AI for All: Adaptive, Accessible, and Inclusive Learning Experiences in the Age of Intelligent LMSs

May combine two claims.

This mapping translated raw interaction data into cognitively meaningful indicators-for instance, passive forum reading was associated with reflective learning, while repeated quiz attempts were linked to active engagement.

“This mapping translated raw interaction data into cognitively meaningful indicators-for instance, passive forum reading was associated with reflective learning, while repeated quiz attempts were linked to active engagement.”

• cites [2] Angeioplastis, A., Aliprantis, J., Konstantakis, M., Varsamis, D., and Tsimpiris, A. (2025). The Learning Style Decoder: FSLSM-Guided Behavior Mapping Meets Deep Neural Prediction in LMS Settings. Computers, 14. sourcePartly supported

  Section 4.2 says the mapping turns Moodle logs into educational actions and classifies quiz submissions as active. The event-grouping passage (section label “–”; page unavailable) associates forum reading with reflective exploration. Section 5.1 tracks the number of quiz attempts, but describes it as an indicator of consistency and strategy; it does not specifically link repeated attempts to active engagement. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

Learning Management Systems (LMSs) remain largely static and administrative, often failing to support personalization and inclusive access to learning resources.

“Learning Management Systems (LMSs) remain largely static and administrative, often failing to support personalization and inclusive access to learning resources.”

• cites [1] Angeioplastis, A., Aliprantis, J., Konstantakis, M., and Tsimpiris, A. (2025). Predicting Student Performance and Enhancing Learning Outcomes: A Data-Driven Approach Using Educational Data Mining Techniques. Computers, 14. sourceNot checked — no full text of the source
• cites [2] Angeioplastis, A., Aliprantis, J., Konstantakis, M., Varsamis, D., and Tsimpiris, A. (2025). The Learning Style Decoder: FSLSM-Guided Behavior Mapping Meets Deep Neural Prediction in LMS Settings. Computers, 14. sourcePartly supported

  Section 4 (Personalization Process; page unavailable) says many existing approaches rely on static learner profiling and identifies a need for more personalized learning pathways. It does not say that LMSs themselves are largely static or administrative, or that they fail to provide inclusive access to learning resources. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Our previous work [2] proposed a hybrid learner profiling methodology that integrates FSLSM questionnaire data with fine-grained Moodle log analytics.

“A representative example is our previous work [2], where we proposed a hybrid learner profiling methodology that integrates FSLSM questionnaire data with fine-grained Moodle log analytics.”

• cites [2] Angeioplastis, A., Aliprantis, J., Konstantakis, M., Varsamis, D., and Tsimpiris, A. (2025). The Learning Style Decoder: FSLSM-Guided Behavior Mapping Meets Deep Neural Prediction in LMS Settings. Computers, 14. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Early ALS implementations often required extensive manual rule design or pre-course diagnostic assessments to initialize learner models.

“Early ALS implementations often required extensive manual rule design or pre-course diagnostic assessments to initialize learner models [13].”

• cites [13] Brusilovsky, 2001, Adaptive hypermedia, User Model. User-Adapt. Interact. sourcePartly supported

  Section 3.2.1 says individual traits were traditionally obtained through psychological tests and describes educational systems that adapted to learning styles. Section 3.1.1 identifies early educational hypermedia systems, but neither passage establishes that early implementations often required pre-course assessments to initialize learner models. The passages also do not establish extensive manual rule-design requirements. (LLM, conf=0.68)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Concerns related to accuracy, bias, and pedagogical appropriateness of AI-generated content remain active areas of research and debate.

“Nevertheless, concerns related to accuracy, bias, and pedagogical appropriateness of AI-generated content remain active areas of research and debate [17,18].”

• cites [17] Kasneci, 2023, ChatGPT for good? On opportunities and challenges of large language models for education, Learn. Individ. Differ. sourceSupported
• cites [18] Ordoumpozanis, K., Konstantakis, M., Caridakis, G., and Zoi, S.K. (2025, January 24–26). Generative AI: A Systematic Review of Related Interfaces and Interactions. Proceedings of the ACM Greek SIGCHI Chapter 3rd International Conference, Syros, Greece. sourceNot checked — no full text of the source

May combine two claims.

This design mitigates hallucination risks and aligns chatbot responses with curricular objectives.

“This design mitigates hallucination risks and aligns chatbot responses with curricular objectives [17].”

• cites [17] Kasneci, 2023, ChatGPT for good? On opportunities and challenges of large language models for education, Learn. Individ. Differ. sourcePartly supported

  Section 4 (page unavailable) recommends customizing model output to align with the teaching style and curriculum, and using authoritative sources and teacher review to check inaccurate information. It does not evaluate the citing paper’s specific design or establish that the design mitigates hallucinations. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The retrieval-grounded generation approach aligns with recent research on mitigating hallucinations and maintaining factual consistency in educational applications of generative AI.

“This retrieval-grounded generation approach aligns with recent research on mitigating hallucinations and maintaining factual consistency in educational applications of generative AI [17,18].”

• cites [17] Kasneci, 2023, ChatGPT for good? On opportunities and challenges of large language models for education, Learn. Individ. Differ. sourcePartly supported

  Section 4, “Key Challenges and Risks Related to the Application of Large Language Models in Education” (page unavailable), warns that students may accept convincing but unverified output and recommends checking model responses against authoritative sources to ensure accuracy. This supports the concern about factual consistency in educational AI, but the paper does not discuss retrieval-grounded generation as a way to mitigate hallucinations. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [18] Ordoumpozanis, K., Konstantakis, M., Caridakis, G., and Zoi, S.K. (2025, January 24–26). Generative AI: A Systematic Review of Related Interfaces and Interactions. Proceedings of the ACM Greek SIGCHI Chapter 3rd International Conference, Syros, Greece. sourceNot checked — no full text of the source

May combine two claims.

No learning resources are hidden; personalization affects prioritization and recommendations rather than access.

“Importantly, no learning resources are hidden; personalization affects prioritization and recommendations rather than access [21,22].”

• cites [21] Nafea, 2019, On recommendation of learning objects using Felder–Silverman learning style model, IEEE Access sourcePartly supported

  The Introduction describes personalized top-n learning-object lists based on predicted ratings, supporting the claim that personalization affects recommendations and prioritization (page unavailable). The Dataset section says students rated learning objects, but neither it nor the retrieved passages says that all resources remain visible or accessible; that part is unsupported. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [22] Fayaza, M.S.F., and Ahangama, S. (2024). Use of Felder–Silverman Learning Style Model in Technology-Enhanced Learning. Proceedings of the 2024 International Research Conference on Smart Computing and Systems Engineering (SCSE), Colombo, Sri Lanka, 4 April 2024, IEEE. Available online: https://ieeexplore.ieee.org/document/10550775. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The chatbot supports multiple pedagogical interaction styles, including on-demand explanations, example generation, and formative questioning, thereby accommodating both reflective learners seeking clarification and active learners engaging in dialogic exploration.

“The chatbot supports multiple pedagogical interaction styles, including on-demand explanations, example generation, and formative questioning, thereby accommodating both reflective learners seeking clarification and active learners engaging in dialogic exploration.”

May combine two claims.

The chatbot is powered by a large language model (LLM) that operates in a retrieval-augmented generation (RAG) setting: the textbook content is indexed, and contextually relevant passages are retrieved and injected into the model prompt to ensure that responses remain grounded in the authoritative course material.

“The chatbot is powered by a large language model (LLM) that operates in a retrieval-augmented generation (RAG) setting: the textbook content is indexed, and contextually relevant passages are retrieved and injected into the model prompt to ensure that responses remain grounded in the authoritative course material.”

Classified as probably not a substantial claim. May combine two claims.

PREPARE combines three complementary pillars: generative AI for automated multimodal content creation, learning analytics and behavioral data analysis for continuous learner profiling, and adaptive learning mechanisms that personalize the presentation and prioritization of learning resources.

“At a conceptual level, PREPARE combines three complementary pillars: (i) generative artificial intelligence for automated multimodal content creation, (ii) learning analytics and behavioral data analysis for continuous learner profiling, and (iii) adaptive learning mechanisms that personalize the presentation and prioritization of learning resources.”

May combine two claims.

The resulting profiles inform adaptive mechanisms that tailor content ordering, recommendations, and interaction styles to individual learners, supporting personalization without restricting access to alternative learning paths.

“The resulting profiles inform adaptive mechanisms that tailor content ordering, recommendations, and interaction styles to individual learners, supporting personalization without restricting access to alternative learning paths.”

Classified as probably not a substantial claim.

These data are analyzed to construct and update learner profiles that reflect both self-reported characteristics and observed learning behaviors.

“These data are analyzed to construct and update learner profiles that reflect both self-reported characteristics and observed learning behaviors.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the reported deployment, approximately 12 courses were produced, each comprising around 13 chapters, with each chapter further decomposed into about 5 learning sub-units, totaling approximately 780 sub-units.

“In the reported deployment, we produced approximately 12 courses, each comprising around 13 chapters, with each chapter further decomposed into about 5 learning sub-units (i.e., ≈780 sub-units in total).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Media-intensive artifacts were generated and uploaded in staged batches due to external service quotas during the NotebookLM pre-release phase, whereas text-based assets, interactive elements, and LMS links could be deployed immediately.

“In practice, media-intensive artifacts (e.g., narrated video lectures) were generated and uploaded in staged batches due to external service quotas during the NotebookLM pre-release phase, whereas text-based assets, interactive elements, and LMS links could be deployed immediately.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

During pilot generation, a multiple-choice item was rejected because the proposed distractors were conceptually implausible and the answer key contained ambiguous phrasing not directly supported by the source text.

“For illustration, during pilot generation, a multiple-choice item was rejected because the proposed distractors were conceptually implausible and the answer key contained ambiguous phrasing not directly supported by the source text.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Course content and AI-generated assets have already been deployed in Moodle, but learner-facing use and data collection will take place during the spring-semester classroom delivery.

“While the course content and AIgenerated assets have already been deployed in Moodle, learner-facing use and data collection will take place during the spring-semester classroom delivery.”

Classified as probably not a substantial claim.

Each sub-unit is delivered as a standardized multimodal package (text, slides, video, podcast, quiz, and chatbot).

“Each sub-unit is delivered as a standardized multimodal package (text, slides, video, podcast, quiz, and chatbot), enabling consistent measurement of cross-modality engagement in the upcoming study.”

• Also stated in: § 5.7. System-Level Validation and Preliminary Evidence · plain

  “Each sub-unit included a standardized multimodal package consisting of textual material, slide decks, video or audio resources, formative quizzes, and access to the AI-powered chatbot tutor.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The choice of k = 4 was informed by pilot experimentation balancing contextual sufficiency and prompt-length constraints, as higher values increased redundancy without measurable grounding gains.

“The choice of k = 4 was informed by pilot experimentation balancing contextual sufficiency and prompt-length constraints, as higher values increased redundancy without measurable grounding gains.”

May combine two claims.

PREPARE demonstrates an automation-first authoring workflow that converts a single textbook into a deployable Moodle course through quota-aware orchestration and HTML-based packaging, enabling repeatable course-scale production without exposing internal code.

“Beyond the educational benefits, PREPARE demonstrates an automation-first authoring workflow that converts a single textbook into a deployable Moodle course through quota-aware orchestration and HTML-based packaging, enabling repeatable course-scale production without exposing internal code [23].”

• cites [23] Pratschke, B.M. (2024). Generative AI and Education: Digital Pedagogies, Teaching Innovation and Learning Design, Springer Nature. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The similarity threshold (0.75) was empirically calibrated during controlled testing to minimize unsupported responses while preserving coverage.

“The similarity threshold (0.75) was empirically calibrated during controlled testing to minimize unsupported responses while preserving coverage.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Media-intensive resources (e.g., narrated videos) were produced using staged batching due to external service quotas.

“Deployment logs confirmed that all core artifact types were successfully generated and published to Moodle, with media-intensive resources (e.g., narrated videos) produced using staged batching due to external service quotas.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Felder-Silverman Learning Style Model (FSLSM) categorizes learners along four dimensions: active/reflective, sensing/intuitive, visual/verbal, and sequential/global.

“A widely adopted framework for capturing individual learner differences is the Felder-Silverman Learning Style Model (FSLSM), which categorizes learners along four dimensions: active/reflective, sensing/intuitive, visual/verbal, and sequential/global [7].”

• cites [7] Felder, 1988, Learning and teaching styles in engineering education, Eng. Educ. sourceNot checked — no full text of the source

May combine two claims.

Our recent study [1] applied multiple classifiers (including k-nearest neighbors, random forests, and neural networks) to Moodle log data from 450 students, achieving high predictive performance (F1 scores exceeding 0.8) for early identification of at-risk learners.

“Building on this line of work, our recent study [1] applied multiple classifiers (including k-nearest neighbors, random forests, and neural networks) to Moodle log data from 450 students, achieving high predictive performance (F1 scores exceeding 0.8) for early identification of at-risk learners.”

• cites [1] Angeioplastis, A., Aliprantis, J., Konstantakis, M., and Tsimpiris, A. (2025). Predicting Student Performance and Enhancing Learning Outcomes: A Data-Driven Approach Using Educational Data Mining Techniques. Computers, 14. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

FSLSM-based personalization has traditionally relied on explicit self-report instruments, most notably the Index of Learning Styles (ILS), administered at the beginning of a course to infer learner preferences.

“Traditionally, FSLSM-based personalization has relied on explicit self-report instruments, most notably the Index of Learning Styles (ILS), administered at the beginning of a course to infer learner preferences.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Multimedia resources (video narration, captions) are revised if transcription errors, mispronunciations of domain terminology, or caption misalignments are detected.

“Multimedia resources (video narration, captions) are revised if transcription errors, mispronunciations of domain terminology, or caption misalignments are detected.”

May combine two claims.

This approach preserves course-level consistency while enabling personalized learning pathways at the individual level.

“This approach preserves course-level consistency while enabling personalized learning pathways at the individual level.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The current PREPARE deployment follows a model-agnostic implementation strategy to support portability and reproducibility.

“The current PREPARE deployment follows a model-agnostic implementation strategy to support portability and reproducibility.”

May combine two claims.

The underlying course structure remains identical for all learners, but presentation order and recommendation cues are dynamically adapted at runtime based on the learner profile.

“While the underlying course structure remains identical for all learners, the presentation order and recommendation cues are dynamically adapted at runtime based on the learner profile, as described in Section 4.2.”

Probably not a substantial claim: may not pass S3 (load-bearing).

AI-driven technologies have increasingly been explored as enablers of inclusivity, supporting automated captioning, text-to-speech, content simplification, and alternative media generation.

“AI-driven technologies have increasingly been explored as enablers of such inclusivity, supporting automated captioning, text-to-speech, content simplification, and alternative media generation [16].”

• cites [16] Bond, 2021, Systematic review of research on artificial intelligence applications in higher education, Int. J. Educ. Technol. High. Educ.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Validation criteria differ slightly by resource type but follow a common quality-control framework including factual alignment with the authoritative textbook, pedagogical coherence, clarity and linguistic precision, cognitive level appropriateness, and accessibility compliance.

“Validation criteria differ slightly by resource type but follow a common quality-control framework including (i) factual alignment with the authoritative textbook, (ii) pedagogical coherence, (iii) clarity and linguistic precision, (iv) cognitive level appropriateness, and (v) accessibility compliance.”

• AssumptionRests on an assumption: The authoritative textbook content is factually correct and pedagogically coherent, so it can serve as ground truth for validating AI-generated artifacts.

May combine two claims.

This workflow reduces repetitive instructor authoring tasks and enables consistent chapter-level packaging of resources across the course while preserving human oversight for validation and final curation.

“This workflow reduces repetitive instructor authoring tasks and enables consistent chapter-level packaging of resources (summary, slides, quizzes, media, and support tools) across the course while preserving human oversight for validation and final curation.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Resources requiring substantial correction are regenerated using refined prompts, and version control metadata records both original and validated outputs.

“Resources requiring substantial correction are regenerated using refined prompts, and version control metadata records both original and validated outputs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Items failing any validation checks are either corrected or regenerated.

“Items failing any of these checks are either corrected or regenerated.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the current deployment, the PDF textbook is treated as the authoritative input source to ensure traceability, reproducibility, and controlled grounding of AI-generated outputs.

“In the current deployment, the PDF textbook is treated as the authoritative input source to ensure traceability, reproducibility, and controlled grounding of AI-generated outputs.”

• AssumptionRests on an assumption: The source PDF textbook is of sufficient quality, completeness, and pedagogical coherence to serve as the single authoritative source for all generated learning assets.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The automation substantially reduces instructor workload while preserving the instructor's ability to modify, extend, or curate the generated materials.

“This automation substantially reduces instructor workload while preserving the instructor's ability to modify, extend, or curate the generated materials.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

For slide decks, resources are revised if visual density is excessive, conceptual sequencing is inconsistent with the chapter structure, or terminology diverges from the source material.

“For slide decks, resources are revised if visual density is excessive, conceptual sequencing is inconsistent with the chapter structure, or terminology diverges from the source material.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The methodological design of PREPARE emphasizes reproducibility and real-world applicability.

“The methodological design of PREPARE emphasizes reproducibility and real-world applicability.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

For quiz items, validation checks target incorrect answer keys, ambiguous phrasing, distractors lacking plausibility, or misalignment with stated learning objectives.

“For quiz items, validation checks target incorrect answer keys, ambiguous phrasing, distractors lacking plausibility, or misalignment with stated learning objectives.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

For textual summaries and notes, rejection criteria include omission of core concepts, introduction of unsupported claims, or oversimplification that alters meaning.

“For textual summaries and notes, rejection criteria include omission of core concepts, introduction of unsupported claims, or oversimplification that alters meaning.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Human review is applied selectively to quiz items to ensure clarity, difficulty balance, and pedagogical appropriateness before learner-facing deployment.

“Human review is applied selectively to ensure clarity, difficulty balance, and pedagogical appropriateness before learner-facing deployment.”

Classified as probably not a substantial claim. May combine two claims.

This enables learners to verify responses against the source material and reinforces transparency and trust in AI-assisted learning support.

“This enables learners to verify responses against the source material and reinforces transparency and trust in AI-assisted learning support.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Validation is exhaustive for core pedagogical artifacts (chapter summaries, slide decks, and quiz items prior to publication), while multimedia artifacts undergo targeted review focusing on terminology accuracy, caption alignment, and accessibility compliance.

“Validation is exhaustive for core pedagogical artifacts (chapter summaries, slide decks, and quiz items prior to publication), while multimedia artifacts undergo targeted review focusing on terminology accuracy, caption alignment, and accessibility compliance.”

May combine two claims.

Video captions are automatically generated and manually corrected to ensure accessibility, in line with Universal Design for Learning principles.

“Video captions are automatically generated and manually corrected to ensure accessibility, in line with Universal Design for Learning principles [15,16].”

• cites [15] Rose, D.H., and Meyer, A. (2006). A Practical Reader in Universal Design for Learning, Harvard Education Press. sourceNot checked — no full text of the source
• cites [16] Bond, 2021, Systematic review of research on artificial intelligence applications in higher education, Int. J. Educ. Technol. High. Educ.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Universal Design for Learning (UDL) principles advocate providing multiple means of representation, engagement, and expression to accommodate diverse learner needs.

“Universal Design for Learning (UDL) principles advocate providing multiple means of representation, engagement, and expression to accommodate diverse learner needs [15].”

• cites [15] Rose, D.H., and Meyer, A. (2006). A Practical Reader in Universal Design for Learning, Harvard Education Press. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

All video resources include captions to support learners with hearing impairments and to enhance comprehension for non-native speakers.

“All video resources include captions to support learners with hearing impairments and to enhance comprehension for non-native speakers.”

Classified as probably not a substantial claim. May combine two claims.

When the retrieved content does not sufficiently support a learner query, the assistant either requests clarification or explicitly states that the information is not covered by the reference material.

“When the retrieved content does not sufficiently support a learner query, the assistant either requests clarification or explicitly states that the information is not covered by the reference material.”

Probably not a substantial claim: may not pass S3 (load-bearing).

All chatbot answers include lightweight internal references to the originating chapter and section of the textbook used for grounding.

“All chatbot answers include lightweight internal references to the originating chapter and section of the textbook used for grounding.”

May combine two claims.

The engineering choices were essential to move from prototype generation to repeatable, scalable course publishing while keeping the codebase internal and the process auditable.

“These engineering choices were essential to move from prototype generation to repeatable, scalable course publishing, while keeping the codebase internal and the process auditable through produced artifacts and LMS-side results.”

Classified as probably not a substantial claim.

To further ensure pedagogical safety, the chatbot operates under a supported-only response policy.

“To further ensure pedagogical safety, the chatbot operates under a supported-only response policy.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Prior research has demonstrated that augmented reality can enhance engagement and conceptual understanding, particularly for experiential and visual learners.

“Prior research has demonstrated that AR can enhance engagement and conceptual understanding, particularly for experiential and visual learners [19].”

• cites [19] Baldiris, 2019, Systematic review and meta-analysis of augmented reality in educational settings, Educ. Res. Rev.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Prior to release, instructors review generated slide decks to verify pedagogical coherence, terminology consistency, and accessibility.

“Prior to release, instructors review the slides to verify pedagogical coherence, terminology consistency, and accessibility (e.g., slide readability and visual density).”

Classified as probably not a substantial claim.

Quiz items undergo validation against the authoritative content to detect inconsistencies or unsupported claims.

“In a second step, items undergo validation against the authoritative content to detect inconsistencies or unsupported claims.”

• AssumptionRests on an assumption: The authoritative textbook content is factually correct and pedagogically coherent, so it can serve as ground truth for validating AI-generated artifacts.

Probably not a substantial claim: may not pass S3 (load-bearing).

In such cases, the system avoids speculative or persuasive responses.

“In such cases, the system avoids speculative or persuasive responses.”

Classified as probably not a substantial claim.

All inputs are normalized into a unified internal representation (textual chunks with metadata and provenance), which enables consistent retrieval-augmented generation, analytics, and LMS deployment regardless of the original modality.

“All inputs are normalized into a unified internal representation (textual chunks with metadata and provenance), which enables consistent retrieval-augmented generation, analytics, and LMS deployment regardless of the original modality.”

May combine two claims.

This automation significantly reduces instructor workload while maintaining consistency across learning resources and supports scalable adoption in institutional settings.

“This automation significantly reduces instructor workload while maintaining consistency across learning resources and supports scalable adoption in institutional settings [3].”

• cites [3] Tan, 2025, Artificial intelligence-enabled adaptive learning platforms: A review, Comput. Educ. Artif. Intell. sourceNot checked — no full text of the source

Classified as probably not a substantial claim. May combine two claims.

The preprocessing step establishes a machine-readable content backbone that supports reproducible AI-driven transformations and mirrors the pedagogical structure of the original material within Moodle.

“This preprocessing step establishes a machine-readable content backbone that supports reproducible AI-driven transformations and mirrors the pedagogical structure of the original material within Moodle.”

May combine two claims.

This decoupling allowed continuous course publication without blocking the entire pipeline on rate-limited steps, reinforcing the operational robustness of the workflow.

“This decoupling allowed continuous course publication without blocking the entire pipeline on rate-limited steps, reinforcing the operational robustness of the workflow.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

All AI-generated artifacts undergo structured human validation prior to publication.

“All AI-generated artifacts undergo structured human validation prior to publication.”

Classified as probably not a substantial claim. May combine two claims.

This design choice enables the deployment of advanced AI-driven personalization capabilities while preserving compatibility with existing institutional infrastructures and pedagogical practices.

“This design choice enables the deployment of advanced AI-driven personalization capabilities while preserving compatibility with existing institutional infrastructures and pedagogical practices.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

All generated resources remain anchored to the same authoritative source, reducing the risk of content drift or inconsistency.

“Importantly, all generated resources remain anchored to the same authoritative source, reducing the risk of content drift or inconsistency.”

May combine two claims.

The system-level validation and deployment readiness suggest that AI-augmented LMS workflows can reduce instructor authoring effort while improving flexibility and inclusivity, provided that human-in-the-loop validation and privacy-aware analytics are embedded from the outset.

“The system-level validation and deployment readiness suggest that AI-augmented LMS workflows can reduce instructor authoring effort while improving flexibility and inclusivity, provided that human-in-the-loop validation and privacy-aware analytics are embedded from the outset.”

• AssumptionRests on an assumption: The generated multimodal resources (summaries, slides, quizzes, etc.) are pedagogically appropriate and accurate enough to support learning without additional expert review beyond the stated human oversight.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

At the time of writing, the intervention has been fully deployed and technically validated in the target Moodle environment, and the evaluation instruments (questionnaires and logging configuration) have been finalized.

“At the time of writing, the intervention has been fully deployed and technically validated in the target Moodle environment, and the evaluation instruments (questionnaires and logging configuration) have been finalized.”

Probably not a substantial claim: may not pass S4 (based).

Adaptation affects only recommendation cues and presentation order, such as highlighting entry-point resources or adjusting chatbot response style, without restricting access to any modality.

“Adaptation affects only recommendation cues and presentation order, such as highlighting entry-point resources or adjusting chatbot response style, without restricting access to any modality.”

May combine two claims.

PREPARE contributes a transferable architectural model for AI-augmented LMSs, combining automation, personalization, and inclusivity within a scalable and institutionally deployable framework.

“PREPARE contributes a transferable architectural model for AI-augmented LMSs, combining automation, personalization, and inclusivity within a scalable and institutionally deployable framework.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The mapping key between participant codes and identities is stored separately and accessible only to the designated principal investigator.

“The mapping key between participant codes and identities is stored separately and accessible only to the designated principal investigator.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Data collection follows strict minimization principles.

“Data collection follows strict minimization principles.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Access to raw data is limited to the core research team explicitly named in the study protocol.

“Access to raw data is limited to the core research team explicitly named in the study protocol.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

PREPARE illustrates how engagement-focused evaluation metrics, learner agency, and ethical safeguards can coexist within a scalable intelligent LMS architecture, providing a practical pathway toward trustworthy AI adoption in education.

“PREPARE illustrates how engagement-focused evaluation metrics, learner agency, and ethical safeguards can coexist within a scalable intelligent LMS architecture, providing a practical pathway toward trustworthy AI adoption in education.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Where multiple inferential tests are conducted, interpretation will adopt a conservative stance, and effect sizes will accompany significance testing to mitigate inflation of Type I error.

“Where multiple inferential tests are conducted, interpretation will adopt a conservative stance, and effect sizes will accompany significance testing to mitigate inflation of Type I error.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Transparency mechanisms, opt-out options, and human oversight contributed to learner confidence and institutional compliance.

“Transparency mechanisms, opt-out options, and human oversight contributed to learner confidence and institutional compliance.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Raw interaction logs and questionnaire data will be retained for a maximum period of three years following study completion for research audit and publication purposes, after which they will be securely deleted.

“Raw interaction logs and questionnaire data will be retained for a maximum period of three years following study completion for research audit and publication purposes, after which they will be securely deleted.”

• Also stated in: § 5.8. Ethical and Data Protection Considerations · plain

  “Data retention is limited to a maximum of three years following study completion for audit and publication purposes, after which raw data will be securely deleted.”

Classified as probably not a substantial claim.

By demonstrating how generative AI, learner modeling, and adaptive delivery can be operationalized within a widely adopted LMS, this work contributes a concrete step toward scalable, trustworthy, and inclusive AI-supported education.

“By demonstrating how generative AI, learner modeling, and adaptive delivery can be operationalized within a widely adopted LMS, this work contributes a concrete step toward scalable, trustworthy, and inclusive AI-supported education.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Primary outcome indicators (engagement intensity, interaction diversity, and perceived personalization usefulness) are predefined to limit interpretive flexibility.

“Primary outcome indicators (engagement intensity, interaction diversity, and perceived personalization usefulness) are predefined to limit interpretive flexibility.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Generative content creation operates exclusively on instructional materials.

“Generative content creation operates exclusively on instructional materials.”

Classified as probably not a substantial claim.

The system demonstrates how AI can function as an enabling layer that supports educators and empowers learners, rather than replacing instructional design or teaching practice.

“Rather than positioning generative AI as a replacement for instructional design or teaching practice, the system demonstrates how AI can function as an enabling layer that supports educators and empowers learners.”

Classified as probably not a substantial claim.

Basic input filtering mechanisms detect common personal identifiers (e.g., phone numbers, email patterns) and prevent submission where detected.

“Additionally, basic input filtering mechanisms detect common personal identifiers (e.g., phone numbers, email patterns) and prevent submission where detected.”

Classified as probably not a substantial claim.

Scale constructs were adapted from previously validated instruments (e.g., Technology Acceptance Model-derived perceived usefulness and ease-of-use measures, established engagement and trust scales), with contextual wording adjustments for the PREPARE deployment.

“Scale constructs were adapted from previously validated instruments (e.g., Technology Acceptance Model-derived perceived usefulness and ease-of-use measures, established engagement and trust scales), with contextual wording adjustments for the PREPARE deployment.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

This retrieval strategy ensures traceability, reduces hallucination risk, and enables reproducibility of context selection decisions across deployments.

“This retrieval strategy ensures traceability, reduces hallucination risk, and enables reproducibility of context selection decisions across deployments.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Identifiers are replaced by random participant codes, and access to the mapping is restricted.

“Identifiers are replaced by random participant codes, and access to the mapping is restricted.”

May combine two claims.

When external LLM services are used for response generation, requests are transmitted without user identifiers, and only the anonymized query and retrieved textbook fragments are included in the prompt context.

“When external LLM services are used for response generation, requests are transmitted without user identifiers, and only the anonymized query and retrieved textbook fragments are included in the prompt context.”

May combine two claims.

Diversity promotion never overrides declared accessibility needs; accessibility settings are treated as hard constraints, whereas learning-style preferences function as soft signals.

“Importantly, diversity promotion never overrides declared accessibility needs.Accessibility settings (e.g., requirements for captions, audio alternatives, or reduced visual complexity) are treated as hard constraints, whereas learning-style preferences function as soft signals.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Fallback modalities ensured that no learner was excluded, reinforcing inclusive design principles.

“While AR access remains dependent on device availability, fallback modalities ensured that no learner was excluded, reinforcing inclusive design principles.”

Probably not a substantial claim: may not pass S4 (based).

The availability of multimodal content-textual, audio, video, and immersive AR-substantially reduced common barriers to participation, particularly for learners with disabilities, language challenges, or attentional constraints.

“The availability of multimodal content-textual, audio, video, and immersive AR-substantially reduced common barriers to participation, particularly for learners with disabilities, language challenges, or attentional constraints.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

The system operationalizes adaptive learning in a manner consistent with responsible AI principles, aligning with GDPR requirements and anticipating emerging regulatory frameworks such as the EU AI Act.

“Importantly, the system operationalizes adaptive learning in a manner consistent with responsible AI principles, aligning with GDPR requirements and anticipating emerging regulatory frameworks such as the EU AI Act.”

Probably not a substantial claim: may not pass S4 (based).

PREPARE demonstrates that personalization based on learner preferences and observed behavior can positively influence engagement.

“Consistent with prior research on adaptive learning systems, PREPARE demonstrates that personalization based on learner preferences and observed behavior can positively influence engagement.”

Probably not a substantial claim: may not pass S3 (load-bearing).

All collected data (LMS interaction logs and questionnaire responses) are stored on institutional infrastructure under restricted access conditions.

“All collected data (LMS interaction logs and questionnaire responses) are stored on institutional infrastructure under restricted access conditions.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

All generated artifacts are associated with a content hash of the source material, a prompt-template identifier, and a timestamped revision ID; this mechanism allows auditing, controlled regeneration when models or prompts change, and comparison across versions.

“All generated artifacts (summaries, slides, quiz items, and narrations) are associated with a content hash of the source material, a prompttemplate identifier, and a timestamped revision ID.”

• Also stated in: § Implementation Details and Traceability · plain

  “This mechanism allows auditing, controlled regeneration when models or prompts change, and comparison across versions.”

Classified as probably not a substantial claim. May combine two claims.

Before publication in Moodle, AI-generated materials pass through a human-in-the-loop validation step where instructors or project researchers review factual consistency with the textbook, pedagogical clarity, and accessibility compliance; artifacts failing validation are corrected or discarded, ensuring only verified content is deployed.

“Before publication in Moodle, AI-generated materials pass through a human-in-theloop validation step.Instructors or project researchers review factual consistency with the textbook, pedagogical clarity, and accessibility compliance.Artifacts failing validation are corrected or discarded, ensuring that only verified content is deployed in the LMS.”

• AssumptionRests on an assumption: The authoritative textbook content is factually correct and pedagogically coherent, so it can serve as ground truth for validating AI-generated artifacts.
• AssumptionRests on an assumption: Instructors or project researchers can reliably assess factual consistency, pedagogical clarity, and accessibility compliance of AI-generated materials.

Classified as probably not a substantial claim. May combine two claims.

This hybrid trigger prevents over-sensitivity to isolated actions while allowing adaptation to evolving engagement patterns.

“This hybrid trigger prevents over-sensitivity to isolated actions while allowing adaptation to evolving engagement patterns.”

May combine two claims.

Based on the current learner profile, PREPARE adjusts content prioritization, recommendations, and interface ordering while still encouraging exposure to alternative learning modalities.

“Based on the current learner profile, PREPARE adjusts content prioritization, recommendations, and interface ordering-for example, emphasizing visual or interactive resources for learners exhibiting strong visual engagement-while still encouraging exposure to alternative learning modalities.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

All learner interaction logs remain within the institutional Moodle environment.

“All learner interaction logs remain within the institutional Moodle environment.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Data are pseudonymized at collection using random participant identifiers.

“Data are pseudonymized at collection using random participant identifiers.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Learner profiles are recalculated periodically rather than continuously in real time; in the current deployment, behavioral indicators are aggregated over a rolling two-week window or after a minimum threshold of 20 meaningful interaction events, whichever occurs first.

“Learner profiles are recalculated periodically rather than continuously in real time.In the current deployment, behavioral indicators are aggregated over a rolling two-week window or after a minimum threshold of 20 meaningful interaction events (e.g., resource views, quiz attempts, chatbot exchanges), whichever occurs first.”

Classified as probably not a substantial claim. May combine two claims.

These comparisons serve two purposes: providing metacognitive feedback to learners through dashboards or summary reports, and informing adaptive system behavior.

“These comparisons serve two purposes: (i) providing metacognitive feedback to learners through dashboards or summary reports, fostering awareness of their learning patterns, and (ii) informing adaptive system behavior.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

When external LLM services are used for conversational support, requests are transmitted without personal identifiers and include only the learner query and retrieved textbook fragments required for grounding.

“When external LLM services are used for conversational support, requests are transmitted without personal identifiers and include only the learner query and retrieved textbook fragments required for grounding.”

May combine two claims.

PREPARE is effective in reducing instructor authoring effort, ensuring consistent multimodal availability, and enabling fine-grained learning analytics and adaptive delivery in real educational settings.

“While these findings do not yet constitute learner outcome evidence, they provide empirical confirmation that PREPARE is effective in reducing instructor authoring effort, ensuring consistent multimodal availability, and enabling fine-grained learning analytics and adaptive delivery in real educational settings.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Grounding responses strictly in course material and promoting verification mitigated concerns related to misinformation, reinforcing responsible AI use in educational contexts.

“Grounding responses strictly in course material and promoting verification mitigated concerns related to misinformation, reinforcing responsible AI use in educational contexts.”

Probably not a substantial claim: may not pass S4 (based).

Discrepancies between self-reported and behavior-inferred signals are monitored over time and interpreted as evidence of evolving strategies rather than model error.

“(iii) dis-crepancies between self-reported and behavior-inferred signals are monitored over time and interpreted as evidence of evolving strategies rather than model error;”

May combine two claims.

The evaluation focused exclusively on secondary education students, highlighting the applicability of AI-driven adaptive LMS workflows in school-level contexts rather than higher education environments.

“It should be noted that the evaluation focused exclusively on secondary education students, highlighting the applicability of AI-driven adaptive LMS workflows in school-level contexts rather than higher education environments.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The system-level validation and the evaluation design suggest the potential for improvements in learner engagement, interaction diversity, and perceived accessibility.

“The system-level validation and the evaluation design suggest the potential for improvements in learner engagement, interaction diversity, and perceived accessibility.”

Classified as probably not a substantial claim. May combine two claims.

The coarse-grained adaptation logic reduces the risk of over-adaptation and preserves learner autonomy.

“While this choice reduces the risk of over-adaptation and preserves learner autonomy, future extensions may incorporate adaptive difficulty and mastery-based progression to further enhance personalization.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

To avoid reinforcing narrow usage patterns, the adaptive logic incorporates a diversity constraint: at regular intervals, alternative modalities are intentionally promoted alongside dominant preferences; this ensures exposure to varied representations and mitigates the risk of preference lock-in or bias amplification.

“To avoid reinforcing narrow usage patterns, the adaptive logic incorporates a diversity constraint: at regular intervals, alternative modalities are intentionally promoted alongside dominant preferences.This mechanism ensures exposure to varied representations and mitigates the risk of preference lock-in or bias amplification.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

By restricting AI processing to instructional content and maintaining learner modeling within institutional boundaries, PREPARE aligns with GDPR principles and anticipates emerging requirements of the EU AI Act for high-risk educational AI systems.

“By restricting AI processing to instructional content and maintaining learner modeling within institutional boundaries, PREPARE aligns with GDPR principles and anticipates emerging requirements of the EU AI Act for high-risk educational AI systems.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

This work highlights a practical pathway toward trustworthy, learner-centered intelligent LMSs and supports the broader vision of inclusive and adaptive education at scale.

“Overall, this work highlights a practical pathway toward trustworthy, learner-centered intelligent LMSs and supports the broader vision of inclusive and adaptive education at scale.”

Classified as probably not a substantial claim.

A central contribution of PREPARE lies in its accessibility-by-design approach.

“A central contribution of PREPARE lies in its accessibility-by-design approach.”

Probably not a substantial claim: may not pass S4 (based).

Learners retain full agency to select any modality, ensuring that adaptivity supports rather than constrains individual learning pathways.

“(iv) learners retain full agency to select any modality, ensuring that adaptivity supports, rather than constrains, individual learning pathways.”

May combine two claims.

The reported system-level validation and the planned classroom study position PREPARE as a deployment-ready framework whose system-level effectiveness can be assessed both in terms of technical feasibility and learner-centered outcomes.

“Taken together, the reported system-level validation and the planned classroom study position PREPARE as a deployment-ready framework whose system-level effectiveness can be assessed both in terms of technical feasibility and learner-centered outcomes.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Mismatches between the initial self-reported profile and the behavior-derived profile are interpreted as indicators of evolving learning strategies or misalignment between perceived and actual preferences.

“Such mismatches are interpreted as indicators of evolving learning strategies or misalignment between perceived and actual preferences.”

Probably not a substantial claim: may not pass S4 (based).

The integration of AR content further supported learners who struggle with abstract representations, offering contextualized and experiential learning opportunities.

“The integration of AR content further supported learners who struggle with abstract representations, offering contextualized and experiential learning opportunities.”

Classified as probably not a substantial claim.

Evaluation instruments include structured pre- and post-use questionnaires composed of validated scale constructs adapted to the PREPARE context.

“Evaluation instruments include structured pre-and post-use questionnaires composed of validated scale constructs adapted to the PREPARE context.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

All learner data collection and profiling processes adhere to strict privacy and data protection principles; behavioral analysis is conducted on anonymized identifiers, and access to individual learner profiles is restricted to the learner and authorized personnel; these practices align with responsible learning analytics and GDPR-compliant data handling.

“All learner data collection and profiling processes adhere to strict privacy and data protection principles.Behavioral analysis is conducted on anonymized identifiers, and access to individual learner profiles is restricted to the learner and authorized personnel.These practices align with established guidelines for responsible learning analytics and GDPR-compliant data handling.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Only pseudonymized interaction events necessary for the study are stored, and no sensitive personal attributes are collected.

“Only pseudonymized interaction events necessary for the study are stored, and no sensitive personal attributes are collected.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A visible on-screen warning is displayed within the chatbot interface prior to interaction.

“Second, a visible on-screen warning is displayed within the chatbot interface prior to interaction.”