Perly Consulting │ Beck Eco

The State of Play

A living index of AI adoption across industries — where established practice meets the bleeding edge
UPDATED DAILY

The AI landscape doesn't move in one direction — it lurches. Some techniques leap from experiment to table stakes in a single quarter; others stall against regulatory walls, technical ceilings, or organisational inertia that no amount of hype can dislodge. Knowing which is which is the hard part. The State of Play cuts through the noise with a rigorously maintained index of AI techniques across every major business domain — classified by maturity, evidenced by real-world adoption, and updated daily so you always know where you stand relative to the field. Stop guessing. Start knowing.

The Daily Dispatch

A daily newsletter distilling the past two weeks of movement in a domain or two — delivered to your inbox while the index updates in the background.

AI Maturity by Domain

Each dot marks the weighted maturity of practices within a domain — hover for a brief summary, click for more detail

DOMAIN
BLEEDING EDGEESTABLISHED

Educational content adaptation & summarisation

BLEEDING EDGE

TRAJECTORY

Stalled

AI that localises, adapts, and summarises educational content for different contexts, languages, and learning levels. Includes textbook summarisation and cultural adaptation; distinct from curriculum design which creates new content rather than transforming existing material.

OVERVIEW

Educational content adaptation and summarisation represents a technical research frontier focused on transforming existing educational material—textbooks, lecture transcripts, articles—into forms suited to different contexts, languages, and learner ability levels. The practice is distinct from curriculum design, which creates entirely new learning pathways; instead it concentrates on the automated modification of existing content to extend its reach and utility.

The field sits at the intersection of natural language processing (particularly abstractive summarization) and instructional design. The core technical challenge centres on factual consistency and quality trade-offs: models that summarise complex educational material frequently introduce factual errors, hallucinations, or lose nuance critical to learning, and compression-based approaches reveal persistent tensions between conciseness and accuracy. This remains the primary barrier to reliable deployment in educational settings. By January 2026, the practice exhibits a fundamental paradox: consumer-scale adoption of summarization tools is mainstream and growing, yet institutional deployment remains constrained by unresolved accuracy, pedagogical outcome, and liability concerns.

CURRENT LANDSCAPE

As of March 2026, the practice exhibits marked polarization between accelerating learner adoption and institutional caution driven by mounting evidence of accuracy and learning outcome risks. Learner adoption has reached saturation in developed higher education markets: UK undergraduates have climbed to 92% (March 2026 HEPI survey, up from 66% in 2024), with summarization of articles and textbooks ranking as the second most-used AI application after concept explanation. Consumer summarization tools continue scaling: Mindgrasp operates at 100k+ users globally with stable revenue; SciSummary serves 700k+ academic users (including Harvard, Stanford, MIT) processing 1.5M+ papers. The AI transcription/summarization market is projected to reach $19.2B by 2034, with 62% of professionals reporting 4+ hours weekly time savings.

Yet evidence quality has become the decisive barrier to institutional deployment. March 2026 academic research from UC San Diego quantifies fundamental content adaptation failures: LLM-generated summaries exhibit a 26.42% "nuance shift" rate (content altered in direction or meaning) and a 60% hallucination rate. Independent testing by ToolHunt (March 2026) on BBC news articles found 51% of AI-generated summaries had significant problems and 19% contained outright factual errors. A Brookings Institution global study (March 2026, drawing on 500+ stakeholders across 50 countries and 400+ studies) concludes that "at this point in its trajectory, the risks of utilizing generative AI in children's education overshadow its benefits," citing impacts on foundational learning capacity, social-emotional well-being, and trust relationships. The critical paradox: the OECD Digital Education Outlook (March 2026) documents that unrestricted AI tools improve immediate task performance (students using LLMs wrote better essays, math exercises scored higher) but simultaneously undermine learning transfer—80% of students using LLMs to write essays could not recall their content afterward, and Turkish mathematics students using ChatGPT performed worse on concept exams than peers despite higher exercise scores.

Duolingo's Vision 2026 roadmap commits to building unique adaptive curricula per user, demonstrating continued large-scale deployment of content adaptation technology. Yet the December 2025 Duolingo failure—a 68% stock decline attributed to user complaints of "robotic lessons" and engagement collapse following aggressive AI-first content generation—signals the fragility of automation-heavy strategies. Institutional concerns have intensified: policies remain nascent (25% of campuses have formal AI policies), and deployment blockers cited include systemic bias, accuracy inadequacy (legal experts warn 80% accuracy is insufficient for liability-sensitive contexts), privacy risk (FERPA violations), and unresolved pedagogical outcome gaps. The Brookings framework distinguishes between "AI-enriched learning" (pedagogically sound design with human oversight) and "AI-diminished learning" (overreliance that undermines capacity), underscoring that tool design and institutional safeguards, not just capability, determine whether content adaptation supports learning or substitutes for it.

The practice remains learner-driven rather than institutionally deployed. Production-ready tooling exists, consumer demand is mainstream, and time-savings are documented. Institutional deployment remains blocked by unresolved accuracy deficits (hallucination rates 19-26% in current systems), learning outcome risks (performance-retention decoupling), bias in training data, and liability concerns. April 2026 evidence reveals no resolution of core barriers: EACL 2026 research identifies Harmful Factuality Hallucination (HFH) failure mode where LLMs misplaced correctness in rephrasing (mitigable ~50% via prompting), peer-reviewed studies document representational and linguistic bias endemic in personalized content generation (>75% of educators acknowledge non-neutral outputs), and K-12 real-world deployment shows cultural erasure risks (AI simplifying Spanish text in student writing). Duolingo's April 2026 content adaptation features (Explain My Answer, Video Call, Roleplay) with 10-fold generation capacity and 148 new courses demonstrate continued large-scale tool deployment but without resolution of the pedagogical barriers that defined the bleeding-edge stall. By April 2026, the bleeding-edge phase exhibits stalled institutional momentum: consumer adoption has plateaued at saturation in leading markets, specialized deployment tools (Diffit, Curipod, NotebookLM) reach 31-40% teacher and learner usage but with moderate satisfaction (52% rate outputs as good/excellent), institutional inflection point remains visible (Cal State 460k+ deployment, $5.88B→$32.27B market expansion 2024-2030) yet constrained by unresolved accuracy, bias, and pedagogical outcome gaps.

By May 2026, institutional infrastructure maturity signals emerging alongside persistent barriers. Moodle LMS integration with Gemini for text summarization reached general availability, signaling mainstream LMS adoption of content adaptation features. Teachers actively use AI tools to translate educational materials for multilingual learners, expanding the practice into localization workflows. US school districts formalized AI acceptable-use policies, moving from early experimentation toward institutional governance. However, core barriers remain unresolved: citation fabrication rates persist at 55% for GPT-3.5 and 18% for GPT-4; systems lack persistent learner models for effective adaptation and require hybrid architectures combining knowledge graphs and retrieval-augmented generation to achieve reliable deployment. Critically, learning outcome risks intensify—passive AI summarization undermines memory formation compared to active retrieval practice, and students delegating written work to AI perform 18-25 percentile points lower on in-person assessments. Pre-service science teachers exhibit low trust in AI-generated explanations despite institutional pressure to deploy, positioning truth assessment as a pedagogically responsible practice requiring explicit verification. Research-backed frameworks (TASU in literature education, Revise–Locate–Justify routine) emerge for integrating content adaptation with human oversight, yet adoption remains limited. The meta-analysis evidence (g_p=0.586 across 72 studies) confirms positive teaching effectiveness when AI organizes and adapts materials, provided implementation includes pedagogical design and human verification infrastructure. By May 2026, the practice exhibits a stable contradiction: LMS integration and instructor adoption climb, consumer tools scale to 57% college usage weekly, specialized K-12 tools reach 31-40% adoption, yet institutional expansion remains constrained by unresolved accuracy, hallucination persistence, learning outcome decoupling, and the pedagogical complexity of deploying content adaptation safely.

TIER HISTORY

ResearchNov-2022 → Jul-2023
Bleeding EdgeJul-2023 → present

EVIDENCE (101)

— Gemini integrated into Moodle LMS for text summarization (product-GA); teachers using AI to translate materials for multilingual learners; US districts formalizing AI policies. Signals movement from experimentation to institutional governance.

— Systematic review of 8,000+ academic records: LLMs lack persistent learner models for content adaptation; hallucinations risk reinforcing misconceptions; hybrid architectures (knowledge graphs, RAG) required for reliable educational deployment.

— YouTube practitioner guide demonstrating AI tools for content adaptation: rewriting paragraphs for different reading levels using ChatGPT, Diffit, Brisk, EduCafe. Shows teacher adoption of AI-assisted content differentiation.

— Practitioner analysis: passive AI summarization for reading replaces active retrieval practice (proven low-effectiveness study technique), undermining memory formation. Highlights learning outcome risks of unguided summarization tool use.

— Frontiers in Psychology: Pre-service science teachers exhibit low trust in GenAI-generated explanations, positioning truth assessment as pedagogically responsible practice requiring explicit verification—signals practitioner adoption barriers.

— Frontiers in Education framework (TASU): seven pedagogical functions including content adaptation/curation role; Revise–Locate–Justify routine required to evidence-ground AI suggestions, protecting factual integrity and student voice.

— Meta-analysis of 72 studies: AI-enabled teaching shows positive effect (g_p=0.586). AI organizing and adapting instructional materials reduces teacher workload and enhances alignment between resources and learning goals.

— Industry analysis: 78% HS, 64% college students use AI; students delegating writing to AI perform 18-25 percentile points lower on in-person assessments than peers. Documents adoption scale alongside learning outcome decoupling.

HISTORY

  • 2022-H2: Foundation research in summarization quality established across major NLP venues (EMNLP 2022 focus on factual consistency, robustness, and evaluation metrics). Educators began early exploration of AI text generation in courses, with cautious and critical perspectives emerging alongside optimism about potential classroom integration.
  • 2023-H1: ChatGPT and GPT-4 driven practical adoption of summarization in mainstream tools (Microsoft tutorials, educational platform experimentation). Concurrent critical assessment identified persistent barriers: cultural bias in LLMs, hallucination risks, citation fabrication. Research documented poor cross-cultural adaptation despite deployment growth in complementary capabilities (tutoring). Content adaptation remains experimental rather than deployed.
  • 2023-H2: Academic research advances personalization metrics in summarization (EMNLP 2023), and commercial summarization tools reach GA (Mindgrasp). Deployment pilots show mixed results: bioRxiv's LLM-generated preprint summaries contain factual errors in specialized domains. Empirical evidence emerges of comprehension harms: Duke study shows complete AI reliance for writing reduces accuracy 25.1%. Institutional adoption remains slow (60% of colleges have no systemic action). Dedicated summarization startup Summari fails due to competition from platform-embedded features. Adoption blocked by technical accuracy, pedagogical uncertainty, and competitive viability.
  • 2024-Q1: Analyst organizations (CoSN, UOC eLinC) recognize content adaptation as emerging 2024 capability; Duolingo scales AI content generation with production deployment and contractor layoffs. Faculty adoption of summarization tools reaches only 5.92% in higher education. Research shows compression rates up to 98% achievable but quality varies across tools and domains. Real-world deployment reveals quality friction: Duolingo users report automated errors (mispronunciations, incorrect translations). Risk warnings increase around algorithm bias and privacy. Adoption remains constrained by technical quality gaps, institutional skill gaps, and absence of killer apps.
  • 2024-Q2: Large-scale deployment momentum accelerates: Duolingo uses proprietary optimization metrics (TSLW) to dynamically adapt content difficulty and pacing in production, demonstrating technical capability at scale. Learner adoption grows rapidly—UK survey shows 77.1% of teenagers now use generative AI with 44.4% for literacy applications. However, educator and institutional adoption lags: instructor use for content design reaches 36% but specialized adoption for adaptation/summarization remains constrained. Quality friction persists in production systems (pronunciation errors, translation inaccuracy), and pedagogical outcomes remain uncertain. Technical barriers (factual consistency, hallucination in specialized domains) and institutional barriers (policy uncertainty, skill gaps) continue to block broader educator deployment.
  • 2024-Q3: International content adaptation pilots show positive targeted results (Oman university deployment of culturally-aware adaptive content achieves 19% performance gains), while high-profile summarization failures surface in Q3: Australian government trial shows AI summarization scores 47% vs human 81% on accuracy, and Google's AI Overviews generates demonstrably harmful inaccurate summaries. Deployment momentum persists but accuracy evidence reveals critical limitations. Learner adoption continues scaling; educator/institutional adoption remains below 6% in most regions. Technical barriers—particularly factual consistency and domain adaptation—remain production blockers.
  • 2024-Q4: Duolingo demonstrates continued production-scale deployment (54% DAU growth, 37.2M users in Q3) attributed to AI-powered content adaptation and personalization. However, evidence converges on quality and outcome limitations: KPMG survey shows 59% student adoption but two-thirds report reduced learning/retention; peer-reviewed research documents systematic bias in LLM summarization across 13 major models; Ellucian survey shows rising institutional concerns (bias 36%→49%, privacy 50%→59% among 445 higher ed professionals). Regulatory barriers emerge: FERPA violations risk from classroom AI integration. Learner adoption continues, but institutional confidence declines as evidence base reveals pedagogical risks alongside technical limitations.
  • 2025-Q1: Student adoption reaches 92% in UK (up from 66% in 2024) and 86% globally for AI in coursework, with summarisation a primary use case (88% use AI for assessments). Institutional leadership perceives adoption as inevitable: 89% of US higher ed leaders estimate at least half students use GenAI. However, quality evidence deteriorates sharply: BBC study documents 70% error rate in AI-generated summaries across major platforms, extending prior evidence of accuracy failures. Pedagogical signals remain negative (66% report reduced retention) and institutional risk aversion increases (bias/privacy concerns rising). Content adaptation remains learner-driven rather than institutionally deployed; educators maintain caution despite overwhelming student adoption.
  • 2025-Q2: Commercial tooling reaches maturity: Azure Summarization service (GA, April 2025) and SciSummary (specialised higher ed tool) signal production readiness. Cengage survey shows 67% of students summarize concepts with AI, 45% of instructors use AI for content creation. However, deployment barriers intensify: Wikipedia halts AI summary pilot (June 2025) due to hallucination/credibility concerns, signaling platform-level rejection despite user demand. K-12 leaders cite systemic model bias and ethical concerns as adoption blockers (CoSN 2025). TEDx speaker emphasizes accuracy paradox: 100% student adoption in sample, zero faculty adoption due to "AI doesn't know truth" concerns. Content adaptation tooling is production-ready but institutional deployment remains stalled by unresolved quality and pedagogical outcome limitations.
  • 2025-Q3: Learner adoption continues scaling (UK: 88% use AI for assessments, primarily summarisation); Duolingo demonstrates continued production scale (40% DAU growth, 10.9M paid subscribers) with "100% automatic" content generation. However, adoption-reality gap widens: 48% of US districts train teachers but only 25% actually use tools. Accuracy failures persist in specialized domains (ChatGPT tested on 500 science papers shows frequent hallucinations and causality inversions). Educator adoption remains cautious; institutional barriers (accuracy, bias, FERPA risk, pedagogical outcomes) unresolved. Practice remains learner-driven rather than institutionally deployed.
  • 2025-Q4: Consumer-scale adoption of specialized summarization tools (Mindgrasp 9k monthly downloads; SciSummary 700k+ users). However, Duolingo's AI-first content generation strategy collapses: 68% stock crash, staff layoffs, user complaints of robotic lessons and engagement decline—critical negative signal of automated content generation risks. Teacher adoption for summarization reaches 38-44% for general AI but specialized tools remain niche. Institutional policies nascent (25% of campuses have AI policies). Accuracy barriers (hallucination in specialized domains) and pedagogical concerns (reduced retention) remain unresolved. Learner-driven adoption contrasts sharply with institutional hesitation.
  • 2026-Jan: Continued consumer adoption momentum: Mindgrasp reaches 100k+ global users; AI summarization market projects to $19.2B by 2034 with 62% professional time savings. Large-scale higher education surveys (LATAM: 30k+ responses) confirm mainstream GenAI integration. However, quality and institutional trust barriers intensify: legal experts highlight liability risks from 80%-accurate summaries; academic analysis warns of hallucinations and surface-level understanding limitations. Institutional deployment remains cautious despite strong consumer-market signals and specialized tool maturity.
  • 2026-Feb: Academic research from TU/e Library documents severe accuracy failures in major LLMs: Gemini 3 Pro 68.8%, ChatGPT 5 61.8%, Claude 4.5 Opus 51.3% under stress testing; overgeneralizations 5x more common than human summaries. Microsoft Azure Summarization service (GA) confirms production limitations: bias in training data, quality degradation for dialects, performance loss on conversational content. Consumer tool adoption continues (QuillBot, Scholarcy, Blinkist) with market maturity, but accuracy evidence strengthens institutional caution. Practice remains learner-driven with significant quality barriers blocking institutional deployment.
  • 2026-Mar: Learner adoption reaches saturation in leading markets with HEPI survey showing 92% UK undergraduate AI use (up from 66% in 2024), with article summarisation ranked second most-used application. Critical accuracy evidence intensifies: UC San Diego research quantifies a 26.42% nuance-shift rate and 60% hallucination rate in LLM-generated summaries; independent testing on BBC news found 51% of AI summaries problematic and 19% factually wrong; a Brookings global study (500+ stakeholders, 50 countries, 400+ studies) concludes AI education risks currently overshadow benefits for children. OECD Digital Education Outlook 2026 documents the performance-learning paradox: AI improves immediate task scores but 80% of students using LLMs could not recall essay content afterward. AI dubbing for eLearning localization cited as a positive adaptation use case with engagement benefits. Consumer adoption continues to plateau at saturation while institutional confidence declines further.
  • 2026-Apr: Specialized deployment tools reach measurable adoption: 31% of K-12 teachers use Diffit/Curipod weekly for reading-level content differentiation (52% rate as good/excellent); UK adoption survey shows 95% of students use AI generally with 94% for assessed work, using tools like NotebookLM for lecture summarization. Pew survey documents 40% of US teens use AI to summarize articles/books/videos. Duolingo scales adaptive features (Explain My Answer, Video Call, Roleplay) with 10-fold content generation capacity, launching 148 courses in Q1 2026. Cal State partnership scales personalized learning to 460k+ students. However, critical quality research continues: a 2026 hallucination benchmark documents domain-specific degradation — AI summarization error rates climb from 0.7% on basic content to 15.6% on medical and 18.7% on legal material, directly constraining deployment in specialist educational contexts; an empirical study of high school students finds AI summaries reduce cognitive load but significantly worsen long-term retention versus full-text reading, adding to the performance-retention paradox evidence; EACL 2026 identifies Harmful Factuality Hallucination where LLMs introduce misplaced correctness in rephrasing (mitigable ~50% via prompting); peer-reviewed study documents endemic representational and linguistic bias in personalized content (>75% of educators acknowledge non-neutral outputs); real K-12 deployment shows cultural bias risks (AI modifying student writing, suggesting language simplification, removing cultural authenticity). Academic research on XR platforms demonstrates proof-of-concept for integrating summarization with translation and sign language rendering (arxiv 2026-04-07). Institutional inflection point visible (market $5.88B→$32.27B by 2030) yet deployment barriers persist: unresolved accuracy deficits, pedagogical outcome gaps, and mounting evidence of representational bias in content adaptation systems.
  • 2026-May: Gemini integration with Moodle LMS for text summarization reached general availability, marking the first mainstream LMS-embedded content adaptation feature at institutional scale. A systematic review of 8,000+ academic records confirms that current LLMs lack persistent learner models for reliable content adaptation and that hybrid architectures combining knowledge graphs with RAG are required — the core technical gap remains unresolved despite rising LMS-layer adoption.