Scaffolding

Principle · All principles
Evidence · 76 claims (23 for, 44 mixed, 9 against) · 54 studies (27 causal, 11 review, 6 quant-synthesis, 4 theoretical, 3 design, 2 associational, 1 qualitative), q1–q4 · 18 of 54 report an effect size · 69 claims rest on one study

Conditional relationship

A learner who cannot yet complete a task alone may learn more from attempting it with support than from attempting it unsupported. Support here means hints, prompts, models, partial structures or tutor questions given during the attempt. The expected change is in later cognitive outcomes on the target task (immediate-performance, near-transfer, conceptual-understanding), not in how well the learner does while the support is present. The relationship is bounded by the learner's starting response on this task (novice for the task, not for the domain), by the kind of support and what it leaves the learner to do, by the setting (most evidence is one-to-one tutoring or computer-based support in STEM problem solving; little is classroom, where the wiki holds one whole-class trial with third graders, a few small group-work studies and two studies of teachers scaffolding their own classes), and by whether the outcome is measured with the support removed.

Two randomised studies in generative-AI settings now test support against its absence directly, and neither settles it. Among novices using an LLM, those given a conceptual roadmap scored higher on a post-intervention exam than unguided novices, at a significance level between .05 and .10, with less spread and with most of the benefit among the weakest at extracting help from the tool [+W]. In a construction-engineering course, learners using AI with instructional prompts, using AI without them, or studying slides did not differ significantly in multiple-choice accuracy [~M]. A non-significant difference does not show the conditions were equivalent. Both studies are under Evidence and qualifications below.

This page owns the general relationship: support versus none. The Scaffolding and Fading principle holds a narrower model inside it, in which support is contingent (raised after failure, lowered after success) and then withdrawn, and the outcome is unassisted performance. Whether support must be contingent and must fade to help is a question that page treats; the evidence below does not settle it, and one synthesis found no difference by fading. The cognitive apprenticeship pattern is a reusable policy built on that narrower model.

Fitting the design to a situation

Ask for three facts when the brief omits them: what the learner does on the task without support, whether the outcome that matters will be measured with the support removed, and who or what gives the support (a tutor, a teacher with a class, software, an AI assistant).

If the brief says…	Then change…	Basis
Learners cannot yet complete the target task alone	Give support during the attempt (prompts, partial structures, models) rather than leaving them to attempt it unsupported, and judge learning on a later task without the support	Scaffolding improves learning [+S]
online-self-paced or computer-based STEM problem solving	Build conceptual, strategic or metacognitive scaffolds into the tasks; do not rely on motivational messages alone, which showed no effect on their own	Computer-based scaffolding has a positive effect in STEM [+M]; Scaffold design features did not moderate the effect [~M]
The objective is applying principles, not recalling concepts	Aim the scaffolds at choosing and applying the principle; the effect was larger at the principles level than the concept level	Scaffolding's effect is greater at the principles level [+M]
The course is project-based, or design-based	Use scaffolds in project-based work; in design-based learning the pooled effect was not distinguishable from zero, so check unsupported performance there	Scaffolding's effect is highest in project-based contexts [~M]
Learners are underperforming or below grade level	Expect a smaller gain from the same scaffolds; check unsupported performance more often and add explicit teaching of the missing knowledge. Scaffolds that structure the domain task can narrow a gap tied to low prior knowledge (one whole-class trial), but heavy regulative scaffolding (goal trees, cues, templates) can crowd out the domain work: in one inquiry-software study lower-achieving pairs built much worse models with the full regulative tool than with an empty one. Aim support at the task the gap is in, keep regulative scaffolds light, and judge them on the domain outcome	Education population moderates the scaffolding effect [~M]; Scaffolding mitigated the limiting effect of low prior knowledge [+W]; Empty-PC dyads built better domain models, concentrated among lower achievers [-M]
classroom, whole-class inquiry with learners who differ in prior knowledge and reasoning	Combine a static structuring scaffold on the task with teacher modelling and adaptive verbal prompts that problematise, and give both procedural and conceptual help where learners stall (for example on graphs); in one trial with third graders this raised posttest reasoning over the same inquiry unit without the scaffolds and almost removed the disadvantage of lower prior knowledge and reasoning. Expect the scaffolded version to lean on different capacities: inhibition predicted learning only under scaffolding, so check who is still left behind	Scaffolded third-graders outperformed control students on irrelevant-evidence reasoning [+M]; Scaffolding almost fully diminished the disadvantage of lower logical reasoning [+W]; Inhibition predicted posttest scores only under scaffolding [~W]; The teacher as adaptive expert in real-time labs [+W]
Groups solve an ill-structured problem together (collaborative problem solving, collaboration scripts)	Build hard scaffolds into the task (prompts for each problem-solving phase) and expect more solution evaluation and better group solutions while they are present; then test the group on a later task without them, because neither a collaboration script nor a reflection scaffold improved interaction in a later unsupported collaboration. In a meta-analysis of collaboration scripts, script effects on collaboration skills did not differ significantly between scripts paired with content information and scripts paired with content scaffolds, at 14% post-hoc power; an editorial reports the same meta-analysis as finding positive script effects on domain learning and collaboration skills. Where scripts differ in how much they prescribe, a fuller script can teach more of the domain-specific skill: in one experiment on mathematical argumentation only the maximal script beat the minimal one. If learners may choose the script level, watch who keeps choosing the maximum: pairs who did so more often scored lower on the social side of argumentation (a correlation, so weaker learners may simply have chosen more help)	Scaffolded tasks elicit more solution-evaluation turns [+W]; Scaffolded tasks had process distributions closer to the ideal [+W]; Final solution scores were higher on the scaffolded task [+W]; No script or reflection effect on later interaction quality [~W]; Content support did not moderate script effects [~W]; CSCL scripts show positive effects on domain learning and collaboration skills (editorial report) [+W]; Only the maximal script outperformed the minimal script on domain-specific argumentation [+M]; More choices of the maximal script correlated with lower socio-discursive argumentation [~W]
Software gives small groups metacognitive prompts (planning, monitoring, reflecting), for example from an on-screen avatar	Prefer problematising prompts, which pose a question the group must resolve, over structuring prompts, which give an example to follow, and judge the prompts on what they target: in one randomised study with collaborating triads, the prompts raised metacognitive activity and metacognitive knowledge but did not change individual domain knowledge (effect sizes near zero), and problematising prompts gave small gains over structuring ones in metacognitive knowledge and transfer and a medium gain in the group's paper. If domain knowledge is the goal, add support aimed at the domain task	Metacognitive scaffolding had no effect on individual domain knowledge [-M]; Metacognitive scaffolds increase metacognitive activities, with no difference between scaffold types [~M]; Problematising scaffolds: small effect on metacognitive knowledge [~M]; Problematising scaffolds: small effect on transfer [~M]; Problematising scaffolds: medium effect on group performance [~M]; Problematising scaffolds elicit constructive, interaction-embedded metacognitive activity [~W]
Software gives individual learners self-regulated-learning scaffolds	Combine more than one kind of scaffold (for example planning plus monitoring prompts) rather than a single one; in a meta-analysis of 29 studies, multiple scaffolds showed a larger pooled effect than a single scaffold. The same synthesis found domain-general scaffolds ahead of domain-specific ones, but the domain-specific group held three studies and the combined group one, and the STEM meta-analysis above found no difference by context-specificity, so do not choose between them on this basis	Multiple kinds of scaffolds are more effective than a single scaffold [+M]; Domain-general scaffolds outperform domain-specific ones [~M]
Learners draw from a text to learn it (learner-generated drawing) and the plan adds support for the drawing	Do not assume the support helps; start with unsupported drawing and test any support against it. In one randomised study (N = 156), neither global nor local drawing support beat unsupported drawing on drawing quality; local support made drawings much less accurate and lowered recognition below the read-only control, and support raised perceived difficulty	Drawing support did not increase drawing quality; local support decreased it [-M]; Local drawing support scored lower than control on recognition [-M]
A teacher scaffolds small groups during class work	Treat prompting as part of scaffolding and plan to diagnose a group's understanding before helping. In one classroom intervention, teachers who held that prompting was not scaffolding gave less support, and the trained teachers' more contingent support did not raise scores on the knowledge assignment; on a multiple-choice test its effect depended on how long groups had worked independently. So check student outcomes, not only how contingent the teacher's help became	Less successful teachers thought prompting was not part of scaffolding [~W]; No contingency effect on the knowledge assignment [~M]; Contingency's effect on achievement depended on independent working time [~M]
Material scaffolds in a class (worksheets, software prompts) will be faded over a unit	Fade them together with teacher moves that take over their work: monitoring the class, prompting peers to share ideas, extending the materials and handing over independence. In one study of two teachers' classes, one teacher's students held their performance as material scaffolds faded while the other's declined across the unit; the teacher whose students held up used those moves more. Teacher and class are confounded in two classes, so treat this as a reason to watch performance as each scaffold is withdrawn	Performance trajectories as material scaffolds faded differed by teacher [~W]; The two teachers' scaffolding moves differed [~W]
Software gives regulative scaffolds (planning, monitoring, report templates) during inquiry	Judge them on the domain outcome, not on the product they template: a full regulative tool gave better-structured, more complete lab reports, while pairs with the empty tool built better domain models	Full-PC dyads wrote better lab reports [~M]; Empty-PC dyads built better domain models [-M]
Learners attempt problems before instruction (productive failure)	Either nudge learners toward suboptimal solutions (failure-driven) or give low-specificity heuristics toward the right one (success-driven); conceptual posttests were similar, and failure-driven support gave higher-quality constructive reasoning, so prefer it when reasoning is the goal	Failure- and success-driven scaffolding produce similar posttest understanding [~M]; Failure-driven scaffolding gave higher-quality constructive reasoning [+W]
One-to-one tutoring is available	Have tutors prompt the learner to produce the next step before they explain; prompting alone matched ordinary tutoring in one small study	Tutoring effectiveness comes from scaffolding [~M]
The support is an AI assistant that can produce the answer	Limit it to hints and questions rather than solutions, and test without it; unrestricted help raised assisted work and lowered unassisted results. Consider teaching learners how to use the tool as a partner rather than an answer source: in one trial structured training of this kind beat problem-solving training on a later problem-solving test. For novices, add a short conceptual roadmap of the topic; its benefit in one study was borderline. Do not leave the level of help to the model's defaults: in one corpus of programming-course tutoring chats, high-assistance replies dominated across request types and assignments. Two second-hand reports (high-school maths with GPT-4, and a skill-building task) describe the same pattern of better assisted and no better or worse unassisted performance	AI assistance impairs unassisted fraction performance [-M]; Supported and independent writing performance diverge [~M]; Structured GenAI training improves problem-solving performance [+M]; Conceptual-map scaffolding and novice LLM performance [+W]; LLM scaffolding level varies, but high-assistance responses dominate [~W]; GPT-4 access: more solved in the moment, worse after removal (review attribution) [-W]; AI-assisted performance does not transfer to unassisted performance [-W]
Second-language readers get reading materials adapted to their proficiency (glosses, sentence unpacking, labels), for example generated by AI	Use them for comprehension of the assigned text, and check reading on a new passage without them: in one randomised EAP study, differentiated materials raised comprehension while present, and on an immediate unsupported passage material condition made no significant difference, so neither a dependence cost nor a carried-over gain was shown (equivalence was not tested)	Differentiated materials exceeded unified ones on comprehension within each proficiency stratum [~M]; No material effect on immediate unsupported application [~W]
Prompts are delivered in a computer module to novices	Write the prompts as text rather than diagrams, and do not judge the format by practice accuracy alone: text prompts gave more correct first attempts, pictorial prompts more correct second attempts, and text prompts gave higher near transfer and stronger reported continuing motivation in the same 51-learner experiment. Far transfer was at the floor in every condition, so the study says nothing about far transfer	Textual prompts give better near transfer than pictorial [+M]; First- and second-attempt accuracy differ by prompt format [~M]; Textual prompts yield stronger continuing motivation [+M]; Far transfer was at the floor in every condition [~M]
The module could show prompts automatically or let learners call them up	Either is defensible on performance: posttest totals did not differ significantly between module-controlled and learner-requested prompts. Novices rated module-controlled prompts more favourably, so default to them and record whether a learner-requested version is used	Prompt presentation had no significant performance effect [~M]; Learners rated externally regulated prompts more positively [+M]
Learners are moving from studying worked examples to solving problems alone	Withdraw the support in steps (complete examples, then increasingly incomplete ones, then independent problems), removing the last worked-out steps first; the evidence is for near transfer, not a fading rate or far transfer. See Scaffolding and Fading	Fading support promotes the transfer of responsibility [+S]
A learner already does as well without the support	Make the support optional or remove it; guidance that helps novices can become redundant	Expertise reversal effect [~S]
Prompts will run across a whole term (learning journals, weekly writing)	Fade each prompt once the learner applies the prompted strategy satisfactorily, rather than presenting it every week: in two longitudinal field studies, permanent cognitive and metacognitive journal prompts helped early in the term and left lower learning outcomes by its end, while adaptive fading sustained strategy use	Journal prompts helped early, became detrimental by term's end; adaptive fading mitigated this [~M]
The strategy can be taught explicitly instead	Consider explicit strategy instruction first; a dynamic-assessment synthesis ranked it above scaffolding	Contingent scaffolding improves learning [~M]
Language teachers plan to support learners mainly by repeating or translating	Plan support aimed at the step the learner is stuck on; repetition and translation used as unplanned support were reported to hinder learning	Repetition and translation as scaffolding can hinder learning [-W]

Observation, state and explanation

What can be observed is a response under stated conditions: what the learner attempted, which support was available or given (record its kind and how much of the task it did), whether the learner used it, and what they produced. "Cannot yet do it alone" and "can do it with help" are inferred states, and each depends on the task and on the support offered. A support can make a task easier by doing part of it for the learner; success then shows what the learner and the support did together, not what the learner has learned.

Response	Plausible explanations	Observation that could change the interpretation
Fails the task unsupported, succeeds with a prompt	The prompt cued knowledge the learner already had; the prompt supplied the missing step; the unsupported attempt failed for a reason unrelated to the step (instructions, access, time)	Give a matched task with a weaker cue (a question, not the step) and record which level is enough; check the unsupported attempt's instructions and response mode.
Fails even with support	The task is too far beyond the learner's current knowledge for this support to bridge; the support is the wrong kind (structuring when the gap is a concept); the learner did not read or use it	Ask the learner to say what the support tells them to do; try a different kind of support on a simpler component of the same task.
Succeeds with support but cannot explain the result	The support did the reasoning and the learner followed it; the learner has the procedure without its rationale; a language or expression barrier	Ask for an explanation in an accessible mode, and pose a changed case the followed steps would not solve.
Does as well without the support as with it, or worse with it	The support is redundant for this learner (expertise reversal); the support adds reading or split attention; the learner ignored it	Offer the task with support optional; record whether it is used and whether accuracy or time changes.
A group produces a better solution on a scaffolded task than on an unscaffolded one	The scaffold structured the product and did part of the process; the scaffolded task was easier or different; the group learned a better problem-solving process	Give the group a comparable later task without the scaffold and record whether solution evaluation and solution quality hold; in one study a script and a reflection scaffold left later unsupported interaction unchanged.
Takes every support offered	The learner needs it; the learner values finishing over independence; using support carries no cost while trying alone does	Ask privately what the learner is aiming for; vary whether support is on request or automatic on a matched task. Record how often the fullest support is chosen and compare with a later unsupported task: in one script study, pairs who chose the maximal script more often learned less on one outcome, which cannot say whether the choice or the learners' starting point did it.

These rows are proposals for interpreting a response; none has been tested with learners as a diagnostic. A probe is itself practice and can change the state it measures, so record each one. Keep the explanations that remain open rather than choosing the one whose name fits.

Evidence and qualifications

• Scaffolding improves learning [+M]. A random-effects meta-analysis of 144 experimental studies (333 outcomes) of computer-based scaffolding for STEM learners from primary school through adult education, working in ill-structured, problem-centred curricula, reports a positive effect on cognitive outcomes (ĝ = 0.46). The effect did not differ by context-specificity or by whether or how scaffolding was faded, and was greatest at the level of principles and among adult learners (adult learners [~W]: that estimate rests on one outcome, so it is no basis for expecting adults to gain more). A systematic review of teacher–student scaffolding (1998–2009) found 8 controlled effectiveness studies, mostly small one-to-one tutoring experiments on simple, well-structured tasks, favouring scaffolded or contingent conditions over no or less-contingent support, with no pooled effect and thin classroom evidence. This is the most direct support for the support-versus-none relationship. A pooled program-level effect forecasts no individual learner's response, and the comparison conditions are not described on the claim page. Not yet checked against its sources. The claim's opening sentence says the benefit holds "provided the support is faded", which its own meta-analysis entry does not show.
• Contingent scaffolding improves learning more than fixed or absent support. [+M]. In one-to-one long-division tutoring (fourth and fifth graders, 8 per condition, coded as a quasi-experiment), fully contingent support beat moderate, high, partly contingent and no support at an immediate and a one-month test. With 11 tutor–student pairs (grade 8, the circulatory system), more interactive tutoring gave similar immediate learning but better transfer. A meta-analysis of dynamic assessment ranked explicit strategy training above scaffolding and scaffolding above coaching. Not settled: the abstracts could not confirm the entries. It supports support over none in tutoring; it does not show a classroom-scale effect, and it does not show that scaffolding beats explicit strategy instruction.
• Human tutoring has a medium effect over no tutoring, and in one small study students learned as well when tutors only prompted them as when tutors also explained and gave feedback [~M]. In one-to-one tutoring of 8th graders on the circulatory system (11 dyads per study), students whose tutors withheld explanations and feedback and only prompted learned as well as students in ordinary tutoring. A review of tutoring experiments reports human tutoring at d = 0.79 against no tutoring, close to intelligent tutoring systems at d = 0.76. The claim page itself says its evidence supports prompting-type scaffolding more than feedback as the active part, so its title overstates. It bears on which kind of support matters, not on support versus none in groups. Not settled: the abstracts could not confirm the entries.
• Instructional guidance that helps novices can become redundant or counterproductive as expertise grows. [~M]. A synthesis of cognitive-load studies: integrated explanations, worked examples and step-by-step guidance help novices by reducing unnecessary search, and can become redundant for more knowledgeable learners, who may learn more efficiently from leaner tasks. This limits the relationship to learners for whom the task is new; it gives no criterion for when an individual has crossed over. Not settled: the abstract could not confirm the entry.
• Conceptual-map scaffolding modestly improves novice LLM-condition performance and reduces outcome variance [+W]. Within the LLM arm of a randomised experiment, novices given a conceptual roadmap with recommended sequencing scored M = .45 (SD = .14) on the post-intervention exam against M = .38 (SD = .22) for unguided novices, a difference the authors call weakly significant (p between .05 and .10). The scaffolding × AIC interaction was negative at the same level, so the benefit was largest for participants weakest at extracting help from the AI. This tests support against none, in favour, but not at conventional significance; the claim's title ("improves") reads more firmly than the result. Whether the exam was taken with the LLM is not stated on the claim page. Working paper.
• Prompted, unprompted, and slide-based learning conditions yield statistically comparable multiple-choice performance [~M]. A randomised experiment (n = 95) in construction-engineering education compared AI use with instructional prompts, AI use without them, and slides. Multiple-choice accuracy means were 0.72, 0.59 and 0.63; the ANOVA was not significant (p = 0.099) and equivalence was not tested. It qualifies the relationship: guidance on how to use the tool did not show a significant gain on a recall-type measure in this study; it does not show the guidance had no effect.
• Structured GenAI scaffolding training improves students' problem-solving performance [+M]. An RCT with 95 sophomore STEM students compared structured training in using generative AI as a collaborative partner with problem-solving training; the trained group scored higher on the posttest (d = 0.814). The comparison is between two trainings, not support against none, and the claim page does not say whether the posttest was taken with the AI; its title speaks of "AI-assisted revision performance" while the quoted result is a posttest score. It bears on how to set up AI support, not on whether support helps.
• Fading support promotes the transfer of responsibility from instructor to learner. [+S]. A systematic review of teacher–student scaffolding names fading as one of its three characteristics, and three experiments (20 ninth graders, 54 and 45 college students) found a gradual move from complete to increasingly incomplete worked examples to independent problems favoured near transfer over static example–problem arrangements, with removing the last steps first better than the first steps first. This compares two ways of giving support, not support against none, and belongs to the narrower model on Scaffolding and Fading; one synthesis found no moderation of the scaffolding effect by fading. Not settled: the text available could not confirm the entries.
• Teacher repetition and translation as unplanned scaffolding can hinder rather than facilitate learning [-W]. A forum piece reports a study of teacher–student exchanges in Cypriot after-school language programmes in which teachers' repetition and translation, used as unplanned support, hindered learning. It is a second-hand report with no design details or outcome measure on the claim page, so it shows only that something called scaffolding can fail; it does not show when. Not yet checked against its sources.
• Scaffolded third-graders outperformed control students on reasoning based on irrelevant evidence after instruction [+M]. In whole-class inquiry lessons (N = 143 third graders), a condition adding a static structuring scaffold, teacher modelling and adaptive problematising prompts scored higher at posttest on reasoning from irrelevant evidence than the same unit without them (0.71 against 0.54 of items solved, d = 0.49, 95% CI 0.12 to 0.85; gain d = 0.73). It is the wiki's one whole-class test of support against less support, at an immediate posttest. In the same data, prior knowledge predicted posttest reasoning, and combined with teacher-judged logical reasoning to predict it, only without the scaffolds (prior knowledge [+W]; logical reasoning [+W]), while inhibition predicted it only with them (inhibition [~W]). The p-values printed for the experimental condition (0.054 for prior knowledge) are near the threshold, so "no relation under scaffolding" is not shown. The support was a bundle, so which part did the work is not known.
• Groups' final solution scores were significantly higher on the scaffolded task than the non-scaffolded task [+W]. Eleven undergraduate engineering groups each did one scaffolded and one non-scaffolded ill-structured task; on the scaffolded one they spent more turns evaluating solutions (turn proportions [+W]), their process distribution was closer to the authors' ideal (Gini [+W]) and their solutions scored higher. The scaffold is confounded with the task, the scales differed (rescaled from 6 to 5 points), the rigour is coded r1, and every outcome was measured with the scaffold present, so this is assisted performance, not learning.
• Neither the external collaboration script nor the reflection scaffold increased the quality of interaction in the subsequent collaboration [~W]. A laboratory experiment compared a script, a collaborative reflection scaffold and no support, then had groups collaborate unsupported: no effect on interaction quality was found. Equivalence was not tested. It is a direct test of whether collaboration support carries into unsupported work, and it did not show that it does. The degree of additional content-related support does not significantly moderate CSCL script effects [~W] is a meta-analytic subgroup contrast (d = 1.13 with content information, d = 0.70 with content scaffolds, not significantly different, post-hoc power 14%), so it cannot show the content scaffold adds nothing.
• Empty-PC dyads built significantly better domain models than full-PC dyads, an effect concentrated among lower-achieving dyads [-M]. In inquiry-learning software, pairs given a regulative tool filled with goals, hints and cues (PC+) wrote better-structured, more complete lab reports (lab reports [~M], d = 1.49 and 1.74) than pairs given the empty tool, but built worse system-dynamics models (d = 1.26 for the empty tool), with the gap about 10 points for lower-achieving pairs and 1 point for higher-achieving ones. The reports were scored on the template the full tool supplied, so they are aligned with the support. It contradicts the reading that more support helps weaker learners most, for regulative support; it compares more support with less, not with none.
• Failure-driven and success-driven scaffolding before instruction produce similar posttest conceptual understanding [~M]. A classroom study (N = 221) of problem solving before instruction compared two kinds of support: nudging learners to generate suboptimal solutions, or giving low-specificity heuristics toward the optimal one. Conceptual posttests were similar (not shown equivalent), and failure-driven scaffolding gave higher-quality constructive reasoning [+W]. It compares kinds, not support against none; no effect size is printed.
• Metacognitive scaffolding had no significant effect on individual domain knowledge [-M]. In a randomised study with collaborating triads in a computer-based learning arrangement (52 triads; control, structuring and problematising conditions), metacognitive scaffolds left free recall, two knowledge tests and transfer unchanged against no scaffolds (r = 0.01 to 0.08), while raising metacognitive activity [~M] in a companion analysis. Problematising scaffolds beat structuring ones on metacognitive knowledge [~M] (r = 0.16), transfer [~M] (r = 0.13) and the group paper [~M] (r = 0.28), one-tailed tests at p < .05; recall and knowledge-test differences between the two were non-significant and negative. This is support against none with a null on the domain, and it fits the reading that support helps the outcome it is aimed at.
• Neither global nor local drawing support increased drawing quality, and local support decreased it [-M]. A four-condition randomised experiment (N = 156) on tablet-based learner-generated drawing: drawing accuracy was 36% unsupported, 33% with global support and 17% with local support (partial η² = 0.35), and on recognition the read-only control scored above local support [-M]. It is a direct test of support against none in which one kind of support did harm, and is the clearest case in the wiki for the revision condition below.
• Multiple kinds of scaffolds are more effective than a single scaffold [+M]. A meta-analysis of 29 studies of self-regulated-learning scaffolds in computer-based environments: ES = 0.58 for multiple scaffolds (18 studies) against 0.21 for a single scaffold (11 studies), Q = 4.42, p = .035. Domain-general against domain-specific [~M]: 0.43 (25 studies) against 0.10 (3 studies), with 1.70 from the one study using both; the small cells make that contrast weak, and the STEM meta-analysis above found no moderation by context-specificity. Subgroup contrasts across studies are not randomised comparisons.
• Only the maximal CSCL script significantly outperformed the minimal script on domain-specific argumentation [+M]. An experiment on mathematical argumentation compared script levels (ANCOVA with pretest, partial η² = .08); in the adaptable condition, choosing the maximal script more often correlated with lower socio-discursive argumentation [~W] (r = −0.36). More prescribed support helped one outcome; learner-chosen maximum support went with less learning on another, which is a correlation, not a test.
• Less successful teachers tended to provide less support because they thought prompting was not part of scaffolding [~W] is an authors' interpretation in a teacher-intervention study of small-group work (rigour r1). In a related classroom study the trained teachers' support became much more contingent, yet the knowledge assignment showed no contingency effect [~M] and the multiple-choice effect depended on independent working time [~M]. Teacher-level change in support did not translate simply into achievement.
• As material scaffolds faded, one teacher's students returned to their starting level while the other's declined [~W]. Two teachers' sixth-grade science classes (28 and 44 students) across High, Medium and Low material support; the teachers' recorded moves differed [~W]. A design study: with one teacher per class, teacher and class cannot be separated.
• Permanent journal prompts helped early in a term but were detrimental by its end; adaptive fading mitigated this [~M]. Reported in a special-issue introduction (Kalyuga & Renkl 2010) from two longitudinal field studies in developmental psychology courses; the fading group applied increasingly more of the prompted strategies. It is the wiki's one across-term test of how long the same support should stay, and it is second-hand.
• Immediate unsupported application on a new passage showed no material-condition effect [~W]. In a 3 × 3 randomised EAP reading study (N = 135), proficiency-differentiated AI-adapted materials raised comprehension while present [~M] (d = 0.60 for low-proficiency readers), but on an unseen passage without glosses or prompts the material effect was not significant (partial η² = 0.013). Assisted and unassisted outcomes again separate.
• Three second-hand or observational AI claims point the same way as the AI row: GPT-4 access in high-school maths [-W] (a review's attribution to Bastani et al. 2025), AI-assisted skill-building gains not transferring [-W] (a theoretical article citing Wiles et al. 2024), and LLM tutors' high-assistance replies dominating [~W] (an observational corpus analysis). None is a primary result read in the wiki.

The wiki has several comparisons of two kinds of support on a common outcome (failure-driven against success-driven before instruction; problematising against structuring metacognitive prompts; global against local drawing support; text against pictorial prompts; maximal against minimal scripts) and one comparing more regulative support with less, but none comparing structuring the task with simplifying it, or procedural with metacognitive prompts directly. It has one whole-class test of scaffolding against the same lessons without it, at an immediate posttest, and none at a delayed horizon. Before transporting these observations, record the learners, setting (one-to-one, small group, computer-based), the support actually given and what it left to the learner, the comparator, outcome alignment and horizon. These studies do not rank one another; their comparisons differ.

The evidence by question

Each question below links a general claim that gathers the studies behind it into one table, one row per distinct study, with each study's design and effect as recorded and no averaged verdict (eval/synthesis/scaffolding.json, scripts/synthesize_claim.py). The count is of distinct studies across the claim and the narrower claims that bear on it.

• Does support during an attempt improve later learning compared with attempting unsupported, and for which learners, contexts and outcomes? Scaffolding improves learning [+S]: 7 distinct studies.
• Must support be contingent on the learner's last response, rather than fixed? Contingent scaffolding improves learning more than fixed or absent support. [+S]: 20 distinct studies.
• Should support be withdrawn, and in what order? Fading support promotes the transfer of responsibility from instructor to learner. [+S]: 7 distinct studies.
• When the support can do the task for the learner (AI assistants), does it cost independent performance? How genAI is used, not who uses it, predicts whether independent performance declines or improves [~M]: 9 distinct studies.
• In what form, and under whose control, should prompts be given? Textual instructional prompts yield significantly higher near-transfer posttest performance than pictorial prompts for novice circuit-analysis learners [+M]: 1 distinct studies.

Further evidence, not yet read against this model

Claims this page cited before it was rewritten as a conditional model. They are evidence about the relationship, but none has yet been re-read against the model above; each says how far its own sources have been checked.

• Peerwise Student Questions 75 Percent High Quality [~M] — does not test the relationship: it reports that 75% of sampled student-authored physics questions met quality criteria, in a course that gave scaffolding activities before authoring; the authors' link to the scaffolding is an untested hypothesis with no unscaffolded comparison.
• Un andamiaje de autoeficacia en un ambiente computacional incrementa el logro de aprendizaje en matemáticas de estudiantes de quinto de primaria que aprenden de forma individual. [+W] — attached 2026-10-15 from Ramírez Ruiz (2018), which proposed "Incorporar andamiajes motivacionales de autoeficacia en ambientes computacionales para favorecer logro equitativo entre estilos cognitivos"; tests this page's relationship. Not read 2026-10-19: this bullet and the two below link no claim page, and the wiki holds none for Ramírez Ruiz (2018), so there is no evidence entry to read against the model.
• El andamiaje motivacional computacional disminuye las diferencias de logro asociadas al estilo cognitivo en la dimensión DIC, favoreciendo equivalentemente a dependientes e independientes de campo. [+W] — attached 2026-10-15 from Ramírez Ruiz (2018), which proposed "Incorporar andamiajes motivacionales de autoeficacia en ambientes computacionales para favorecer logro equitativo entre estilos cognitivos".
• El andamiaje de autoeficacia aumenta la carga germánica y reduce las cargas cognitivas intrínseca y extrínseca en estudiantes que aprenden con un ambiente computacional. [+W] — attached 2026-10-15 from Ramírez Ruiz (2018), which proposed "Incorporar andamiajes motivacionales de autoeficacia en ambientes computacionales para favorecer logro equitativo entre estilos cognitivos"; tests this page's relationship.
• Effects of collaboration scripts combined with heuristic worked examples on mathematical argumentation skills depend on prior knowledge and working-memory capacity, consistent with a Matthew Effect [~W] — attached 2026-10-18 from Carolyn P. Rosé et al. (2017), which proposed "Design scaffolding with attention to over-scripting risk and plan for fading scripting support over time". Read 2026-10-19: its only entry is a journal editorial describing the design of a 108-student factorial study (fading or not, order of scaffold types); it quotes no result, so it cannot inform the model until the study itself is recorded.
• AI scaffolding differs from human scaffolding in availability, feedback timing, and dynamic adjustment, making its developmental impact conditional [~W] — read 2026-10-19 (wave 2): a conceptual argument (q1) that always-available AI help may reduce chances for independent goal maintenance; it tests nothing, and the AI row already rests on experiments.
• Scripted roles scaffold students' cognitive responsibility for building common knowledge in knowledge-building environments [~W] — read 2026-10-19 (wave 2): an editorial's summary of a blended undergraduate study with no comparison condition or learning outcome on the claim page; relevant to collaboration scripts, does not change the model.
• Provided support does not influence task regulation between conditions [~W] — read 2026-10-19 (wave 2): in collaborative inquiry, communication support changed team regulation but not task regulation (a null, equivalence not tested); a process measure that does not bear on later learning.
• Metacognitive scaffolding plus executive attention training yields significant increases in the N2 ERP component associated with conflict processing in preschoolers [~W] — read 2026-10-19 (wave 2): a neural marker in a preschool attention-training study, no statistic printed; not a learning outcome this model predicts.
• Scaffolding targeting analysis-level skills yields much larger effects (g = 0.537) than scaffolding targeting synthesis or evaluation [~W] — attached 2026-10-22 from Nam Ju Kim et al. (2018), which proposed "Differentiate scaffolding design according to which level of higher-order thinking is addressed in problem-based learning".

Objective and learner-valued goal

The designer's objective is usually a capability the learner shows without the support, on a representative task, at a stated horizon and with stated aids. The learner may value something else: getting this task done, not looking stuck, or the real-world result the task stands for. Ask what they want and why, and record it apart from the objective. A learner who values finishing may use every support; one who values independence may refuse support they still need. In both cases how support is used reflects the goal as well as competence.

For example, an adult returning to study may want to submit a correct statistics assignment this week, while the designer wants them to choose and justify an analysis alone in next term's project. A step-by-step prompt sheet serves the first aim and may hide whether the second has been met. Agree which aim the support is for, and check the second with an unsupported task later, not with the prompted one.

What would revise this model?

The expectation should weaken if supported attempts fail to beat unsupported ones on aligned outcomes measured without the support, for comparable learners, outside one-to-one tutoring and computer-based STEM problem solving. It should also weaken if explicit strategy instruction matches or exceeds the gain at lower cost, as the dynamic-assessment synthesis suggests. If the benefit appears only while the support is present, the page is describing assisted performance, not learning, and should say so. If one kind of support (for example, support that does part of the task) consistently fails where another succeeds, the relationship belongs to that kind, not to support in general; local drawing support, which lowered accuracy and recognition below unsupported drawing in one experiment, is the wiki's one instance so far, and metacognitive prompts that left domain knowledge unchanged show support helping only the outcome it targets. Do not protect the model by relabelling every failure as "the wrong support" or "outside the zone of proximal development" after the outcome is seen.

Assisted performance, unassisted immediate performance, delayed retention, transfer and use in the learner's own setting are separate claims. The evidence above establishes no best kind or amount of support, no individual readiness criterion and no classroom-scale effect.

Claims

• Contingency is not treated as a necessary condition of scaffolding in bilingual education research, unlike broader educational research [~W]
• Scaffolding shows a moderate positive effect on conceptual learning outcomes in STEM pedagogical interventions other than direct instruction [+W]
• Fading-out prompts yield more cognitive strategy use and higher learning outcomes than permanent prompts [+W]
• Computer-based scaffolding improves cognitive outcomes in problem-based learning in STEM education with an overall effect of g = 0.385 [+W]
• Metacognitive and strategic scaffolding show larger effects in PBL-STEM than conceptual scaffolding [~W]
• In the replication, children's posttest outcome scores did not significantly differ across the four instructional conditions, failing to replicate the contingent-instruction advantage of D. Wood et al. (1978) [-W]
• Tutor accuracy in applying the contingency rule showed no significant association with children's posttest outcome, efficiency, or autonomy, with near-zero correlations [~W]

Related Principles

• Scaffolding and Fading

Examples

• Hints, worked examples, modeling, and coaching that are reduced as competence grows.
• Backward-faded worked-example computer module for series and parallel circuit analysis

Key Sources

• Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89-100. https://doi.org/10.1111/j.1469-7610.1976.tb00381.x
• Bates, S. P., Galloway, R. K., Riise, J., and Homer, D. (2013). Assessing the quality of a student-generated question repository. https://arxiv.org/abs/1308.2202. doi:10.48550/arxiv.1308.2202
• Reisslein, J., Atkinson, R. K., & Reisslein, M. (2004). Exploring the Presentation and Format of Help in a Computer-Based Electrical Engineering Learning Environment. Arizona State University. https://eric.ed.gov/?id=ED484994

Page metadata

Field	Value
type	principle
id	scaffolding
aliases	pre-assessment-scaffolding-hypothesized-quality-driver, text-system-controlled-prompts-for-novices
title	Scaffolding
description	Temporary support given while a learner attempts a task they cannot yet complete alone may improve later cognitive outcomes compared with unsupported attempts, qualified by the learner's task-specific starting response, the kind of support, the setting and whether the outcome is measured without the support.
status	review
canonical	True
generated	by claude/unspecified on 2026-10-19
trust	unverified — no one has confirmed this page's content
sources	3 — wood-1976, bates-2013, reisslein-2004

Frontmatter is the page's machine-readable record — see the Schema & Guide.