Scaffolding

Scaffolding — structured support that helps learners accomplish tasks they cannot yet complete independently, with support fading as competence grows. In AI in Education, scaffolding is the primary design principle for ensuring AI tools support learning rather than replace it.

Questions to Consider

• Think of a time someone 'helped' you with something you were learning, and the help did the work so well you learned less. Where's the line between support that lets you grow and support that replaces you? The page argues this is the central design question for AI tutors.
• Scaffolding is rooted in the Zone of Proximal Development — enough support to enable progress, not so much that learning is bypassed. What does 'too much support' look like in an AI tutor, and can you detect it from the student's behavior alone?
• A key finding: students often prefer the more directive AI tutor roles even though they perform better with collaborative peer and Teaching-assistant roles. Does learner preference reliably track what's best for learning — and what does this divergence imply for letting students choose their own scaffolding?
• Scaffolding that never fades creates dependency. The page notes automated scaffolds risk staying static instead of being withdrawn as competence grows. Why is 'fading' essential, and why might an AI system fail to do it if it isn't deliberately designed to?
• The design principle is 'scaffold, do not substitute.' Students themselves asked for AI that 'does not provide any solutions for you — you still learn as you have to find the correct answer yourself.' Does that match how you've experienced effective help, or have you preferred the shortcut even while knowing it cost you?
• Set a goal before reading: pick a task you teach, and sketch what a hint looks like that preserves the learner's effort versus an answer that removes it. How will you know your hints are in the productive-struggle zone?

Introduction

How scaffolding appears in AIED

• Prompt-based scaffolding: Guided LLM scaffolding teaches structured prompting as a learning intervention. Critical engagement scaffolding uses culturally responsive approaches.

• Mixed-initiative scaffolding for gaps the learner cannot name: Sun et al. (2026) keep the writer in charge of narrative synthesis while the system takes initiative only over domain discovery, proactively flagging knowledge gaps writers cannot formulate as queries.

• Socratic scaffolding: Socratic AI dialogue withholds direct answers, using questions to guide discovery — a form of Desirable Difficulties scaffolding. As a boundary condition, questioning helps at onset but decays with depth, so the fade rule shifts from eliciting to addressing the error: the benefit falls from a 20-point advantage at onset to 3 points by the fourth stuck turn, while addressing the error rises from +5 to +12 Ahtisham et al. (2026).

• Adaptive fading: Intelligent tutoring systems adjust scaffolding based on Knowledge Tracing estimates, providing more support for unmastered concepts and less for known ones.

• The function, not the tool, decides what counts as success: Vasileiou et al. (2026) propose a three-function framework for AI support in special education. Compensation reduces an executive demand so the learner can participate (success = access while support remains). Scaffolding supports performance just beyond current capability (success = decisions with targeted assistance). Development builds strategies usable without the tool (success = strategy use after fading or transfer). Each function carries a different evaluation standard, and the field's recurring error is reporting task completion during support as development. It is a narrative review with interpretive categories, not a pooled effect estimate.

• Need-triggered adaptive scaffolding in clinical interview training: the MeduAI-SP randomized trial (Yang et al., 2026; N = 100 third-year medical students) operationalized scaffolding as adaptive, need-triggered support: a turn-level evaluator agent flagged when a learner was stuck, omitted key history, risked premature closure, or impaired rapport, and only then did a tutor agent issue a Socratic prompt. Expert annotation of 207 consultations showed that scaffolding need was strongly phase-dependent, rising from about 16.1% of student utterances early in the encounter to 34.8% late (24.1% flagged overall) — empirical evidence that novice consultation support is most needed during integration and diagnostic reasoning, not initial information gathering. The scaffolding condition outperformed a structured progressive-disclosure control on the final OSCE-aligned examination (71.8% vs. 55.6%; β = 16.4 percentage points; P = 3.30e-4), with the largest gain in communication (Hedges' g = −0.79).

• Hint systems: AI tutor hint research examines when hints help versus when they encourage Over-Reliance.

• Scaffolds that raise post-mistake recovery. In a 6,000-student middle-school field experiment, AI support slowed progress and cut questions attempted but increased next-attempt correctness after errors and reduced the attempts needed to return to a correct answer — a "productive slowdown" rather than answer supply (Oreopoulos et al. (2026)).

• Authentic teacher feedback is mostly scaffolding. Lee et al. (2025) built a ten-type taxonomy from 100 teacher-authored feedback messages on open-ended middle school math responses and found most clusters were facilitative moves — Hinting, Rhetorical Question, Show Your Work, and Worked Example — rather than bare correctness feedback. The factor structure sharpened the contrast: Factor 7 "Suggestive Hinting" loaded positively on Mathematical Suggestions (N=4) and Hint (N=3) but negatively on Explanation (N=2), marking hints as indirect support that scaffolds understanding without providing the near-complete solution.

• Training the learner is itself a scaffold. A 45-minute session teaching students to prompt a model by SRL phase, and to remediate its visual and reasoning errors, raised physics revision accuracy (d = 0.81) without increasing reported total cognitive load (Huang et al. (2026)).

• Conceptual and representational scaffolds: Concept Catalyst and LLM tutor rethinking explore design patterns for cognitive support. Davor (2026) supplies a domain where representational support and answer-giving have to be pulled apart: geometry proof is rich in Visualization resources, yet the step from a relationship that can be seen to a valid deductive argument is the one learners fail to make, so the scaffold has to carry them across representations rather than hand them another picture. His quasi-experiment with 86 Ghanaian senior high school students configured generative AI to prompt and question rather than solve — asking learners to justify what they observed, name the relevant theorems, and order their logical steps, with the teacher using the AI's representations to steer discussion — and the AI-supported class outscored conventional instruction on proof construction even after controlling for pre-test performance. Perceptions of the feedback and explanation quality were favorable, though only the experimental group was surveyed. The scaffold is positioned here as the bridge from visual intuition to deductive reasoning, and the paper never names the model or a collection window, so it offers more of a design warrant than a replication recipe.

• "Scaffold, do not substitute" as a design principle: Favero et al. (2026) argue that the central risk of AI in education is misalignment — AI that substitutes for human effort erodes the capacities education is meant to build — and derive a single design principle, scaffold, do not substitute. Scaffolding must be a first-class capability of AI systems: knowing when to withhold an answer, ask a question, surface uncertainty, or present alternative perspectives. Their analysis of student essays shows learners themselves converge on this — asking for AI that "does not provide any solutions for you, you still learn as you have to find the correct answer yourself." The principle positions scaffolding as the alternative to a self-reinforcing harm cycle of substitution across cognition, Learner Agency, emotion, and Ethics.

• Mentoring as scaffolding for a research process. Wang et al. (2026) treat a systematic review as an intentional learning experience in which mentoring supplies integrated methodological and emotional scaffolding, while automation is confined to procedural load such as abstract screening, and data extraction, reconciliation and synthesis stay with human reviewers.

• Scaffolding embedded in the medium, not bolted on: Wang, Du and Jin (2026) operationalize four scaffolding principles inside a video player — ZPD-based adjustment of comment depth, fading scaffolding (knowledge support thins over the timeline), distributed scaffolding (every comment is either knowledge or emotional support), and cognitive-load-derived timing — by computing frame-level video entropy and inserting comments only in low-information intervals. Their four-condition ablation with 20 learners found the entropy-timing module produced the strongest and most robust effect on perceived quality when removed (Z = −2.85, r = 0.45, p = .004), which makes the scheduling of scaffolding a measurable design variable rather than a packaging detail. The study also carries a warning for automated scaffolding: ChatGPT-generated comments were consistently harder to read, less lexically diverse, and less topically aligned than instructor comments, with the largest relevance gap in emotional support (feedback quality).

• Emotional scaffolding as a prompt layer, measured against cognitive scaffolding alone. Holding the interface and lesson fixed, the emotional version asked students to report feelings at three points and used those reports to vary tone, affective scaffolding and cognitive challenge along control and value appraisals, with temperature at 0.7 so the adaptations guided rather than dictated replies. In the dialogue traces it bought more performance-phase and self-reflection work than SRL prompts alone, and more enjoyment and less boredom. It added nothing to learning gains or self-efficacy in a single 25-minute session (Diordiev, Stenberdt & Makransky, 2026).

• Preferred scaffolding is not always the most effective: Zhu, Yang and Yang (2026) found in a within-subjects experiment that students performed best with Peer and Teaching Assistant AI roles (which foster collaborative reasoning) yet preferred the more directive Tutor and Excellent Student roles — a divergence between preference and performance that cautions against equating learner preference with effective scaffolding in AI-supported mathematical modeling.

• A more structured scaffold is not automatically a more agentic one: an expert design bot raised interaction length and self-authored prompts relative to plain ChatGPT, but collaborative Problem Solving agency fell in both conditions, with agreement instead of critique the lowest-agency behavior observed (Krushinskaia et al. (2026)).

• Automated scoring as a deliberate scaffold: Chen and Liu (2026) treated an automated interpreter-scoring system as formative scaffolding rather than as a measurement instrument, in a 14-week comparison with 46 English Translation and Interpreting sophomores. The scaffold converted into a gain only where the deficit it targeted was decomposable, reliably scored and sensitively scaled: linguistic accuracy and logical coherence rose, while information fidelity and delivery fluency stayed flat, fidelity being the dimension on which automated and human ratings disagreed most (r = 0.12). The cycle it reinforced was also partial. Only practice-phase execution and monitoring correlated with score gains (r = 0.42), while pre-learning planning sat near the scale midpoint (M = 3.01) and students set goals from the previous score rather than from the task ahead. A scaffold can be well placed inside the performance phase and still leave the planning that would make it unnecessary untouched.

• Closed-loop scaffolding on the learner's current boundary: Zhu, Luo and Li (2026) built a music-education loop in which Multimodal AI error detection over performance audio and score becomes the reward that steers reinforcement-generated practice tracks, so difficulty follows the learner's current boundary instead of a fixed syllabus: generated practice trajectories matched learner skill profiles at a peak cosine similarity of 0.962, and the Group x Time interaction favored the scaffolded group in a 12-week quasi-experiment with 120 undergraduates (beta = 0.52, 95% CI [0.31, 0.73]). Two limits follow for automated scaffolding. Detection is triage rather than judgment, since rhythm error precision of 89.7% means roughly one flagged error in ten is a false alarm, and blind expert review rated support for musical expression the system's weakest area, leaving interpretation to the teacher.

• Pacing is scaffolded too. Englmeier (2025) pairs content scaffolding with a temporal layer of weekly time allowances, progress dashboards and end-of-section checks; without pacing, nearly 25 percent of over 600 learners began in the final week and 95 percent of them failed.

The ZPD connection

Vygotsky's Zone of Proximal Development provides the theoretical foundation: scaffolding targets the space between what learners can do independently and what they can achieve with support. AI tools should operate in this zone — enough support to enable progress, not so much that learning is bypassed. A configuration that inverts the usual direction of adaptation appears in Sidorkin's (2026) graduate course, where the learner rather than the system set the support level: weekly readings were generated on demand and students dialed comprehension level through iterative prompting (pacing, definitions, vocabulary density, depth). Analysis of three reading logs found definitional markers 3.4x to 8.7x more frequent in AI responses following comprehension-oriented prompts than in baseline explanatory text, with the clearest cases building a definitional layer and then a numbered procedural one, which makes scaffold density a measurable property of the learner's request rather than only of a system's mastery estimate. Requiring at least three follow-up questions per reading made that dialing routine, turning the text into an interaction that surfaced comprehension gaps the instructor otherwise would not see.

Scaffold form has to match what the learner can already hold, not only what the task requires. In a systematic scoping review of 24 evidence sources on creative thinking in children aged 6 to 15, Niu et al. (2026) report that younger children lacked the precise linguistic and metacognitive control that text-based prompting demands and needed multimodal, adult-facilitated interfaces, while text-based LLMs showed stronger reported outcomes with older children and young adolescents. Prompt dependence was strongest in the lower grades, and no included study examined developmental readiness thresholds, so the authors treat modality and scaffolding choices as an open design question to be settled by the learner's capacity rather than by convenience.

Connections

Scaffolding connects to Over-Reliance (scaffolding that doesn't fade creates dependency), Cognitive Load Theory (scaffolding manages cognitive load), Feedback Loop (scaffolding provides formative feedback), and AI Literacy (learners must recognize when scaffolding is beneficial vs. when it displaces learning).

Agents must scaffold dynamically, not statically: Woollaston et al. (2026) identify that automated scaffolds risk staying static instead of being withdrawn as competence grows, and recommend dynamic scaffolds that adapt and fade — a key guardrail for Agentic AI. CoMeT (Hou et al. 2026) supplies both the separation and the warrant for when to fade: its support climbs one rung each time a learner does not use it and drops to the lightest rung on take-up, holding metacognitive demand statistically equivalent to a withholding tutor (p_TOST = .004) while delivering an artifact in 48.1% of sessions against 23.7% — more system labor, not less. The trigger it validated is aim rather than depth: after a full demonstration the tutor later conceded 30.3% of what was still open, after a pasted artifact 25.9%, after a bare assertion 20.4% and after a request to build 33.7%, whereas turns not aimed at the decision under support drew a later concession 40.3% of the time against 28.8% for aimed turns (11.5-point difference, 95% bootstrap interval [2.1, 22.4]). Take-up was sparse — 37.0% after the first ask and 21.8% after the third — so a fade rule keyed to learner effort would read a non-answer as readiness.

Fading can be trained out of a tutor. A multi-turn RL tutor rewarded only on the current problem and a masked near-transfer variant stopped fading support and assigning independent work, though nothing penalized either, because their payoff lies beyond one dialogue episode (Macina, Kapur & Sachan, 2026). A reward horizon that ends with the problem will erase the fading a design specifies.

A fourth zone makes the source of the scaffold visible: Tsim and Gutoreva (2025) add "known to GenAI" to Vygotsky's zones, so a task falls into Substitute, Complement, Aid or Non-negotiable — and Non-negotiable holds what needs a person because it is normative, tacit or too risky to hand over.

• Place AI by phase of the idea, not by course. The six-move placement frame holds AI out of the first hard attempt and final unaided check while licensing it between them for hints, examples and drill; if AI makes the task feel effortless, it is in the wrong place (Brcic & Frljic, 2026).

Scaffolding must be situation-appropriate, not maximal: Zhang et al. (2026) introduce TutorMoments, which evaluates whether LM tutors scaffold only when support is needed, push for rigor when the student is ready, and avoid over-scaffolding (reducing cognitive demand more than the situation requires). Minimally prompted frontier models default to over-scaffolding at the expense of productive struggle.

• AI that scaffolds productive struggle. Kim et al. (2026) derive AI design principles (non-directive support, reflective design, Human-in-the-Loop) that keep scaffolding in the productive-struggle zone rather than collapsing to answer-giving; Puech et al. (2025) show Large Language Models (LLMs) tutors can be steered to give help only when strictly necessary — scaffolding that preserves the learner's own effort.

Scaffold withdrawal as the enforcement mechanism for verification. Kumar, Wongsirichot and Nanthaamornphong (2026) synthesize 72 computing-education studies and locate scaffold withdrawal, alongside guardrail tools and Self-Regulated Learning designs, as one of three ways courses enforce critical engagement with AI output — structurally (constraining what the tool returns), procedurally (reflection logs, self-testing) and temporally (progressively restoring conditions under which independent reasoning is required). Their evidence is that efficiency gains under AI assistance do not transfer to unaided performance, and that the failure mode — the pseudo-apprenticeship pattern, where students watch AI generate code without performing the task — is exactly modeling without whole-task practice. Graduated access therefore functions as a fading schedule for a powerful new form of support, and the review grounds it in 4C/ID: assistance helps only when the learner already has enough schema to engage critically with it (the zone of proximal development, Cognitive Offloading).

Sequencing is the scaffold's other lever: Vendrell and Johnston (2026) recommend requiring a learner's independent attempt before any AI consultation and then using the model to generate counterarguments rather than answers, with deliberate AI-free phases, so the support amplifies reasoning instead of replacing it.

Rule-Guided vs. Ad-Hoc Scaffolding

• Rule-guided vs. ad-hoc scaffolding. Looi, Liu, and Sun (2026) formalize a distinction central to scaffolding design: rule-guided scaffolding, in which tutoring is governed by an auditable three-layer architecture (diagnosis → intent selection → constrained response generation), versus ad-hoc scaffolding, where helpful moves are difficult to audit and replicate. Their primary-school math study showed rule-guided scaffolding improves interactional consistency, reduces premature answer-giving and early closure, and sustains cognitive engagement — evidence that explicitness and auditability of scaffold moves matter for both consistency and learning in procedural domains.
• Scaffolding as the constrained pathway between bypass and offloading. The Neuroplasticity-AI Interaction Model names scaffolding as the third of three pathways for LLM help, alongside direct bypass and cognitive offloading, and defines it by whether the model preserves the effortful processing the task is meant to train (Bypass, Offload, or Scaffold: A Conceptual Model of How Large Language Models Shape Learning). The model's calibration evidence is a natural experiment in scaffold constraint: unrestricted GPT-4 access in a study of nearly 1,000 high-school mathematics students produced a 48% practice gain but a 17% deficit on the unassisted exam, while hint-constrained GPT Tutor produced a 127% practice gain with the exam deficit largely eliminated. The design lesson matches the rule-guided versus ad-hoc distinction above at a coarser grain: it is the constraint on what the tutor is allowed to supply, not the presence of a tutor, that determines whether the scaffold is removed successfully.

Timing can be the rule. Rotter and colleagues trained a reinforcement-learning agent to decide, question by question, whether a student could open an unrestricted LLM at all, and tested it against always-on and never-on access with 105 university students. Timed access scored 5.11 on the unassisted post-test against 4.03 for always-on (t(71) = 2.51, p = .014, d = 0.59) and matched full restriction at 5.25, while self-assessment accuracy rose to 5.95 against 5.17 (p = .023). The mechanism was rationing rather than coaching: students in the timed condition made 2.46 requests on average against 11.25, 43.2% never used the model at all, and declared metacognitive awareness did not move. The design also exposed a loophole worth designing against — one participant considered answering wrong on purpose to unlock the AI sooner (Rotter et al., 2026).

The rule-guided versus ad-hoc distinction reaches past the tutor and into the assignment itself. A cross-institutional faculty collaboratory found that critique of AI output did not happen on its own, so requirements to audit, compare, revise and justify generated content had to be written into the task, and those who protected independent disciplinary analysis before AI entered found candidates could evaluate AI output more critically. Rehearsal evidence agrees: pre-service teachers practicing with AI partners used more probing and exploring questions when structured post-rehearsal feedback was provided. In both cases support is specified in advance rather than improvised, separating designed guidance from the ad-hoc help that is hard to audit and replicate.

• A correctly constrained scaffold still fails if the constraint is not administered. Azimi (2026) built what the distinction above prescribes — a budget of 25 hints per session, a 15-minute cap on AI use, a required end-of-session reflection, and no generated code — and randomized 33 master's students between it and unrestricted Generative AI use across seven weeks. Assignment performance and concept-inventory gains did not differ between conditions. The hint budget did not act as a rationing mechanism: some students spent most of it on the first problems and had none left for the demanding ones, others finished with most unused, and within the Coach condition it was the students who already had a deliberate strategy for spending a hint who scored higher. The constraint raised reported confidence and followed students out of the classroom as a self-questioning habit (whether a question was worth asking the tool), but it rewarded existing self-governance rather than developing it. Auditability of a scaffold and a learner's capacity to use it are separate conditions.

• A scaffold the learner cannot verify is a fluent substitute. Because chemistry reasons across observable phenomena, particulate models and symbolic notation, a generated answer can be locally persuasive and globally wrong. Vega-Baudrit and Rivera Álvarez (2026) place scaffolding on the productive side of their Presage-Process-Product analysis and uncritical copying on the failure side, and require scaffolds to demand representational translation in both directions, since a response that describes neutralization correctly can still claim that every equivalence point has a pH of 7. Verification is designed into the task rather than announced as a rule: students identify a false assumption, correct a unit or mechanism error, compare a symbolic structure with a submicroscopic model, or justify rejecting a generated answer. Because students cannot verify what they do not yet understand, prior knowledge and scaffolding come first, and prompting is treated as an epistemic act in which the learner specifies the constraints the answer must satisfy.

The scaffold can sit in the assignment rather than in the tool. In a 43-student mechanics section, GPT-5 exchanges were mostly recipe requests during discussion sessions, then shifted on a graded end-of-term assignment that required at least three iterative exchanges and warned against pasting problem statements. Recipe-seeking fell from 52% to 22% of coded phrases, and think-aloud requests about physical meaning rose from 0% to 25% (Stone, Zhang & Utter, 2026). The model did not change between the two settings; the specification students worked to did.

Common Misconceptions

• "Scaffolding is a constant level of support." Expert annotation of 207 consultations showed scaffolding need was strongly phase-dependent, rising from about 16.1% of student utterances early in the encounter. A single scaffold level is misallocated at both ends of the same consultation.
• "Any structured control condition is a fair comparison for scaffolding." On the examination, the scaffolded arm still beat a genuine progressive-disclosure alternative (71.8% versus 55.6%), so the gain was measured against a real pedagogical alternative rather than against nothing.
• "Scaffolding effects are general." The gain was 16.4 percentage points (P = 3.30e-4), with the largest improvement in communication (Hedges' g = -0.79). The effect concentrated in one outcome, so an average across outcomes understates where it works and overstates where it does not.

Connected Concepts

• Pedagogical Patterns — Where in the sequence support belongs, and why attempt-before-help is the load-bearing rule
• Problem-Based Learning — PBL embeds fading scaffolds around ill-structured problems
• Learning by Teaching — Scaffolding knowledge building through explanation
• Sociocultural Learning — Vygotskian foundation: ZPD and socially mediated learning
• Cognitive Offloading — Scaffolding that never fades creates over-reliance
• Feedback — Scaffolding delivers formative feedback as support fades
• AI Literacy — Recognizing when scaffolding supports vs. displaces learning
• Intelligent Tutoring — ITSs adapt scaffold intensity to mastery estimates
• Socratic Method — Questioning that withholds direct answers
• Metacognition — Scaffolds that build self-monitoring and self-regulation
• Adaptive Learning — Adaptive systems modulate support within the learner's ZPD
• Learning Design — Scaffolding is a core instructional-design strategy
• Help-Seeking — Scaffolding shapes when and how learners request help
• Teaching — Teachers scaffold, then fade as competence grows
• Pedagogies and Teaching Strategies — Umbrella: pedagogies and teaching strategies in AI education
• Agentic AI
• Productive Failure — Productive Failure

Connected Articles

• AI is ready, but are the students. From answer-making to sense-making — a graded rubric, not a better model, moved student dialogue from recipe-seeking to think-aloud (Stone, Zhang & Utter, 2026)
• Generative AI in computing education: A systematic review and a framework for responsible integration — Scaffold withdrawal as the mechanism enforcing verification (VIE framework)
• Evaluating Scaffolding-Oriented Multi-Agent Large Language Model System for Clinical Interview Training — Evaluating Scaffolding-Oriented Multi-Agent Large Language Model System for Clinical Interview Training
• Assessing ChatGPT-Generated Comments for Video-Based Learning Content to Enhance Knowledge and Emotional Support Based on Scaffolding Theory — Entropy-timed AI comments as in-video knowledge and emotional scaffolding (Wang, Du & Jin 2026)
• Beyond Access: Guided LLM Scaffolding for Independent Learning in Undergraduate Statistics — Guided LLM prompting as a structured learning intervention
• Scaffolding critical engagement with GenAI: Transforming ethnic minority preparatory students' collaborative discourse in prompt engineering tasks — Culturally responsive critical-engagement scaffolding with GenAI
• Rethinking Scaffolding in LLM Tutors: The Interactional Mismatch Between Benchmarks and Real-World Deployments — Design patterns for scaffolding in LLM tutors
• Creating Learning Scaffolds for Engineering Design Using Concept Catalyst — Concept Catalyst scaffolds for conceptual change
• Catching The Correct Answer Trap: Characterising AI Tutor Blind Spots When Analysing Student Reasoning — When hints help vs. when they encourage over-reliance
• Scaffolding critical thinking with generative AI: Design principles for integrating large language models in higher education — Scaffolding critical thinking with GenAI
• VeriForge: Mitigating Latent Knowledge Gaps in Narrative Drafting via Mixed-Initiative Scaffolding — Scaffolded narrative drafting with Veriforge
• From Substitution to Scaffolding: Breaking the Self-Reinforcing Harm Cycle of AI in Education (and Beyond) — From Substitution to Scaffolding: Breaking the Self-Reinforcing Harm Cycle
• Preferred scaffolding does not lead to better learning performance: Empirical evidence from AI-supported mathematical modelling — Preferred scaffolding in AI-supported mathematical modeling
• Agentic AI and Pedagogical Best Practice: The Tension Between Automation and Learning — Dynamic (fading) scaffolds as a guardrail for agentic AI
• When Help is Unhelpful: Evaluating AI Tutors for Productive Struggle — When Help is Unhelpful: evaluating AI tutors for productive struggle
• Designing AI systems to support a productive-failure-based learning: insights from adult learners on AI applications and AI system design principles — Designing AI Systems to Support Productive-Failure-Based Learning
• Towards the Pedagogical Steering of Large Language Models for Tutoring: A Case Study with Modeling Productive Failure — Pedagogical Steering of LLMs for Productive Failure
• Making AI tutoring productive: Evidence from a mastery-based math practice experiment — Making AI tutoring productive: mastery-based math practice
• The effortless trap: Productive struggle, AI, and the illusion of learning — Guarded vs. unguarded AI: the placement rule (Brcic & Frljic 2026)
• Supporting undergraduate students' learning in practical chemistry courses through AI-supported experimental design — AI-supported experimental design in practical chemistry
• Scaffolding systematic reviews in learning design and technology through mentoring and AI integration — Scaffolding Systematic Reviews with Mentoring and AI (Wang 2026)
• Pre-service teachers' agency during their interactions with generative AI while designing for learning – a process view on Intelligent-TPACK — Pre-service teacher agency during GenAI interactions in design for learning (Krushinskaia, Elen & Raes 2026)
• Responsible integration of artificial intelligence into pedagogies of practice in mathematics teacher education — AI across the pedagogies of practice in mathematics teacher education: structured rehearsal feedback raised probing questions (Bondurant & Shaughnessy 2026)
• AI integration as instructional design: Lessons from a cross-institutional faculty collaboratory in teacher preparation — AI integration as instructional design: critique of AI output must be assigned, and disciplinary analysis comes first
• AI literacy over tool design: a mixed-methods study of scaffolded versus unrestricted generative AI in programming education — A hint-budgeted AI Study Coach: scaffolded vs unrestricted GenAI use, and why the constraint alone did not produce learning (Azimi 2026)
• SCAN: A Decision-Making Framework for Task Assignment with Generative AI — SCAN: scaffolding reframed as deciding which sub-zone a task belongs to
• Impact of automated scoring on interpreting performance and self-regulated learning: evidence from a pedagogical experiment — Automated scoring as formative scaffolding: it converts into a gain only where the targeted deficit is decomposable, reliable and sensitively scaled (Chen & Liu 2026)
• Adaptive teaching assistance model combining generative AI and big data analytics — Closed-loop adaptive scaffolding that tracks the learner's current boundary in music practice, with error detection as triage (Zhu et al. 2026)
• The role of generative AI in facilitating children’s creative thinking and cognitive development: a systematic scoping review — Scaffold form must match children's developmental capacity, not only the task (Niu et al. 2026)
• Generative artificial intelligence in university chemistry education: a critical review using Johnstone’s chemistry triplet and Biggs’ 3P model — Scaffolds must require representational translation and verification, since students cannot verify what they do not understand (Vega-Baudrit & Rivera Álvarez 2026)
• Generative AI-supported instruction as a cognitive scaffold: Effects on senior high school students' geometric reasoning and proof construction — GenAI as a prompting scaffold that bridges visual intuition to deductive proof in geometry, where visualization alone is insufficient (Davor 2026)
• Examining Variation in How Guided AI Tutors Resolve Student Impasses — Examining Variation in How Guided AI Tutors Resolve Student Impasses
• From answer machine to collaborative partner: impact of structured generative AI training on scientific problem solving in STEM education — Training the learner as scaffolding for GenAI use in STEM problem solving (Huang et al. 2026)
• A Conversational AI Framework for Teacher Training — Conversational AI diagnostic for teaching misconceptions with evidence-grounded dialogue (Miranda & Calvani 2026)
• Scaffolding Autonomy: AI-Augmented Self-Paced Learning Environments for Sustainable Learning Outcomes — Content and temporal scaffolding layers in a self-paced course, with late-start and failure rates by study timing
• The Assistance Dilemma - Learning to Teach via Multi-Turn Reinforcement Learning — The assistance dilemma: training a multi-turn RL tutor whose reward horizon ends with the problem trains fading out (Macina, Kapur & Sachan 2026)