reduction), validating human-AI collaboration as effective deliberate practice model.
— Stanford SCALE Initiative RCTs across two districts: AI tutors alone produced near-zero engagement and learning gains; human support increased usage 71-80% but remained insufficient without baseline engagement, establishing implementation and human involvement as binding constraints.
— Synthesis of Pew, RAND, OECD, and University of Chicago research on AI learning outcomes: productivity gains consistently fade or reverse when tool removed (exam performance advantage vanishes), establishing tool-removal decay as systematic limitation of unguarded AI assistance.
— PNAS study (Bastani et al. 2025): Randomized trial comparing AI tutoring with and without pedagogical guardrails showed GPT Base group achieved +48% on assisted exercises but regressed 17% on unassisted exams, while guardrailed version preserved learning, establishing desirable difficulty as critical design parameter.
— UK government procurement document and Education Endowment Foundation research call explicitly identify cognitive offloading as measurable harm; DfE Safety Standards now require tracking 'cognitive offloading rate,' establishing pedagogical design (cognitive preservation) as procurement-level requirement.
— World Bank analysis of 26,000 Chinese students: autonomous AI use for homework improved practice 18% but reduced exam performance 20% (1.4 SD effect), entrance exam scores drop 18-24%, establishing metacognitive laziness at population scale.
— Peer-reviewed study of 1,498 undergraduates: gap between AI-assisted output quality (7.62/10) and independent knowledge mastery (5.55/10) documents illusion of competence, showing AI assistance improves perceived but not actual skill acquisition.
— Multi-site RCT (1,222 participants) by Liu, Christian, Bakker, Dubey: unguided AI assistance reduces persistence and undermines independent problem-solving; effect occurs within ~10 minutes, establishing behavioral mechanism critical to deliberate practice design.
— Multi-site RCT (1,222 participants) shows unguided AI assistance reduces persistence and impairs independent problem-solving within ~10 minutes, establishing behavioral dependency as intrinsic risk directly undermining durable skill acquisition in deliberate practice contexts.
— Brookings synthesis of 20+ years EdTech research: success requires three conditions (teacher involvement, infrastructure readiness, rigorous evaluation); cites RCTs from Nigeria and Ghana showing gains only when teachers remain present; establishes deployment prerequisites.
— Peer-reviewed framework directly addressing AI-supported deliberate practice: EFFORT-AI specifies six phases to preserve cognitive effort and retrieve, explain, monitor, and transfer as learner-owned processes rather than delegated to AI.
— Meta-analysis of 4 RCTs (268 surgical trainees): AI tutoring showed small OSATS skill gain (0.20 SD) but significantly higher cognitive load, with authors concluding findings do not support replacing human tutors and indicating hybrid model required.
— Khan Academy founder candid assessment: AI tutoring outcomes are 'mixed bag' / 'neutral-to-marginal positive'; disconfirms 2023 leading-edge claims; acknowledges cheating concerns and low adoption despite infrastructure scaling.
— Large independent survey (1,660 teachers, 10 countries): gap between adoption (50% positive) and meaningful impact (13% 'very positive'); teacher training critical (75% well-trained vs 38% untrained report positive impact).
— Large-scale RCT (770 students, 10 Spanish schools) shows realistic outcomes scaling tutoring: math grades +0.15σ, standardized tests +0.11σ—one-third of pilot effects. Critical voltage-drop evidence for real-world deployment.
— RCT of Tutor CoPilot (AI augmenting human tutors) shows largest gains for less-experienced tutors, leveling expertise gaps. Frames AI as instructor augmentation not replacement, demonstrating collaborative model for deliberate practice support.
— EdWeek opinion cites UPenn finding (48% grade boost then 17% crash post-AI removal) and Stanford report (gains vanish in independent assessment). Sets four conditions for durable learning: domain knowledge, productive struggle, critical thinking, human interaction.
— Gallup survey: fewer than 1 in 5 teachers (18%) received formal AI guidance; 48% informal; 33% none. RAND data: majority use AI but fewer than half of principals have AI policies. Signals institutional barriers to structured deployment.
— US Department of Education IES grant ($3.75M, 3-year 2024–2027) funding CAIT system development with deliberate practice design: adaptive assessment, personalized sequencing. Federal R&D commitment signals institutional confidence; pilot RCT 2026-27.
— Critical practitioner analysis: Khanmigo adoption remains ~15% despite 2-year pilot and free launch; text-based bot limitations; system architecture deficit ('anterograde amnesia'—no cross-session memory); cost barrier and competitive displacement persist.
— Independent analysis of 502K+ customer reviews: Duolingo's April 'AI-first' announcement triggered trust erosion (0.27%→3.71%) and quality concerns persisting 11 months. Customers interpreted AI pivot as quality-control removal, predicting later bookings deceleration.
— Enterprise learning survey (2,000+ learners): 85% cannot apply AI training to their roles; 78% trained in systems disconnected from work context; evidence that skill acquisition fails without environmental alignment.
— Large-scale longitudinal study (3.2M math problem interactions): ChatGPT release caused 26.9% collapse in study time on AI-susceptible problems and 25% decline in learning gains on proctored assessment, establishing behavioral cognitive surrender.
— OECD 2026 finding: practice performance gains (+127% with generic AI) paradoxically reduce exam scores (-17%), establishing 'metacognitive laziness' as core mechanism undermining durable skill acquisition.
— RCT in Sierra Leone (1,800 students) and Italy (9,000 students) demonstrated Gemini grounded in learning science improved math mastery +0.26 to +0.38 SD; positive signal that well-designed AI tutoring with pedagogical grounding supports skill development.
— RCT (770 students): adaptive problem sequencing (not better answers) improved exam performance +0.15 SD; proactive pedagogical design outperforms reactive AI assistance; strongest gains for beginners, narrowing skill gaps.
— Critical market assessment: Sal Khan admitted Khanmigo became 'a non-event' for most students; 95% engagement exclusion rate and engagement cliff (60% drop after 3 weeks) documented; teacher tools bifurcated market from student tutors.
— Turkey RCT (1,000 students): unguarded AI raised practice accuracy 48% but lowered exam performance 17%; safeguarded AI version preserved skills; OECD identifies cognitive offloading as mechanism and documents pedagogical guardrails required.
— Research from Carnegie Mellon, MIT, Oxford, UCLA shows 10 minutes of AI assistance impairs analytical thinking and problem-solving persistence; identifies cognitive dependency as foreseeable risk directly undermining durable skill development.
— AIED 2026 peer-reviewed analysis of 10,235 code submissions from deployed AI tutors reveals engagement-based behavioral signals more strongly predict student learning than pedagogical quality alone; identifies substantial differences in tutor effectiveness in driving student action.
— Practitioner documentation of classroom Khanmigo implementation: students skip guiding prompts, seek direct answers, avoid productive struggle; reveals behavioral adoption gaps and weak knowledge transfer despite system design for deliberate practice.
— 6-month A/B testing (Oct 2025–April 2026) on 1.35M+ tutoring threads shows: learning history +3.4%, prerequisite review +2.7%, conversation context +5.09% engagement; demonstrates iterative refinement driving measurable skill transfer gains.
— Critical deployment signal: Khan Academy's own Chief Learning Officer admits only 15% of students with Khanmigo access actually engage with it, triggering redesign from passive to proactive AI; reveals adoption gap despite 108M interactions.
— Official metrics: 56.5M DAU (+21% YoY), 12.5M subscribers (+21%); product pivot to Speaking Adventures, Video Call, and spoken tokens demonstrates sustained deliberate practice investment; 20,500 AI-generated course units show content velocity for skill progression.
— Harvard RCT (Kestin et al., Scientific Reports 2025) documents AI tutors outperforming active learning on engagement and motivation; emphasizes effectiveness contingent on pedagogical alignment and is primarily demonstrated in STEM-structured domains.
— Frontiers in Psychology meta-analysis of 72 studies shows AI-enabled teaching yields effect size g=0.586 but is fundamentally conditional on implementation type and application context; emphasizes importance of appropriate AI deployment over presence alone.
— Engineering firm documentation of production ITS deployments (Khan Academy, Carnegie Learning, Duolingo) at scale (600K–1M+ students). Documents required five-layer architecture (learner model, curriculum, pedagogy, interface, evaluation) and effect sizes d=0.66–0.79 only when guardrails are applied.
— Khanmigo deployment case study documenting Khan Academy's four-phase measurement evolution from intuition-based to production A/B testing, revealing iterative improvement in pedagogical design.
— Working paper establishing cognitive harm from AI tutoring as foreseeable documented risk, with neuroimaging evidence and institutional liability framework; distinguishes adult cognitive atrophy from childhood cognitive foreclosure.
— Khanmigo deployment data: 108M+ interactions since 2023, but only 15% student engagement despite access; introduces 'next-item correctness' metric measuring independent learning transfer, not AI-assisted performance.
— Quasi-experimental study (60 EFL learners, Indian university) showing AI-mediated writing gains sustained at 5-week follow-up without AI, indicating skill internalization not tool dependency.
— Large-scale RCT (14,892 students across 4 countries, 62 schools) found MathMentor-GPT produced measurable math achievement gains with teacher integration amplifying effects; named mechanism.
— OECD's authoritative 2026 policy framework establishes pedagogical design (augmentation, Socratic questioning) as determinant of learning gains, not tool access alone; defines effectiveness conditions.
— Meta-analysis contrasting Harvard RCT (pedagogically-designed AI tutor, 0.73-1.3 SD gains) with Wharton RCT (unguarded ChatGPT access, -17% exam decline); demonstrates design determines outcome.
— Synthesis of peer-reviewed studies documenting critical limitation: AI assistance creates cognitive dependency, reversing 48% practice gains to -17% exam failure when tool removed.
— Peer-reviewed umbrella review synthesizing 102 systematic reviews on AI in K-12: confirms diverse applications but identifies critical gaps in teacher training, student AI literacy, ethical frameworks, and theoretical grounding of effectiveness research.
— Systematic review of 28 studies (4,597 K-12 students): ITSs show medium-to-large positive effects vs traditional instruction, but mixed results vs non-intelligent tutoring; effectiveness contingent on pedagogical design and teacher guidance, not technology alone.
— Khan Academy founder admits Khanmigo deployment 'was a non-event' for most students; uptake and learning gains fell short of 2023 promises; teacher adoption at featured schools waned despite administrator enthusiasm.
— RCT (N=1,222) shows AI assistance improves short-term performance but reduces persistence and impairs unassisted task performance, directly undermining skill acquisition; effect emerges after only 10 minutes of AI interaction.
— Independent comparative testing: Khanmigo (Socratic method) achieved 23% learning outcome improvement vs ChatGPT's 16% with direct answers; pedagogical approach significantly affects skill acquisition effectiveness in math and science.
— Critical assessment of Khanmigo deployment: effective for guided problem-solving when students are in 'productive zone' but ineffective for foundational instruction; weak for open-ended writing; named Hobart HS teacher reduced usage citing student frustration.
— University of Sheffield analysis documents fundamental AI tutoring limitation: models deliver errors with same confidence as correct answers; 45% of AI responses contain significant inaccuracy but users cannot distinguish; pedagogical mismatch: immediate resolution of confusion prevents learning.
— Stanford preprint identifies systematic bias in deployed AI tutor feedback: high-achieving/White students receive development-focused critique; Hispanic/ELL students receive grammar-only feedback; low-achieving students receive withholding bias.
— Harvard RCT with 194 physics students: AI tutor median learning gain (2.75→4.5/6) more than doubled vs active learning (3.5); completion 18% faster; pedagogical guardrails (active problem-solving first, cognitive load management) were critical design features.
— African Educational Research Journal (2026): quasi-experimental study showed AI personalized tutoring with adaptive pathways and learning analytics significantly improved academic performance and reduced at-risk student proportion; enables early detection of learning difficulties.
— 5-month RCT in Taipei high schools (770 Python students): reinforcement learning-enhanced AI problem sequencing improved final exam performance 0.15 SD with larger gains for beginners; isolates adaptive sequencing as scalable AI enhancement.
— Khan Academy 2022-23 annual report documents organizational deployment scale: Khanmigo launched; added 13.7M new learners; 945K US K-12 students in Districts Program; 40% YoY growth in very-active learners (18+ hours/year); 5-year target to accelerate 5M students.
— PRISMA systematic review (Frontiers Education, 2026) meta-analyzing 2013-2025 literature: AI adaptive learning systems yield medium-to-large effects (g=0.50-0.70) moderated by study quality, implementation duration, learner characteristics; identifies professional development as essential.
— Georgia State University's Jill Watson AI teaching assistant answered 10K+ questions in single semester with 95%+ satisfaction; response time <3 seconds; demonstrates infrastructure scalability for one-on-one tutoring to thousands of students per educator.
— RCT with 334 university economics students: unrestricted GPT-4 access raised exam performance 0.23 SD vs textbook control, 0.21 SD more than restricted access; contradicts premature AI reliance concerns when students adopt scaffolding strategies.
— Peer-reviewed controlled experiments (Chua & Pan, 2026) with adult learners: pretesting on word-picture matching yielded recall improvements (d=0.18-0.40) and multiple-choice gains (d=0.25-0.67); extends deliberate practice methodology to digital learning platforms.
— Quasi-experimental study (Frontiers Psychology, 2026): 383 primary students in AI self-study rooms vs traditional study showed significant effects on motivation (η²=0.171), self-regulation (η²=0.238), engagement (η²=0.220) over 10 weeks.
— Interview with Sal Khan and district admin: India study showed 0.44 effect size over 7 months; implementation threshold ~18 hours/year student engagement and 60-skill mastery; explicit requirement for teacher oversight and intervention—not autonomous AI.
— IZA Discussion Paper: RCT of 334 students shows unrestricted AI tutor access raises exam performance 0.23 SD; unrestricted outperforms restricted access 0.21 SD; benefits concentrated among low-baseline-knowledge students with strong self-regulation.
— AI for Cause case study shows Khanmigo growth from 68K (Mar 2023) to 700K users across 380+ US school districts, using Socratic method pedagogy; Harvard/Stanford RCT confirms effect size favoring AI tutor over traditional instruction.
— Duolingo's AI-first strategy drove growth but stock correction reveals risks: revenue growth slowing to 18-20% for 2026, margin compression from R&D spending, and competitive threats from free LLMs; signals unsustainability concerns in AI-powered platform economics.
— Rose Luckin critique argues AI tutors address only ~16% of learning (knowledge delivery), neglecting metacognition and social intelligence; cites research showing AI-assisted learners exhibit metacognitive laziness; documents narrow scope of current AI tutoring approach.
— Nationally representative AmeriSpeak survey finds 57% personal AI use and 20% professional use, with adoption sharply increasing by education level; documents broad AI tool penetration including learning support among educated demographics.
— Industry synthesis of 2,400+ enterprise AI initiatives documents 80.3% overall failure rate and 95% GenAI pilot failure-to-scale; identifies leadership and organizational barriers; critical evidence of adoption challenges constraining AI skill tool deployment.
— Digital Education Council survey of 30,000+ LATAM higher education respondents shows 92% student and 79% faculty AI engagement, with 50% of students supporting AI-assisted feedback; demonstrates rapid regional adoption surpassing global trends.
— SIAI analysis examines AI tutoring scaling risks: cites 2025 Nature study showing AI tutor outperformed traditional active learning, but documents persistent inequality risks and error concerns; balances efficacy gains with critical adoption barriers.
— SIAI research finds AI course bots handle 60-90% of routine questions in first year; Tutor CoPilot study shows 4pp higher mastery achievement with AI-augmented tutors; validates scaled AI tutoring infrastructure for deliberate practice support.
— Anthropic study of developers learning new library shows heavy AI assistance decreased library-specific skill mastery by 17% (two grade points); critical evidence that AI delegation reduces conceptual learning despite modest productivity gains.
— Critical analysis cites PwC survey: 56% of CEOs report no effect from AI investments on costs/revenue; 87% of workers report <10% productivity savings; documents persistent gap between AI adoption and real-world benefit realization.
— Duolingo case study of 25,000 test-takers shows AI-powered practice test improves confidence and certified test performance through adaptive item generation; demonstrates deliberate practice efficacy at scale in language assessment.
— Khan Academy announced Khanmigo deployment across New Mexico districts with features for personalized student support and teacher insights; signals continued geographic expansion of AI tutoring infrastructure into additional U.S. regions.
— McNorton analysis of Khanmigo efficacy across 350,000 students shows 20% higher learning gains for students meeting recommended usage; validates AI tutoring effectiveness and establishes baseline engagement-outcome linkage.
— Critical assessment examining limitations of Duolingo's grammar-drill mechanics in preparing learners for real-world conversational deployment, highlighting gaps between structured practice outcomes and authentic communication readiness.
— Review of June 2025 Harvard physics study showing AI tutoring with verified solutions outperformed human instruction (effect sizes d=0.73-1.3) and achieved 70% efficiency gain, addressing hallucination concerns in pedagogically-designed tutors.
— Systematic literature review of 48 peer-reviewed studies comparing AI tutors against human tutors and traditional instruction, establishing comparative effectiveness and global adoption patterns across diverse educational contexts.
— Empirical classroom trial validating safe and effective AI tutoring implementation with rigorous safety audits, providing evidence-based support for pedagogically-sound AI tutoring deployment in educational settings.
— Editorial review establishing that properly-designed AI tutoring surpasses traditional pedagogical best practices in controlled settings, with RCT evidence published in Scientific Reports providing rigorous validation for deployment efficacy.
— Multi-university qualitative adoption study across 4 Australian institutions with 79 focus groups and 8,000+ student survey responses analyzing student experience and institutional integration of AI tutoring tools.
— Duolingo achieved 51% year-over-year DAU growth to 47M users and $1B revenue forecast through AI-powered conversational practice, lesson generation, and video tutoring; demonstrates production-scale deployment of deliberate practice mechanics.
— Survey of 2,000+ US K-12 teachers finds 60% use AI weekly reclaiming 6 hours on grading, but 42% of untrained users report anxiety over errors; highlights adoption acceleration paired with critical training and confidence gaps.
— LinkedIn survey: 51% of professionals report AI training feels excessive; 33% embarrassed about understanding, 35% nervous discussing AI; linked to broader enterprise AI ROI challenges (95% pilot failure rate).
— Khan Academy's Khanmigo user base scaled from 40K to 700K in one year with 1M+ expected in 2025-26; critical assessment notes lack of RCT efficacy data and MIT study documenting cognitive risks from AI tool reliance.
— University of Iowa pilot of Khanmigo Teacher Tools found faculty used tool <1/week with no reported teaching impact due to manual copy-pasting friction; provides negative deployment signal on institutional integration barriers.
— Educator Elliott Hedman documents that Khanmigo fails underrepresented students lacking self-efficacy; argues AI tutors only work for already-motivated students and raise equity concerns about skill development access.
— Critical analysis questioning whether Duolingo's gamification and habit-formation mechanics drive actual language fluency or primarily psychological dependence on streaks and rewards.
— Preregistered experimental study demonstrating that personalised AI dialogue effectively corrects persistent misconceptions in psychology and education, validating dialogue-based deliberate practice for conceptual learning.
— Critical assessment of Duolingo's AI-accelerated course generation (150 new courses in one year) and job displacement of human language teachers; raises concerns about quality trade-offs and labor impact.
— University of Nebraska Medical Center integrated Khanmigo Teacher Tools into Canvas LMS with faculty training; demonstrates institutional adoption of AI tutoring infrastructure across higher education.
— Duolingo engineering deep-dive on Lily AI video call feature design for conversational practice: structured guardrails, predictable patterns, and adaptive difficulty to enable realistic speaking partner interaction.
— Research review showing 1,000-student Penn study where guarded AI tutor improved mastery 127% vs unrestricted ChatGPT's 17% loss on retention; validates deliberate practice guardrails including hints-not-answers methodology.
— Production engineering improvements to Khanmigo's math tutoring—calculator integration, GPT-4 Omni upgrade, improved error detection—demonstrate active refinement of AI deliberate practice infrastructure addressing numerical reasoning.
— Coverage of Class Companion and Tutor CoPilot adoption in K-12 classrooms, with Tutor CoPilot showing 4pp higher mastery in math for 1,800 underserved students; signals real-world deliberate practice deployment at scale.
— Product manager interview reveals Duolingo's 'Lily' AI video call feature design for conversational speaking practice with personality and adaptive difficulty; demonstrates production deployment of realistic speaking partner mechanics.
— Critical analysis by Dan Meyer documenting specific Khanmigo interaction failures (2-second premature intervention) and argues AI tutors fail to approximate human teaching; provides negative signal on current AI tutor effectiveness.
— Named school deployment of Khanmigo in Geometry and broader math classes with documented improvements in engagement and achievement, including benefits for ELL and special education students via data-driven grouping.
— Critical analysis by microschool educator argues AI tutors fail to address core learning motivation and meaningful context; highlights limitation of curriculum-aligned problem-solving in fostering durable skill development.
— Philippines Department of Education partnered with Khan Academy and Smart Communications for nationwide free AI-powered learning access; 5-10GB weekly data allocation demonstrates government-scale commitment to skill acquisition infrastructure.
— CBS 60 Minutes coverage documents Khanmigo deployment across 266 school districts including Hobart High School; pilot students and teachers report effectiveness; $15/student annual cost demonstrates scaling economics.
— Large-scale efficacy study of ~350K students in grades 3-8 found 30+ minutes weekly use associated with ~20% greater learning gains (effect size .36), with longitudinal analysis linking skill practice progression to incremental gains.
— Stanford RCT with ~1,000 students and 900 tutors found AI-augmented tutoring improved student mastery by 4-9pp, with stronger gains for novice tutors; validates human-AI collaboration in skill development.
— Appen/Harris Poll survey documents declining enterprise AI deployment (47.4% vs 55.5% in 2021) and ROI realization (47.3% vs 56.7%), highlighting implementation challenges that constrain AI skill tool adoption.
— Pilot RCT (n=37) tested 8-week structured deliberate practice course with peer feedback for therapist skill development; demonstrates efficacy of systematized deliberate practice methodology with embedded feedback in professional training context.
— Quasi-experimental study with 40 EFL students compared Duolingo-integrated and control groups, demonstrating AI-powered practice improves language learner engagement and willingness to communicate in classroom settings.
— Khan Academy expanded free Khanmigo access to Arizona, New Hampshire, Ohio, Louisiana, and Oklahoma in partnership with state education departments, demonstrating geographic scaling of AI deliberate practice across U.S. regions.
— Clever survey of 100K+ schools found rising educator optimism about AI in classrooms while identifying critical gaps in inclusive edtech; reflects Q3 2024 shift in teacher sentiment toward deliberate practice tools.
— KQED investigation documented AI hallucinations in mathematics tutoring, including specific Khanmigo failures on algebra problems; provides critical assessment of reliability limitations in AI-based deliberate practice systems.
— Microsoft and Khan Academy released Khanmigo for Teachers free across 49 countries, powered by Azure OpenAI; signals continued institutional backing and intent to scale AI-powered deliberate practice infrastructure globally.
— Production case study detailing Duolingo's four-step course creation combining human curriculum experts with AI content generation, exercise synthesis, and evaluation; demonstrates scaled human-AI collaboration in deployed skill practice platform.
— Research identified algorithm aversion as a critical adoption barrier in AI tutoring despite proven reliability; highlights trust as a bottleneck for broader deployment of deliberate practice systems.
— Peer-reviewed controlled experiment demonstrated that AI-generated evaluative feedback enhances skill acquisition and transfer in sequential decision-making tasks, with empirical validation via NSF-funded research.
— Microsoft partnership scaled Khanmigo for Teachers to all U.S. educators via free pilot access, covering $44 annual cost; signals ecosystem expansion and institutional backing for AI deliberate practice infrastructure.
— RAND survey of 1,020 teachers found 33% overall AI adoption, 18% regular use for skill-building tasks (content adaptation, material generation); adoption higher among higher-poverty schools and middle/high school instructors.
— Temple University faculty integration of AI for formative feedback and deliberate practice showed students shifted from viewing AI as cheating to collaborative learning tool; demonstrated via structured assignments with immediate AI feedback.
— Student survey and interviews found Khan Academy's AI tutor Khanmigo significantly improved comprehension and motivation through immediate feedback and autonomous learning support.
— Controlled study of 77 dental students showed ChatGPT learners achieved higher exam grades (P=.045) than literature-research group, demonstrating effective structured practice with immediate feedback.
— Duolingo deployed GPT-4 for lesson personalization and AI-powered conversation practice, demonstrating production deployment of deliberate practice mechanics at scale.
— IEEE study of ChatGPT, Gemini, and Copilot in mechanical engineering found AI struggles with numerical problem-solving but succeeds in theory-based skill instruction; 172-student survey.
— User evaluation of Duolingo Max's new AI-powered roleplay feature at scale; demonstrates deployed AI practice mechanics including scenario-based conversation with challenge escalation.
— Emerging product addressing core friction in language skill acquisition: finding conversation partners; demonstrates turn-based voice and text chat tutor design pattern for deliberate practice.
— Critical assessment of Duolingo's 21M+ daily active user base questioning whether gamified practice drives durable language proficiency or mainly engagement; identifies core tension between addictiveness and learning efficacy.
— CIPD survey of 1,108 L&D professionals found only 5% currently use AI for learning, with 6% planning adoption; documents very low corporate training adoption despite widespread interest.
— Randomized controlled trial of 1,800 K-12 students from underserved communities showed AI-guided tutoring improved mastery by 4pp (p<0.01), with 9pp gains for students of lower-rated tutors; validates AI-augmented deliberate practice at scale.
— Early pilot in Newark Public Schools documented critical accuracy issues: Khanmigo gave direct answers instead of guidance and fabricated misinformation; provides essential negative signal on deployment reliability risks.
— CEO announced Duolingo Max live product with GPT-4, early rollout stage across 20.3M daily active users; signals real-world deployment of AI-powered skill practice at consumer scale.
— Official product general availability of Duolingo Max subscription with Video Call and Roleplay AI features for conversational practice, powered by GPT-4 across 188 countries; signals consumer-scale deployment of deliberate practice mechanics.
— Randomized controlled trial with 165 students showed AI peer dialogue improved physics post-test scores by 10.5pp despite 40% intentional inaccuracy; demonstrates effectiveness of dialogue-based deliberate practice methodology.