An Accessibility-First Generative AI Chatbot to Support Inclusive MOOC Discovery and Early Onboarding

MOOC accessibility barriers and discovery-layer constraints

MOOCs have expanded participation opportunities, yet access remains uneven for learners who rely on assistive technologies or alternative sensory, cognitive, or linguistic supports (Bozkurt et al., 2024; Farrow et al., 2024; Hamideh Kerdar et al., 2024; Ingavélez-Guerra et al., 2023; Królak et al., 2022). Platform-level accounts commonly describe variability across courses and content types within the same provider rather than treating accessibility as a uniform platform property, reflecting differences in authoring practices, partner institutions, and media formats (Rodriguez-Ascaso et al., 2024).

Two barriers recur across recent analyses. First, video lectures remain a dominant format, yet captions and transcripts are inconsistently available or vary in quality; automated captions may degrade under real-world conditions, reducing reliability for deaf and hard-of-hearing learners and for learners who benefit from text alternatives (Burgstahler, 2023; Lee, 2024; Risi, 2025). Visual learning resources also frequently lack equivalent alternatives (e.g., meaningful alternative text or audio description), limiting usability for blind and low-vision learners. Second, accessibility is often difficult to identify during course selection. When platforms do not support structured filtering by accessibility-relevant features (e.g., captioning, transcripts, language options, screen-reader compatibility indicators), learners may need to manually inspect courses to determine fit an added burden in low-bandwidth contexts and for multilingual users (Aguis, 2024; UNESCO, 2021). Overall, barriers are documented at both content level (missing alternatives) and interaction level (weak accessibility signals during search and early navigation).

Table 1 summarizes what platform documentation and recent evaluations describe about accessibility targets, auditing, and recurring gaps (2024–early 2025), emphasizing reported/observed provisions rather than platform-wide conclusions (Coursera Support, 2025; EdX, 2025; FutureLearn, 2025; Support-Khan Academy, 2025). Across the reviewed platforms, captions are frequently present but inconsistently implemented; alternatives for visuals are uneven and often depend on course teams; and real-time transcription is not described as a standard feature in the sources reviewed. These patterns motivate support earlier in the learner pathway (discovery and early navigation), not only course-by-course remediation.

GenAI Affordances for Multimodal and Multilingual Access

Natural Language Processing (NLP) has advanced from symbolic systems to transformer-based architectures that support context-sensitive generation and multilingual processing, enabling conversational interaction at scale (Goel & Polepeddi, 2018; Kuddus, 2022). When integrated with speech and translation technologies, LLM-based GenAI can support interaction-level accessibility through speech-to-text (STT), text-to-speech (TTS), multilingual mediation, and structured explanation/summarization (Holmes et al., 2022; Khong et al., 2024; Mrayhi et al., 2025a). These capabilities are particularly relevant to MOOC discovery and onboarding, where learners may require modality alternatives and language support to understand requirements, navigate interfaces, and identify courses that match accessibility needs.

Educational AI has long emphasized personalization through tutoring systems and recommender algorithms, but accessibility-specific outcomes are less consistently evaluated (Rane, 2023). GenAI extends the design space by enabling assistive interaction supports that can reduce friction during early navigation (e.g., speech-enabled assistance; stepwise conversational scaffolding) (Grover, 2024; Kumar & Nagar, 2024). At the same time, the inclusivity of GenAI is constrained by ethical and technical factors, including bias, uneven performance across languages, and underrepresentation of low-resource languages (Educause, 2025). These constraints increase the importance of explicit scope statements and transparent reporting of what is implemented and evaluated in accessibility-oriented GenAI systems.

Educational GenAI chatbots: reported supports and recurrent limitations

GenAI chatbots are increasingly deployed to support tutoring, navigation, and engagement, yet accessibility-oriented design and evaluation are inconsistently specified. Targeted studies indicate that chatbots can support accessibility-relevant interaction (e.g., speech output for navigation support; structured conversational scaffolding for comprehension) (Grover, 2024; Kumar & Nagar, 2024). However, published accounts often provide limited detail on multimodal coverage (speech input/output across contexts), multilingual robustness, and whether discovery-layer accessibility attributes are incorporated into guidance (Nataraj et al., 2024; Sarangam, 2024).

Mainstream deployments illustrate similar constraints. Khanmigo offers speech output and multilingual support, but reported coverage does not necessarily extend across platform contexts and may provide limited speech input and disability-sensitive adaptation (Caelen & Blete, 2024; Khan Academy, 2025). Jill Watson has been presented as a scalable virtual teaching assistant for courseware-grounded Q&A in LMS/forum contexts; published accounts foreground teaching-presence support and automation rather than accessibility-oriented interaction supports as evaluation targets (Goel & Polepeddi, 2018; Kakar et al., 2024). Coursera’s AI features emphasize recommendation and learner assistance, yet published descriptions do not consistently indicate that accessibility metadata or disability-responsive supports are incorporated into interaction flows (Coursera Support, 2025; Rao & Lin, 2024). Duolingo’s chatbot incorporates speech and feedback supports within language-learning tasks (Jiang et al., 2024), but domain specificity limits transferability to cross-disciplinary MOOC discovery and navigation (Ouyang et al., 2024).

Because accessibility outcomes cannot be attributed to a chatbot in isolation, Table 2 summarizes these systems using capability-oriented descriptors (what is reported as an interaction support and what is emphasized in evaluation), rather than implying platform-wide accessibility claims. Across studies and deployments, three gaps recur: (1) accessibility-oriented discovery remains limited when accessibility attributes are not consistently surfaced during search and onboarding; (2) assistive interaction capabilities (speech input/output, multilingual mediation, structured scaffolding) are implemented unevenly or described at a high level; and (3) UDL is referenced more often than it is operationalized through reported design logic mapping learner needs to checkpoints and triggerable supports.

Positioning an accessibility-first chatbot for discovery and early navigation

This literature indicates that improving access in MOOCs requires attention to interaction friction during discovery and early navigation, in addition to content-level remediation. It also shows that existing chatbots rarely foreground accessibility as an evaluation target or document how interaction supports operationalize UDL beyond general claims. The RECMOOC4All Chatbot is therefore positioned as an accessibility-first interaction interface that targets discovery and early navigation through multimodal interaction (speech-to-text and text-to-speech), multilingual mediation, and accessibility-aware discovery when accessibility attributes are available via metadata and/or audit outputs. This scope supports a focused evaluation of interaction-layer accessibility supports while acknowledging that content-level accessibility depends on course materials and provider implementation.

Business Understanding

Learners with visual, cognitive/neurodiverse, motor and multilingual needs encounter barriers during MOOC discovery and onboarding, including limited multimodal options, inaccessible navigation pathways and high cognitive load in early decision-making. RECMOOC4All positions the chatbot as the primary interaction interface within an ecosystem that also includes a MOOC aggregator, learner modeling, a recommendation engine and an accessibility checker (Figure 1). Within this scope, the chatbot serves two functions:

1. Interaction and guidance: Supporting discovery-oriented dialogue (course search, clarification, onboarding help and first-step guidance) through text and voice modalities, with multilingual capabilities.

2. Signal capture for personalization: Logging declared preferences and interaction signals (modality choice, reformulations, navigation loops) and passing them to the learner-modeling module to refine recommendations and subsequent interactions.

Figure 1

This delineation clarifies that the chatbot manages the conversational layer, while personalization and accessibility assessments are handled by connected services.

System architecture

The RECMOOC4All ecosystem comprises five coordinated modules (Figure 1):

• GenAI conversational chatbot (interaction layer): manages real-time dialogue, supports discovery for non-registered users and personalization for registered users via backend services, and orchestrates dialogue routing.

• MOOC aggregator: consolidates course metadata and attaches available accessibility-related attributes (e.g., subtitle/caption indicators when available, language, and content descriptors).

• Learner modeling: maintains and updates adaptive profiles using declared preferences and interaction telemetry.

• Recommender system: generates recommendations using content-based retrieval (e.g., TF-IDF) and ranking strategies appropriate to the available signals, then filters or annotates outputs based on accessibility constraints and available indicators.

• Accessibility checker and remediation (decision support): performs selected WCAG-relevant checks and produces structured outputs (e.g., indicators and remediation hints) that can be surfaced to the chatbot to support informed discovery decisions.

The chatbot integrates GPT-4o-mini for intent routing and response planning, Whisper for speech-to-text, and OPUS-MT for multilingual translation, orchestrated via LangGraph/LangChain to support modular conversational workflows. Pilot sessions were conducted on a live prototype (not a simulated environment), and system logs captured latency, timeouts/failures, and fallback behavior to characterize feasibility under realistic network conditions.

Runtime workflow and orchestration

Figure 2 illustrates the state-aware routing workflow: language detection, optional STT for audio input, optional affect/engagement components if enabled, query classification, and dispatch to task-specific chains (recommendation, assistance, feedback, quiz or roadmap), with optional TTS for voice output. State transitions are logged to support traceability of routing and adaptation decisions.

Figure 2

Datasets and data preparation

Following CRISP-DM, the chatbot pipeline used multimodal, multilingual and accessibility-tagged resources supporting three functions: intent routing, assistance dialogue and MOOC discovery prompts (Table 4). Training and benchmarking resources were derived from publicly accessible sources (e.g., provider support materials and open multilingual benchmarks); no proprietary platform logs were used for training. Pilot interaction traces were collected under consent and used only for evaluation and telemetry-driven refinement.

Data preparation

A reproducible pipeline transformed raw inputs into standardized, accessibility-tagged records (Figure 3). The pipeline sequences (1) missing-field identification; (2) cross-referencing external public sources; (3) consolidating provider metadata; (4) normalizing fields; and (5) validating interaction-layer accessibility indicators. The following steps ensured data readiness:

• Normalization and harmonization: unify schemas across text, transcripts and feedback logs; standardize encodings, timestamps and language codes; resample audio to 16 kHz mono.

• Accessibility-aware enrichment: add structured indicators used by the interaction policy (e.g., readability proxies; response density/chunking controls; STT/TTS readiness; keyboard-first pathways).

• Validation and remediation hints: apply rule-based checks to interaction-layer metadata and selected WCAG-relevant indicators (e.g., subtitle/transcript flags; keyboard-path readiness) alongside UDL-informed readability/structure indicators. Missing or inconsistent attributes trigger structured remediation hints (e.g., request transcript; reduce density; provide keyboard-first guidance).

• Missing/uncertain data handling: impute accessibility-critical fields only when defensible; otherwise mark as unknown to avoid overstated assumptions.

• Deduplication and quality control: remove duplicates/corrupted entries; apply conservative imputations for non-accessibility fields while preserving accessibility tags.

Figure 3

This preparation ensures each record can be routed by intent and will activate the appropriate interaction supports (e.g., simplified responses, voice-first workflows, multilingual mediation) under consistent and auditable rules.

Pilot Study Design

Measures and Instruments

Modeling and evaluation

Runtime and Deployment Considerations

The prototype integrates third-party AI components (STT, NMT, and LLM-based routing/assistance). To assess feasibility under realistic conditions, pilot sessions were conducted on a live prototype and telemetry captured latency, timeouts/failures, and fallback behavior (e.g., retry, simplified response, or alternative modality suggestion). These measures informed scalability discussion by grounding performance claims in observed runtime behavior rather than assuming ideal network conditions.

Ethics, Privacy, and Data Governance

This feasibility pilot was conducted as a minimal-risk usability study. Participants were recruited on a voluntary basis and provided informed consent after receiving a plain-language summary of (i) what data would be collected (text queries, optional voice input, system telemetry and feedback), (ii) the purpose of the pilot (testing early-stage MOOC discovery and onboarding), (iii) storage and retention practices, and (iv) their right to withdraw at any time without penalty. Only the data necessary for assessing interaction feasibility were collected; no video, facial imagery or other biometric signals were recorded. Interaction logs were pseudonymized prior to analysis and stored on access-restricted servers. No formal institutional research ethics committee (REC/IRB) review was available or required for this type of minimal-risk pilot; therefore, the study adhered to prevailing research integrity and data-protection standards. Any sharing of derived materials (e.g. de-identified excerpts, aggregated telemetry or evaluation scripts) is restricted to datasets from which individuals cannot be re-identified and will occur only with participant consent and appropriate data-protection safeguards.

Data Availability

All third-party datasets and pre-trained models used for development and benchmarking (such as public MOOC support corpora and open multilingual resources) are publicly available. Study-generated interaction traces are not made public because they may contain sensitive or personally identifying content. Anonymized excerpts or aggregated summaries can be provided on reasonable request, subject to participant consent and compliance with data-protection requirements.

RQ1: reliability and runtime feasibility across text and voice

Using the GPT-4o-mini routing component integrated via LangChain/LangGraph, intent detection achieved 92.1 % agreement with expected scenario intents. Whisper (tiny) transcribed voice queries with a Word Error Rate (WER) of 9.2 % on the controlled speech set described in the Methodology. When transcription confidence was low, the system prompted users for clarification.

System logs recorded mean end-to-end response times of 823 ms for text queries and 1.24 s for voice queries (including STT). Timeouts and fallbacks (e.g., retry prompts, simplified responses, modality-switch suggestions) were logged as feasibility indicators. Together, routing accuracy, STT performance, and runtime latency indicate that the system can reliably and quickly support discovery/onboarding tasks across text and voice. Building on these modality-level findings, we next examine how well the chatbot mediates across languages.

RQ2: robustness of multilingual mediation

The ensemble language detector (LangDetect, LangID, Polyglot) achieved 95.82 % precision on the Zarajamshaid benchmark, enabling dependable language identification. Translation fidelity was evaluated with BERTScore on domain queries containing MOOC and accessibility terminology. OPUS-MT outperformed M2M-100 (F1 = 0.8962 vs. 0.8155); therefore, OPUS-MT was chosen as the default translation layer for the pilot’s main languages (Arabic, French, Urdu), with M2M-100 retained for fallback. Glossary constraints stabilized key accessibility and pedagogical terms (e.g., captions, screen reader, prerequisites). These results indicate that multilingual mediation preserves domain meaning for MOOC queries in the evaluated settings. Since effective translation alone is not enough to ensure equitable access, we next evaluate whether the chatbot can surface relevant accessibility information during course discovery.

RQ3: accessibility-aware course discovery using available signals

During discovery dialogues, the chatbot surfaced accessibility-related attributes when they existed in course metadata or structured outputs (e.g., language, caption/subtitle indicators, modality descriptors). When signals were absent, the system explicitly displayed “unknown/not specified” and asked clarifying questions (e.g., “Do you require captions?” or “Is keyboard-only navigation essential?”). WCAG 2.2 is treated as a design reference for the chatbot’s interaction flows rather than as a blanket conformance claim; thus, the interaction layer surfaces available signals and preserves uncertainty when attributes are missing. This behavior shows that the system can support accessibility-aware discovery at the interaction layer. To understand how users experience these features and to gauge pilot-level utility, we are now transitioning to user feedback and telemetry.

RQ4: pilot utility evidence from user feedback and telemetry

Post-session Likert scores averaged 4.4/5, indicating positive perceived usefulness and usability. Accessibility-specific reactions differed by profile: participants with visual impairments valued the voice/TTS interface and screen-reader-friendly navigation; those with cognitive processing challenges highlighted simplified outputs and reduced information density; multilingual participants emphasized clarity and stability of translated responses.

Open-ended responses were thematically analyzed as described in the Methodology. The resulting themes (friction points, clarity/cognitive load, modality usefulness, multilingual clarity, accessibility-signal usefulness) were coded with substantial inter-rater reliability (κ = 0.81).

Telemetry provided additional pilot-level insights: relative to an internal text-only baseline under the same tasks, session duration increased by about 25%, bookmarking by about 15%, and session completion among participants with disability-related needs by about 31%. These deltas are exploratory and should be interpreted cautiously given the small sample and short observation window, but they are consistent with reduced friction during early discovery and onboarding.

Taken together, self-reports, coded feedback, and interaction proxies provide early evidence of improved access-related task completion and engagement behavior across diverse learner profiles. No video, facial, or biometric data were collected; therefore, no affective-tracking outcomes are reported.

Scenario-based interaction cases

Scenario-based sessions illustrate how the system behaves across varied accessibility needs:

Scenario 1: Sighted learner (English text)

A neurotypical learner asked “I want to learn SQL”; the chatbot parsed the intent and delivered targeted course recommendations with low latency (Figure 4).

Figure 4

Scenario 2: Visually impaired learner (Arabic voice)

A blind learner interacted via Arabic speech; after STT processing, the chatbot returned concise Arabic responses, surfaced available caption/subtitle indicators, and offered a voice-first, keyboard-compatible flow. The user reported improved independent navigation compared to prior experiences (Figure 5).

Figure 5

Scenario 3: Learner with cognitive processing challenges (simplified UI)

• An adult learner with executive-function limitations used a simplified text interface; simplified language, reduced information density, and controlled pacing enabled successful search and bookmarking of courses without assistance (Figure 6).

Figure 6

Scenario 4: Multilingual learner with combined needs (Arabic voice + translation)

A learner with visual impairment and mild learning difficulties used Arabic voice input; language detection and glossary-constrained translation maintained stable terminology, while the interface adapted outputs to both visual and cognitive accessibility requirements. The learner felt understood and included (Figure 7).

Figure 7

Scenario 5: First-time MOOC learner (English onboarding)

A novice user asked how to create an account; the chatbot delivered clear, step-by-step instructions that enabled independent account creation (Figure 8).

Figure 8

These scenarios show that the chatbot can accommodate varied modality preferences and accessibility profiles through voice and text, multilingual mediation, simplified phrasing, and structured guidance, all within an inclusive interaction design.

Personalisation and adaptivity

During pilot sessions, the chatbot adapted outputs based on declared preferences and interaction signals (e.g., modality choice, reformulations, clarification loops). Observed behaviors included chunked summaries and simplified phrasing, modality-aware scaffolding, and accessibility-aware annotation of discovery results when signals were available. Offline comparison of retrieval configurations showed a 38% precision gain for accessibility-tagged queries relative to a TF-IDF-only baseline in offline evaluation, supporting the retrieval design choice.

Accessibility signal monitoring

The integrated accessibility checker generated course-level accessibility indicators and remediation hints from available metadata (e.g., caption/subtitle flags, modality tags, transcript/alt-text fields when present) and lightweight page parsing where feasible. In the pilot recommendation set, 87.5% of courses contained sufficient signals to be flagged as “indicators present”; this does not imply full WCAG conformance of the course experience. When signals were missing or ambiguous, courses were labelled “unknown/not specified”, deprioritized for accessibility-constrained queries, or flagged for follow-up.

Positioning against existing GenAI chatbots

Most existing educational chatbots (e.g., Khanmigo, Jill Watson) prioritize tutoring and general personalization, providing limited and inconsistent accessibility features (Caelen & Blete, 2024; Goel & Polepeddi, 2018). RECMOOC4All differs by integrating STT, TTS, NMT, simplified phrasing, and accessibility metadata into a unified interaction layer. This aggregator-level design may reduce dependence on individual instructors and improve consistency of accessibility signals across courses. To contextualize this distinction, Table 5 compares key interaction-layer features across several educational GenAI systems. The comparison emphasizes descriptive differences rather than formal conformance.

Architectural Scalability and Modularity

The RECMOOC4All prototype uses a modular architecture, built around LangGraph orchestration, LangChain prompts, ChromaDB semantic retrieval and SQL-Alchemy learner modelling, to support horizontal scalability and real-time personalization (Holmes et al., 2022). By decoupling components, it permits substitution of translation, STT or recommendation modules without interrupting service, which is essential for long-term evolution and potential cross-platform deployment. This modularity also enables continuous refinement: logged state transitions and feedback loops allow the system to adapt to emergent user needs.

As Schmitt notes, feedback-aware architectures enhance personalization by interpreting implicit user behavior, such as navigation loops and query reformulations. The present pilot validates the feasibility of this approach but does not yet test scalability across multiple providers (Schmitt, 2020); future work should examine how component substitution and feedback loops perform at larger scale and in diverse MOOC ecosystems.

Limitations

Several factors limit the generalizability of this pilot. The sample (n = 15) is small and heterogeneous, preventing robust statistical comparisons across subgroups. The absence of a randomized control condition precludes causal attribution of engagement gains. Language support is incomplete: OPUS-MT and M2M-100 cover Arabic, French and Urdu well, but low-resource languages and dialects are insufficiently supported. Speech-recognition accuracy declines in noisy or accented conditions, and the evaluation focused only on discovery and onboarding tasks. Future research should therefore avoid overgeneralizing these preliminary findings.

Future Directions

To enhance inclusivity, future work should expand language coverage and improve robustness for dialects and noisy inputs. Sign-language support could aid Deaf learners, as suggested by (Yarrow et al., 2023), while emotion-aware pacing may personalize interaction for diverse cognitive profiles. Larger, controlled, multi-site studies such as those advocated by (Tlili, 2024), are needed to evaluate long-term learning outcomes and equity impacts. Integrating domain-specific glossaries and paraphrasing models may assist learners with limited literacy or cognitive constraints. Finally, any expansion should be accompanied by transparent privacy safeguards and independent accessibility audits.

In sum, this pilot provides preliminary evidence that an accessibility-first, modular GenAI chatbot can enhance discovery and early navigation in MOOCs. The approach complements existing literature (Mrayhi et al., 2025b; Meyer et al., 2014; Tlili et al., 2025) by demonstrating a practical implementation of UDL principles within a Generative AI framework. However, its broader effectiveness and scalability remain to be validated through larger and more diverse evaluations.