Artificial Intelligence as Cognitive Scaffolding for Executive Functioning in Students with Special Educational Needs

Synthesis: Vasileiou, Mantsos, Krokou, Kofa and Polydoros (2026) conducted a critical narrative review asking a sharper question than the usual "does AI help students with SEN": what function does the AI actually perform? Their answer is a three-function framework — compensation (reducing or externalizing an executive demand so the learner can participate), scaffolding (supporting performance just beyond current independent capability), and development (building strategies the learner can later use without the tool). The review's central claim is that improved task completion while a tool is available is not evidence of executive-function development — the field's strongest AI evidence concerns learning outcomes from intelligent tutoring systems (Ma et al. 2014), not direct change in inhibition, working-memory updating, or flexibility. The framework is interpretive (no pooled effect sizes, no PRISMA flow diagram), and most AI-in-education studies measure achievement, engagement, or task completion rather than executive functions with validated instruments — so the authors call for withdrawal phases, transfer tasks, and direct EF measures before any developmental claims. Tool vintage: literature through March 2026; spans ITS, adaptive platforms, conversational agents, and generative AI. The pedagogical upshot is a teacher-mediated cycle: identify the barrier, define the goal, pick the least intrusive support, keep the learner in the decision loop, review process as well as performance, then maintain, adjust, or fade the assistance — with support faded or transferred before any "development" claim is made.

Summary

Executive functions — inhibition, working-memory updating, cognitive flexibility, and the planning and monitoring they support (Diamond 2013; Miyake & Friedman 2012) — are what classroom learning demands: following multistep instructions, beginning an assignment, revising an unsuccessful approach, checking completed work. Students with ADHD, autism, or specific learning disabilities may struggle in one or more executive domains, but the groups are heterogeneous and a difficulty completing work cannot automatically be attributed to an internal executive deficit — instructions may be ambiguous, materials may impose unnecessary cognitive load, or the classroom may lack cues and guided practice. The review brings together three literatures — executive functioning and self-regulated learning, inclusive education, and AI in education — because improved task performance while a tool is available may reflect external support rather than a change in the learner's skills. Self-regulated learning (Zimmerman 2008) overlaps with executive functioning but is not identical: it adds motivational beliefs and metacognitive knowledge. AI systems can intervene at several points in the regulation cycle — before a task (clarifying goals), during performance (prompts, feedback), and after (comparing outcomes against criteria) — and the review's contribution is distinguishing what each kind of intervention actually accomplishes.

The evidence base is useful but incomplete. Reviews show rapid growth in intelligent tutoring systems, adaptive platforms, automated feedback, learning analytics, and conversational agents, but much research targets higher education, technical performance, or general attainment rather than special education and executive-function outcomes (Zawacki-Richter et al. 2019; Zhai et al. 2021). A systematic review of AI in special education found potential benefits alongside substantial heterogeneity (Hopcan et al. 2023), and the strongest adjacent evidence concerns learning outcomes from intelligent tutoring systems rather than direct change in executive functioning (Ma et al. 2014). Evidence of better achievement or engagement therefore cannot be used as a proxy for gains in inhibition, working-memory updating, or flexibility — which is why the review treats many technology-to-process links as theoretically plausible mechanisms requiring direct testing, and positions its framework against Ouyang and Jiao's (2021) AI-directed / AI-supported / AI-empowered paradigms, which place learner agency at the center of system design.

Key Findings

• The same tool can serve three different functions, and the function — not the tool — determines what counts as success. The framework: compensation reduces or externalizes an executive demand (persistent reminders, automated sequencing; success = access and participation while support remains; risk = support mistaken for underlying improvement); scaffolding supports performance just beyond current independent capability (graduated hints, planning prompts, feedback questions; success = the learner makes decisions with targeted assistance; risk = excessive prompting reduces agency and productive struggle); development promotes independent strategy use and transfer (learner generates a plan, compares it with AI suggestions, reflects, repeats with less help; success = strategy use persisting after fading or transfer to a new task; risk = short-term performance reported as durable development). The framework is a progression, not a hierarchy — compensation is not an inferior outcome, since some learners require continuing supports to participate equitably — but problems arise when the purpose is unclear.
• The field's strongest AI evidence is about learning outcomes, not executive functions. The meta-analytic support for intelligent tutoring systems concerns learning relative to comparison conditions (Ma et al. 2014), and even when attention or planning appears to improve during tool use, three alternative explanations remain: the system may have reduced task difficulty, externally performed part of the executive work, or temporarily increased motivation. The authors conclude that "direct executive-function measures, withdrawal phases, transfer tasks, and longer follow-up periods are needed to distinguish these possibilities."
• Intelligent tutoring and adaptive systems support executive demands, they do not train executive functions. These systems structure problem solving — reducing the need to hold an entire procedure in working memory, directing attention to a relevant step, prompting error monitoring — and adaptive difficulty adjustment can support persistence and reduce frustration. But if a platform supplies every next step, the learner may complete more problems without developing an independent planning routine. The review's design prescription: graduated prompts, learner-selected strategies, and assistance that reduces as competence increases.
• Planning assistants are compensation or instruction depending on the objective. Breaking an assignment into subtasks, estimating time, and displaying progress directly address task initiation, prospective memory, sequencing, and time management — a legitimate accommodation for a learner who cannot begin work. But if the objective is independent planning, the system should progressively transfer responsibility by asking the learner to propose steps, justify priorities, and review time-estimate accuracy: "The teacher must decide which supports should remain available and which should be faded."
• Learning analytics are one input to professional judgment, never inference of cognitive state. Time-on-task, repeated errors, help-seeking, and irregular progress can locate where a learner needs support, and dashboards can support self-monitoring. But interaction traces are not measures of internal states: slow responding may reflect careful thought, language processing, distraction, or inaccessible design. "Algorithms should not infer an executive deficit or diagnostic status from behavioral data alone."
• Generative AI's affordances and risks are two sides of the same coin. Conversational and generative systems can produce outlines, explanations, contrasting representations, and alternative explanations that reduce barriers for students who struggle to organize ideas or transform intention into written action, and may support cognitive flexibility through contrasting explanations (Grassini 2023; Kasneci et al. 2023). But generated content may be inaccurate, culturally narrow, developmentally inappropriate, or inconsistent across prompts; a systematic scoping review found limited transparency, replicability, privacy protection, and stakeholder involvement across educational LLM applications (Yan et al. 2024). Ready-made plans and prose may replace the very planning and monitoring processes education seeks to develop.
• Accessibility support must not be confused with remediation. Text-to-speech, adaptive text presentation, and structured prompts can improve access for learners with reading or writing difficulties, with effects varying across learners and implementations (Wood et al. 2018). Whether a tool is used as an accommodation, as part of explicit strategy instruction, or as temporary practice support "should be stated in the learning plan."
• Cognitive offloading is not inherently harmful; the question is calibration. Notes, calendars, and calculators are ordinary tools, and for learners with disabilities external support may be essential for equitable participation. The question is whether the offloaded process is necessary for the intended outcome: if the goal is to construct an argument, generating the structure removes planning practice; if the goal is to demonstrate scientific understanding, reducing transcription load may enable a more valid demonstration. Support "should be calibrated rather than categorically accepted or prohibited."

What this means for practice

• Start with the barrier and the objective, and name the function. The first two of six educational principles: "begin with a clearly defined learner barrier and educational objective rather than with an available tool," then specify whether the intended function is compensation, scaffolding, or development — each has a different success indicator (access and participation vs. decisions-with-support vs. strategy use after fading), and a tool should be evaluated against the function it declared: developmental claims require demonstrated retention, transfer, or independent performance.
• Preserve learner agency through design, not policy. The third principle: require meaningful choices, verification, and reflection. Useful designs preserve decision points, ask learners to justify choices, and incorporate deliberate fading — the direction Molenaar (2022) describes as hybrid human-AI learning, where human and artificial capacities are deliberately combined rather than treated as substitutes. Such designs also let educators inspect the support provided and override it when it conflicts with instructional aims.
• Run the teacher-mediated cycle. The review's implementation procedure: (1) educator and learner identify the functional barrier and define an observable goal; (2) select the least intrusive support likely to help; (3) the system provides support while the learner remains actively involved in decisions; (4) review both performance and the process used; (5) maintain, adjust, or fade assistance according to the original purpose. The cycle "places educational judgment, learner voice, and ongoing assessment above technological novelty."
• The teacher's role expands rather than shrinks. AI cannot determine the educational meaning of a student's behavior without contextual knowledge — teachers understand classroom relationships, curriculum aims, prior instruction, and the task's place in a longer learning trajectory. Their role "is therefore not reduced to supervising technology. It expands to include selecting appropriate functions, teaching critical use, interpreting data, and protecting learner agency." Students should also participate in decisions about support, "especially when systems collect behavioral data or make recommendations about them" — learner voice reveals whether prompts are distracting, whether feedback is understandable, and whether a tool increases confidence or dependence.
• Apply data minimization and contestability. Institutions should minimize data collection, define retention and access, evaluate vendors, and communicate with learners and families; data used for one educational purpose should not silently be repurposed for diagnosis, discipline, or high-stakes prediction. Algorithmic bias can emerge from unrepresentative training data or outcome definitions that treat conformity as success — a system may read atypical communication, longer response times, or frequent help-seeking as low ability rather than as a learner–environment mismatch. Human review is necessary but not sufficient; institutions also need procedures for contesting automated outputs.

Limitations

• Narrative, not systematic, review. The critical narrative design does not permit an estimate of pooled effectiveness and is more susceptible to selection and interpretive bias than a systematic review; the paper reports no PRISMA flow diagram and does not claim systematic-review completeness.
• The underlying evidence base measures the wrong outcomes. Most educational AI studies measure achievement, engagement, usability, or task completion rather than executive functions with validated instruments, so "conclusions about executive-function development must therefore remain cautious."
• The three-function framework is interpretive, not validated. The categories were produced by comparative reading and concept mapping; they are "proposed as a framework for research and practice rather than as the result of a quantitative effectiveness analysis."
• Terminology varies across disciplines and the evidence is developing rapidly, which limits comparability across the synthesized literature.
• No primary data. The review synthesizes published literature only; the framework's predictions (e.g., that fading or transfer distinguishes development from compensation) await direct empirical test.

Research priorities

The review calls for: precise reporting of the technology's function, the support provided, and the target executive process; validated executive-function or self-regulation measures alongside academic outcomes; comparison conditions that distinguish AI from equivalent non-AI scaffolding (since benefits may arise from feedback or structure rather than AI itself); longitudinal and single-case experimental designs for maintenance and individual variability; transfer to unfamiliar tasks; learner agency, teacher workload, and unintended effects; participant contributions to design and evaluation "rather than appear only as recipients of a finished system"; and transparent reporting of privacy arrangements, model limitations, and human oversight.

Citation

Vasileiou, I., Mantsos, E., Krokou, Z., Kofa, O., & Polydoros, G. (2026). Artificial Intelligence as Cognitive Scaffolding for Executive Functioning in Students with Special Educational Needs. Journal of Educational Technology and E-Learning Innovations, 2(2), 40–46.