AIEd Knowledge BaseAIEd KB
On this page

Concept

Physics Education

created 2026-08-09 · updated 2026-11-04

Physics Education — the study of how students learn physics and how to teach it more effectively, spanning Socratic AI tutoring, computational thinking assessment, student Trust and AI adoption patterns, automated scoring validity, and teacher preparation. The physics education articles in this knowledge base are notable for their domain-specificity: they explore how AI tools interact with the unique cognitive demands of physics reasoning — visual-spatial thinking, mathematical modeling, abstract systems thinking, and multi-step Problem Solving.

Questions to Consider

  • Physics problems often require visual-spatial thinking, mathematical modeling, and multi-step reasoning. Why might these be precisely the cognitive demands that current AI tutors struggle with?
  • Students report a big trust-utility gap—91% use AI for coursework but only 41% trust it. Have you experienced using a tool you didn't fully trust? What drove the gap?
  • AI scoring systematically underestimated linguistically weak students' physics explanations. What does that suggest about how an AI grades an explanation versus a correct numeric answer?
  • One study found a Socratic AI chatbot dramatically improved question specificity in a live physics course, but students also frequently 'ceded strategic control' to the tutor. When does handing over strategic control help learning, and when does it harm it?
  • Why might physics be a 'proving ground' for AI in education—what makes its problems ideal for studying how AI affects reasoning and assessment?
  • Would you trust an AI to grade your physics problem set or reason through a force diagram with you? What would need to be true about the AI—and about your course—for you to say yes?

Introduction

Physics education research has become a proving ground for AI in education because physics problems are well-structured yet cognitively demanding, making them ideal for studying how AI tools affect learning, reasoning, and assessment. The 26 articles in this knowledge base collectively paint a picture of a field grappling with both the promise and the limits of AI — from Socratic chatbots that improve student question quality to systematic scoring biases that penalize linguistically diverse learners.

Key research themes

Socratic AI tutoring in physics is the most developed theme, with three articles deploying Large Language Models (LLMs)-powered Socratic dialogue in real physics courses. Hashmi et al. demonstrated that sustained Socratic interaction with an AI chatbot dramatically improves question specificity in introductory mechanics, with 150 STEM majors in a live course. Hashmi & Rebello built a bottom-up taxonomy of 357 student discourse categories from the same deployment, revealing that meta-procedural turns — where students cede strategic control to the tutor — dominate student interactions. Both contribute to broader Socratic Method research and connect to AI Tutoring and Intelligent Tutoring frameworks.

Student AI adoption and trust explores how physics students actually use AI tools. Fouad & Bentley found a 50-point trust-utility gap: 91% use AI for coursework but only 41% trust it, with students spontaneously identifying AI failure modes in visual-spatial reasoning and circuits. Becker et al. developed a two-profile typology — 70% "Pragmatic Users" and 30% "Skeptical Non-Users" — from 1,189 survey responses, showing both groups make calculated risk-utility trade-offs. These studies advance AI Literacy and Trust Calibration research, and challenge one-size-fits-all AI policies.

Use is selective, and approval outruns it. In a survey of 209 engineering students at six Kazakh universities, 36.1% reported regular general AI use, but only 20.2% used it very frequently in physics and specialized disciplines (Nauryzbayeva et al. (2026)). Use concentrated on explanation-seeking tasks — 55.1% for clarifying difficult topics and 51.4% for homework — ahead of written assignments at 43.5%. Engineering thinking drew the strongest perceived contribution at 61.1%, yet the instrument measured only use patterns and perceived support, not learning. Nothing here shows that AI improved physics learning, and the authors say so.

Structured use, not merely adoption. The adoption studies above describe who uses AI and how much they trust it. A separate strand asks what instructional design should do with that use, and multi-institutional evidence on it has been scarce. Paul et al. (2026) ran the four-step Physics AI-Replication protocol in 11 algebra- and calculus-based sections at four institutions: identify a hard problem, get an AI explanation and solution, request similar practice problems, then solve them independently before checking.

Perceptions were consistent across campuses. Of 358 enrolled students, 235 were surveyed (65.6%); 90.9% credited improved conceptual understanding and 97.7% found the replicated practice helpful; post-engagement confidence averaged 7.18 on a 1–10 scale with institutional means from 6.86 to 7.52; and the dominant qualitative theme held between 60.6% and 67.6% at every site. Confidence tracked those two protocol benefits (r_s = 0.37 and 0.35) but not AI usage frequency (r_s = 0.02) — the variable the adoption literature usually measures.

Epistemological beliefs track chatbot-behavior preference. In a large calculus-based course, 52% of students preferred a chatbot beginning with guided inquiry and answering only when asked, and they scored higher on the EBAPS total (p = 0.029) — differences that did not survive a Bonferroni correction, making the association suggestive, not conclusive (Sirnoorkar and Mamidpalliwar (2026)).

Acceptance has now been modeled rather than described: a 340-student structural equation study of Grade 10 physics classes in Hangzhou explains 58% of the variance in intention to use DeepSeek. Its notable result is that beliefs about physics — utility value, expectancy, cost — predict how useful and how easy the tool seems, not only whether students intend to use it (Xue, Xu, Qi & Lu (2026)).

Perception change without behavior change is what O'Brien et al. (2026) add to this picture. A reflective lesson on how LLMs work, taught in a required first-year course for physics majors, raised skepticism sharply — agreement that LLMs can leave students with a false sense of confidence rose from 58% to 88%, and the belief that an LLM outperforms the average physics student fell from 54% to 32% — while convenience (71% agreement) and deadline pressure (65%) remained the dominant reasons for use. The lesson's limits are as informative as its effect: AI Literacy instruction shifted what students said about these tools and not the pressures that make them reach for one, which is why an intervention of this kind belongs alongside problem and policy design rather than in place of it.

Problem posing turns the learner into the problem author: 76% of 49 physics students perceived a positive change in their GenAI interactions after prompt-engineering training, novices benefiting most, and about 68% viewed the technique positively as self-study (Dawson & Rebello (2026)).

Assessment and computational thinking examines how AI can evaluate physics learning. Savage et al. used LLMs to assess computational thinking growth in introductory physics, finding LLMs can scale CT assessment but struggle with complex constructs like Systems Thinking. Feser & Tschisgale demonstrated that AI scoring systematically underestimates linguistically weak students' physics explanations — a finding that connects to Assessment Validity, Bias Mitigation, and Equity.

An AI grader's disagreement can be a retrieval failure, not a reasoning gap. In experimental physics laboratory reports a dropped sign or misread axis may never reach the model, so document conditions — high-resolution PDFs, editor-written equations — cap any score, and processing belongs behind teacher verification (Abreu et al. (2026)).

Psychometric infrastructure for diagnostic assessment. Le et al. (2026) invert the usual AI-in-physics question: rather than asking whether a model can solve or grade physics, they ask whether the field's own research-based assessments can be made to diagnose it. Mapping FCI, FMCE and EMCS items onto 14 fine-grained learning objectives and fitting a DINA cognitive-diagnostic model to 24,394 posttest responses from 807 courses at 79 institutions via the LASSO platform, they built the Mechanics Cognitive Diagnostic — reported as the first cognitive diagnostic computerized adaptive test in physics. The FCI and EMCS fit well (RMSEA² = 0.033 and 0.022) while the FMCE fit only marginally (0.065), and the authors trace that misfit to instrument design rather than modeling: 42 of 43 scored FMCE items share scenario stems in chained sets, creating the local item dependence that DINA's conditional-independence assumption forbids (the FCI blocks 13 of 30 items; the EMCS none). Classification accuracy met the low-stakes formative benchmark for 19 of 22 objective–assessment combinations; the three failures were EMCS energy objectives whose items overlap by roughly 70 percent, so mastery of one cannot be separated from the others. The significance for physics teaching is that instruments courses already administer can be repurposed to deliver actionable, objective-level feedback during instruction rather than a retrospective posttest score — provided the diagnostic claims are pitched at the resolution the item bank can actually support.

Benchmarking multimodal AI on authentic physics problems. Chen et al. (2026) introduce OmniPhys, a large-scale Multimodal AI Benchmark (15,246 questions, 19,850 images) spanning middle-school through university-level physics from Chinese educational corpora. Unusually, it evaluates not just multimodal input comprehension but multimodal output generation — whether models can synthesize structured physics diagrams, a core component of authentic problem solving. Extensive evaluations reveal critical gaps in current multimodal LLMs, especially in complex reasoning and visual generation.

Instructional-design frameworks for AI-augmented instruction. Kuhn et al. propose the AIRIS framework (Activate–Inquire–Reflect with Intelligent Support) — a three-phase structure for cognitively activated AI use in physics: students predict and sketch expected outcomes before AI (Activate), delegate computational and representational steps to AI while critically comparing output to their own predictions (Inquire), and interpret, check consistency across representations, and reflect on what the AI contributed afterward (Reflect). Grounded in Self-Regulated Learning, Cognitive Load Theory, multiple external representations, and Human AI Collaboration, it frames the central challenge as instructional design rather than cheating or tool choice, and calls for "withdrawal condition" experiments testing whether learning survives the removal of AI support.

  • Teach critique of AI solutions with a rubric. Twenty-four introductory physics lab groups that critiqued an AI solution with MAPS-guided reflection identified skipped numerical integration, undefined notation and qualitative-instead-of-quantitative plotting, while groups that only solved a related problem accepted the polished output or rejected it from their own misconceptions (Borse et al. (2026)).
  • Energy preconceptions decide whether the AI becomes concept talk or a lookup. Wang et al. (2026) ran fourteen pre-service teacher groups through energy forms and conservation in Go-Lab with an on-demand metacognitive chatbot. Groups whose members answered an energy pretest differently asked the same kinds of questions as matched groups, but put AI query negotiation at the center of their interaction network and sustained positive or neutral evaluation of its answers, while matched groups ran a linear query-then-read path. Concept confusion ran f = 87 against 36, with 88.9% against 57.1% case coverage.

Generative video as synthetic experimental data. Alvarado-Cruz et al. (2026) generate video scenarios with PixVerse, Grok Imagine and Pippit for three resistive-force regimes — constant friction, linear drag and quadratic drag — extract the kinematics with the Open Source Tracker tool, and fit the analytical models by non-linear least squares. The synthetic data agreed with the classical equations of motion and recovered physically meaningful parameters, and the recurring practical finding is that prompt specificity governs physical coherence: more detailed descriptions produced more coherent dynamics. The workflow mirrors experimental practice from model construction to quantitative validation, and reframes prompt formulation as a stage of experimental design rather than a convenience. What it does not yet demonstrate is learning: validation here is agreement between generated motion and the authors' models, not students' measurement judgment, so the approach inherits the validity question any generated data used as evidence must answer.

Bespoke simulations for the topics PhET does not cover. LLM-generated HTML/JavaScript models let instructors produce the simulation a course needs rather than the nearest published one, subject to technical and physical validation; in a pilot, 53 students in 26 pairs built and refined their free-fall models with no programming background required (Ben-Zion et al., 2025).

A reusable four-element prompt — tools, display, hand controls, optimization — lets a no-code teacher generate a browser-based AR simulation whose pinch-and-spread gestures drive the physics; in a 29-student pilot the gesture helped all 29 "feel" wavelength (mean 4.52) and 86% reported greater engagement (Levy et al. (2026)). AI-generated instrumentation from a prompt. A rotation lab generated by natural-language prompting of an AI assistant, with no manual coding, measured uniform and accelerated circular motion to within 1% of independent Tracker video analysis, letting instructors build the interface around their pedagogical variables rather than adapt to a precompiled sensor app (Suñer et al. (2026)).

Assisted performance vs. unaided knowledge in a redesigned course. A 2026 redesign of the introductory nuclear and particle physics course at Ruhr University Bochum (Mikhasenko et al.) allowed generative AI on ten deliberately AI-resistant, research-shaped homework sheets designed so that naive prompting would not suffice. Engagement and ambition were high — 24 of 42 students earned credit on all ten sheets, and one derivation filled more than two meters of blackboard — but an unaided 90-minute written exam was a "serious warning": a mean of 20.6/80, with only two of 27 examinees reaching 40. The authors conclude that assisted performance and independently retrievable knowledge are distinct achievements that cannot be assumed to train or demonstrate each other, and that physics courses must reserve some practice for unaided work — reinforcing the knowledge base's broader transfer evidence.

A 2026 PRISMA review of 11 empirical studies across six countries reaches the same conclusion at the level of the field: short-term gains from AI feedback and simulation did not persist into later examinations or transfer to independent performance (review of AI in university physics education).

Agent role design as an instructional variable. Wang et al. (2026) hold the model (DeepSeek R1), platform, and temperature constant and vary only the prompt-specified role: a teacher-centered agent answering authoritatively from a bounded textbook knowledge source, versus a student-centered agent configured as an empathic teacher with knowledge of students' understanding, scripted to diagnose the cause of Misconceptions about AI, name the relevant concept, and transfer to an analogous phenomenon. Across 59 high-school graduates working two conceptual items, the student-centered agent produced higher post-test scores (9.67 vs. 7.93; r = 0.38), lower extraneous and higher germane cognitive load, stronger flow experience (d = 0.92), and higher empathy perception (r = 0.53) — evidence that role framing, not just answer accuracy, is what makes a physics agent instructionally effective (Pedagogical Agent, Prompt Engineering).

  • Benchmark scores understate what models can already do in physics. Re-grading six widely used physics benchmarks with domain experts found that most of the reported shortfall was an artifact of defective items and restrictive automated graders: of 250 audited rejections, 143 (57.20%) were benchmark defects and 95 (38.00%) grader errors, with only 12 (4.80%) genuine model errors. Corrected, HLE-Physics mean@4 rose from 47.28% to 78.66% and CritPt from 32.29% to 87.50%. For physics instruction this cuts both ways: it means students can already obtain expert-level text solutions to many canonical problems, so assessment of physics reasoning needs to move toward items that resist benchmark contamination and toward process evidence rather than final answers. (How Good Are Frontier Models at Physics? Expert Re-Grading Reveals Broken Evaluations and Near-Saturation of Leading Benchmarks)

  • AI grading can reproduce a high-stakes physics outcome without matching every part. Grading 10,364 handwritten Olympiad and university pages, a multimodal LLM recovered the same five-student Olympiad team as examiners and correlated with totals at r = 0.93–0.96, but exact question-part agreement reached 70% — examiner-controlled second reading, not replacement (Pathak et al. (2026)).

  • Physics benchmarks can be curriculum-grounded and non-English. Jahin et al. (2026) built 1,834 Bengali questions from the national Grade 9-10 physics textbook and a 1,760-node knowledge graph; one fixed fine-tuning recipe gained 5.5, 15.0 and 23.3 points at 0.6B, 1.7B and 4B parameters.

Physics education sits within the broader STEM Education domain but has distinctive connections: to Socratic Method through the strong tradition of Socratic dialogue in physics problem-solving; to Computational Thinking through the increasing role of computation in physics; to Assessment Validity through the challenges of scoring physics explanations; and to Workplace Learning through Simulation-based preparation. The Student Experience and AI Literacy concepts are essential for understanding how physics students navigate AI tools, while Educational Measurement and Automated Grading connect to the assessment dimension.

Implications for physics instructors

  • Design for student trust, not just adoption. Fouad & Bentley document a 50-point trust-utility gap (91% use, 41% trust), with students identifying AI failures in visual-spatial reasoning and circuits — create opportunities to expose and discuss these limits rather than assume acceptance.

  • Use Socratic AI to deepen question quality, but watch for strategic ceding. Socratic chatbots improve question specificity, yet taxonomy research finds meta-procedural turns dominate — students hand strategic control to the tutor. Intervene to keep students the decision-makers.

  • A well-engineered tutor beat in-class active learning. Kestin et al. (2025) found that 194 Harvard life-sciences students learned more from the course's own AI tutor than from research-based active-learning lessons covering the same content (median post-test 4.5 vs. 3.5), with higher reported engagement and Motivation, crediting expert-crafted prompts and platform-sequenced problems rather than the technology alone.

  • Structure AI use cognitively, not just permissively. AIRIS (Activate–Inquire–Reflect) shows the value of having students predict/outline before AI, delegate computational steps while comparing output critically, and reflect afterward — treat AI integration as an instructional-design problem, and test whether learning survives AI removal. In a randomized trial with 95 undergraduates, 45 minutes of SRL-aligned training on mechanics problems lifted accuracy when students later revised electromagnetism items with a model: 69.3% of their incorrect answers became correct against 41.3% for untrained controls (d = 0.81) (Huang et al. (2026)).

  • Guard against scoring bias. AI scoring systematically underestimates linguistically weaker students' explanations; use language-aware or human-moderated scoring for conceptual assessment.

  • Use simulated classrooms for teacher preparation. Simulated multi-agent classrooms give prospective teachers rare practice responding to authentic student reasoning — a low-cost complement to live microteaching.

  • Reserve unaided practice and assessment. The Bochum redesign (Mikhasenko et al. 2026) shows AI-permitted, research-shaped homework completed with high engagement can leave students far behind on an unaided exam (mean 20.6/80) — treat assisted performance and independently retrievable knowledge as distinct, and build deliberate unaided practice and a written exam into the course.

  • Verify facilitator fidelity; keep the model-building phase AI-free. Pressed to "just tell us" which can weighed more, a purpose-configured Gem invented mass values no one had measured, so a language-based facilitator's fidelity must be checked rather than assumed, and the AI-free build phase preserves the hands-on work it cannot reach (Tufino & Damiani (2026)).

Common Misconceptions

  • "Students trust AI enough to rely on it uncritically." Fouad and Bentley found a 50-point trust-utility gap: 91% use AI for coursework but only 41% trust it. Widespread use coexists with low trust, so an intervention aimed at building enthusiasm addresses the wrong variable.
  • "Physics students are one population." Becker et al. developed a two-profile typology from 1,189 survey responses: 70% "Pragmatic Users" and 30% "Skeptical Non-Users". A single classroom policy has to serve two groups with opposite stances.
  • "The gap between use and trust is irrational." The two figures measure different things — whether the tool helps you finish work, and whether you believe it does so correctly. Students are distinguishing utility from reliability, which is the analytically correct distinction and not a contradiction.

Connected Concepts

Connected Articles

Connected Resources

  • teacher.dev
    A non-profit studio publishing one free classroom app for every day of a public build streak, from printable figure generators to a quiet-time safari and a local-first hall pass.

Metadata

DisciplinePhysics Education, Stem Education
LevelHigher Ed
Intended audienceLearners, Instructors
Learning and instructionSocratic Method
Technologies and techniquesGenerative AI, Intelligent Tutoring
Foundations of AI in educationAI Literacy
SourceHistoryEditChat

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.