Learning Sciences

Learning sciences — the interdisciplinary research field that studies how people learn and how to design environments in which learning happens, drawing on cognitive psychology, learning theory, computer science and linguistics, and judging its designs with empirical evidence rather than theory alone. In this knowledge base it is the research field around AI in education rather than one of the school subjects: it supplies the mechanisms that AI systems operationalise (knowledge components, mastery thresholds, transfer), the design objects they are embedded in (planned course sequences, tutors, Feedback regimes) and the standards by which they are judged (Learning Gains, Assessment Validity, equity). Its organising question is not whether a tool performs well but whether a learner changed.

Questions to Consider

A learner passes every practice item, so the mastery threshold ends the set — then misapplies the rule where the action should be withheld. Whose error is that: the learner's, the model's, or the stopping rule's?
Sequence mining can describe 554 courses as patterns without observing a classroom. What does the field gain, and lose, by studying designed intentions instead of enacted activity?
Demographic sensitivity in an LLM's feedback looks like adaptation when it tracks a learner's stated education level and like bias when it shifts sentiment. Should a field that cannot separate the two keep using open-ended models to assess?
Does the learning sciences' appetite for causal design — randomised assignment, counterfactual audits, executable models of the learner — narrow what counts as evidence in AI in education?

Introduction

The learning sciences study learning and the design of learning environments, and they are defined by their methods as much as their topics: experiments, classroom trials, quantitative modelling of student data, qualitative analysis of designs and contexts, and design-based research that builds an intervention and revises it in use. That breadth separates this page from the neighbours that supply the frameworks, the practice and the instruments; the section below sets out each boundary and what the field has established on the other side of it.

This page covers the substantive knowledge such methods have produced — what learners do with a generative model, which arrangements change outcomes, and where the field's own instruments fail. AIEd in the Disciplines takes the opposite cut, holding that subject matter changes what support should do; the learning sciences take the cross-cutting view, and the mechanisms they test are gathered under Cognitive Psychology.

How AI appears in the learning sciences

Mechanism first, then the model. An, McLaren and Stamper (2026) ran eleven experiments (N = 192) with Intelligent Tutoring systems for Riichi Mahjong and showed that learners who compiled an overgeneralised production — the action without its application constraint — misapplied it on the first "do-not-act" item at 61.5%–100%, against 12% expected error under Bayesian knowledge tracing. With a 95% mastery threshold the system stopped practice before learners met a case requiring the action to be withheld, so the defect went undetected. Short do-not-act practice with Feedback naming the missing constraint cut misapplication to 0.0%–23.1% (Cohen's h 1.70–2.44), and a secondary analysis of thirteen K-12 Decimal Point datasets found the same structure in whole-number bias (84%–88% of comparison errors).
The optimum depends on the content. Rachatasumrit, Koedinger and Carvalho (2025) treat the example–problem ratio as a content–treatment interaction: in a 2×2 experiment with 95 participants on geometry-area material, practice-only training produced larger learning gains for verbatim facts while example-integrated training produced larger gains for generalisable skills (β = 0.41, p = .038, d = 0.38). A simulated learner (Apprentice Learner) reproduced the crossover only when given an ACT-R-style memory-and-forgetting mechanism. More practice is not uniformly better: memory-oriented content warrants retrieval, and induction-oriented skills warrant integrated examples.
Design as an analysable object. Divjak, Svetec and Horvat (2026) turned Learning Analytics on Learning Design itself, coding 29,064 activities across 554 courses planned in a free course-design tool. Acquisition was the most common learning type and the most common entry point; the strongest Markov transition was Assessment → Discussion (0.332) and the highest-confidence rule was Acquisition → Assessment → Practice → Practice (0.743, lift 1.45). Learning type tracked intended outcome level, Acquisition falling from about 50% of activities at Bloom level 1 to around 20% at level 6. The authors stress these are pre-implementation designs: resemblance to flipped, inquiry-based or project-based sequences is not evidence of intent.
Auditing the models that assess. Rooein, Benedetto and Hovy (2026) audited six LLMs across essay scoring, formative feedback and question answering, holding the task input fixed while varying only demographic context (192,480 calls). Scoring was stable under explicit personas, but Llama-70B inflated its own scores by 1.57 points under implicit conversational history (p < 0.001), and higher education drew less readable and more positive responses — a sentiment gap of roughly four standard deviations. Readability effects shrank while length effects grew, and some coefficients changed sign between conditions. The authors offer it as an audit instrument, not a deployment verdict, and read the entanglement of demographic and topical signal as a threat to validity and equity.
Measuring competence, and its limits. Verí (2026) organises instruments for competent Generative AI use into four domains — knowledge and use, epistemic oversight, reliance calibration, and control of tool-using agents — refusing to collapse them into one proficiency continuum. Three same-sample correlations between self-rated and demonstrated AI literacy pooled to r = .055 (95% CI [-.047, .156], reported N = 2,765), which the author reads as enough to reject treating self-ratings as interchangeable with performance scores, though not enough to set a cutoff. No validated instrument covered the full set of decisions that tool-using agents create; the proposed layered battery is a design hypothesis.
Self-report about the learner's own offloading. Alam (2026) translates the Cognitive Offloading literature into PAUSE, a browser-only self-check with four domains, every LLM-era item anchored to a source, no composite, no storage and no model in production; its bands are descriptive rather than normed, and a reading must not justify assessment, admissions or hiring decisions. Its stated limits matter: self-report of offloading is vulnerable to the faculty it concerns, a respondent who deliberately uses AI as a scaffold reads as offloading on several items, and whether AI-associated offloading is distinct from general technology dependence remains open.
Where human expertise sits. Wang et al. (2024) report the clearest division of labour: in a two-month randomised controlled trial with about 900 novice K-12 tutors and roughly 1,800 students, real-time suggestions drawn from experienced tutors' reasoning raised topic mastery by 4 percentage points (62% to 66%, p < 0.01), and by 9 points for lower-rated tutors, at about $20 per tutor per year, shifting tutoring toward guiding questions. Gains were proximal — year-end tests did not move. Reichert et al. (2026) find teachers arriving at the same position by design: six secondary teachers prototyping chatbots specified a bounded expert, holding authority boundaries (responsibility for learning and safety is not delegable) and expertise boundaries (the model lacks their knowledge of individual students), and delegating content presentation, practice and corrective Feedback while reserving objective-setting and summative assessment.
Capability at the level of the field. Sutedjo, Chowdhury and Liu (2026) surveyed 127 faculty with a TPACK instrument adapted for generative AI: strong content and pedagogical content knowledge (M = 4.70–5.15) beside markedly lower technology-integrated knowledge, with holistic TPACK lowest at 2.55, content knowledge uncorrelated with any technology-integrated domain, and the three integrated domains correlating so highly (r = .81–.91) that they may function as one factor. Perrotta (2026) reads the governance layer through a discontinued UK civil-service prototype whose codebase was a system prompt plus a retrieval pipeline over commercial models, arguing that the generality of foundation models both enables rapid repurposing into policy tools and makes aberrant output a permanently only-mitigable risk — oversight that peers over the loop rather than sitting inside it.

How the learning sciences relate to their neighbours

Design-based research is the method this field developed rather than borrowed: an intervention is built and revised inside a working classroom, with its theoretical rationale revised alongside it, so one study yields both an artefact and a design principle. That is what separates it from a laboratory experiment, which isolates a cause by holding the context still, and it is why the field's findings arrive as design knowledge rather than as effect sizes. Research Methods in AIED takes the other cut: that page surveys the whole repertoire — experiments, surveys, qualitative work, benchmarks, reviews, consensus methods — as a choice among instruments, weighed for the validity of the claim each can support. This page reads the same corpus from the substantive side, asking what the repertoire has established about learning and judging a method by whether its design claim survives contact with learners.

Learning Theories collects the candidate frameworks — behaviourism, cognitivism, constructivism, sociocultural accounts, motivation and self-regulation — as lenses for reading AI. The learning sciences share that vocabulary but not that stance: here a theory is a claim about mechanism that a design must either instantiate or refute, and the field's standing rests on empirical and design work rather than on the coherence of a framework. The theory page is the one to open for what a framework asserts; this page is the one for the evidence a framework has accumulated.

The field also builds theory rather than only testing borrowed frameworks: Theory Development in AI in Education covers the conceptual work that explains how learners, teachers and AI systems interact, and it is where the field's own constructs are argued before they are measured. What its designs are usually asked to produce is transfer — knowledge and skill that survive past the tutor, subject or task they were learned in — which is why a gain measured inside a tool counts as a weaker claim than one measured without it. And because a designed environment is a compound intervention, attributing an outcome to one component is the field's standing measurement problem: Educational Measurement supplies the psychometric apparatus that makes the attribution arguable at all, which is why measurement questions arrive early here rather than after the fact.

Cognitive Psychology is the mechanism-level discipline the field draws on most heavily, supplying bounded working memory, encoding and retrieval, decomposable knowledge components and the diagnostic language of learner modelling. The learning sciences use those mechanisms without reducing to them: their unit of analysis is a designed environment carrying social, motivational and contextual variables that a laboratory account of memory does not, and their tests are run on whole interventions rather than on isolated cognitive effects.

Pedagogies and Teaching Strategies and Learning Design cover practice — which teaching strategy to use, and how to sequence objectives, activities and assessment into a course. Both are what the learning sciences study from the outside, as objects of description and evaluation; the field does not tell a teacher which tactic to reach for next, it reports what the tactics have been shown to do. Learning design is the closer relative, since both produce something that can be implemented and tested, but the designer's output is a teachable course and the field's output is knowledge about designs in general.

The findings only matter once they reach teaching, and that journey runs through three pages. Educational Development is the institutional practice that carries them — faculty development, standards, policy and identity work decide whether a validated design ever reaches a classroom, which is why the field's evidence routinely leads what institutions have implemented. Professional Development is where the knowledge has to land before a teacher enters the room, and Teaching is where it lands afterwards, in the moment-to-moment judgment about when to intervene, which instrument to use, and when to leave a learner alone. None of the three produces learning-science findings; all three decide whether those findings change practice.

Connected Concepts

Connected Articles

Beyond AI Literacy: A Structured Review and Exploratory Meta-Analysis of Measures for Competent Generative-AI Use — Review and exploratory meta-analysis of measures for competent generative-AI use (Verí 2026)
Deceptive Overgeneralization: When Adaptive Learning Enables Systematic Misapplication — Correctness masking an incomplete rule: mastery stopping rules in adaptive learning (An, McLaren & Stamper 2026)
The Role of Implicit and Explicit Demographic Signals in Large Language Model-based Student Assessment — Counterfactual audit of demographic signals in LLM student assessment (Rooein, Benedetto & Hovy 2026)
Learning paths and patterns in learning design: Insights from 500+ courses — Markov chains and pattern mining over 29,064 activities in 554 courses (Divjak, Svetec & Horvat 2026)
PAUSE: A Privacy-Preserving Self-Reflection Tool for AI-Associated Cognitive Offloading — A privacy-preserving, non-diagnostic self-check for AI-associated offloading (Alam 2026)
Zero-Shot Governance: General-Purpose AI in Policy — Zero-shot governance: general-purpose AI in policy, read through the Redbox codebase (Perrotta 2026)
Evidence and Theory for why the Best Example-Problem Ratio To Optimize Learning Gain Depends on Knowledge Content — Why the best example–problem ratio depends on content (Rachatasumrit, Koedinger & Carvalho 2025)
Human-Centered Design of LLM-Powered Educational Chatbots: A Study with Secondary Teachers — Teachers design bounded-expert chatbots, with selective delegation of instruction (Reichert et al. 2026)
Assessing faculty self-perceived knowledge in using generative AI to teach 21st-century skills — Faculty GenAI TPACK: strong content knowledge, weak technology-integrated knowledge (Sutedjo, Chowdhury & Liu 2026)
Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise — Tutor CoPilot: a randomised trial of human–AI live tutoring at scale (Wang et al. 2024)

Metadata

Discipline	Learning Sciences
Level	K-12, Higher Education, Adult Learners
Intended audience	Researchers, Instructional Designers, Learning Designers, Teaching, Policymakers
Learning and instruction	Cognitive Psychology, Learning Theories, Pedagogies and Teaching Strategies
Technologies and techniques	Intelligent Tutoring, Learning Analytics
Research methods and evaluation	Research Methods in AIED
Foundations of AI in education	Learning Design
Page kind	Framework, Synthesis

Source History Edit Chat