Introduction
Artificial intelligence in education is often discussed through generative tools: writing assistants, tutors, image generators, learning analytics, and automated feedback. This is Digital AI operating mainly through information. A second development is equally important. Physical AI connects computation to sensors, actuators, machines, and material environments. It may take the form of a robot, autonomous instrument, intelligent laboratory platform, or adaptive cyber-physical system.
The two categories overlap and are not a hierarchy. Digital systems can create serious real-world consequences, while a moving robot may be fixed automation rather than AI. The distinction is useful because it clarifies what learners must know, what teachers must govern, and what evidence institutions need when intelligence moves from informational output to physical action.
Working Definitions
Digital AI primarily transforms informational inputs - such as text, images, audio, records, or software states - into classifications, predictions, recommendations, generated content, or code. Its educational risks include error, fabricated support, bias, privacy loss, insecure use, inappropriate delegation, and weakened independent performance.
Physical AI connects information processing with repeated sensing and action in a material environment: Sense → Estimate → Decide → Act → Observe again. It inherits the digital risks and adds sensor error, delay, unstable feedback, infeasible motion, collision, energy and thermal limits, hardware failure, and the need to stop or recover safely.
Education should therefore teach both epistemic responsibility - how knowledge claims are produced and justified - and operational responsibility - how decisions become bounded action. Neither is achieved by prompt technique alone.
Why the Curriculum Must Expand
UNESCO's 2024 student framework organizes AI competence around a human-centered mindset, ethics, techniques and applications, and AI system design, across the levels Understand, Apply, and Create.1 Its teacher framework emphasizes human agency, ethics, AI foundations, pedagogy, and continuing professional learning.2 These frameworks support a move from tool familiarity toward system understanding and responsible creation.
The empirical evidence is design-dependent. In a high-school mathematics field experiment, general-purpose GPT access improved assisted practice yet reduced subsequent unassisted performance; learning-oriented safeguards largely mitigated the decline.3 A separate randomized study in undergraduate physics found higher immediate learning gains from a carefully structured AI tutor than from an in-class active-learning lesson.4 These results are not directly comparable or universally generalizable. They demonstrate why an institution should specify the pedagogy, learner population, task, outcome, duration, and assessment condition before claiming educational benefit.
The OECD Digital Education Outlook 2026 reaches a compatible conclusion: task performance with generative AI should not be treated automatically as learning, while pedagogically designed uses can support learning.5
A Five Stage Learning Progression
A practical curriculum can be organized as Understand → Model → Build → Evaluate → Govern. The sequence is cumulative rather than rigid; learners may return to an earlier stage when evidence exposes a flawed assumption.
| Stage | Curriculum question | Digital AI evidence | Physical AI evidence |
| Understand | What is the system, and who is involved? | Explain output, sources, and delegation. | Identify sensors, actuators, and human control. |
| Model | Which variables, assumptions, and constraints matter? | Map input, model, output, and evidence. | Map state, feedback, uncertainty, and limits. |
| Build | Can the idea be implemented at appropriate risk? | Configure a bounded tutor or analysis workflow. | Prototype first in simulation or low-power hardware. |
| Evaluate | What evidence supports learning and system claims? | Check accuracy, transfer, bias, and privacy. | Test error, robustness, safe stop, and recovery. |
| Govern | Who decides, intervenes, and remains accountable? | Set authorship, access, appeal, and data rules. | Set supervision, operating limits, and override. |
Table 1 | The five-stage progression connects curriculum questions to evidence for Digital AI and Physical AI.
Age-Appropriate Practice
At foundational levels, learners can identify inputs, decisions, and actions in familiar systems; compare a rule with a learned pattern; and use teacher-mediated demonstrations or simulations that avoid unnecessary personal data. The goal is conceptual language, cause and effect, and safe participation.
At secondary level, learners can audit a generated explanation against sources, document corrections, and reproduce the reasoning without AI. In a robotics task, they can program a simulated or low-power rover, model motion, vary sensor noise, measure error, and implement a safe stop. A fixed-rule controller should be labeled automation unless learning or adaptive inference is genuinely present.
At university and professional levels, students can integrate perception, estimation, planning, control, and governance; compare simulation with hardware; perform failure injection; document operating limits; assess data and cybersecurity risks; and construct a safety or assurance argument for a defined context.
Assessment Must Follow the Claim
A polished AI-assisted product is not sufficient evidence of learning. Assessment should distinguish at least four layers: the quality of the submitted output; the learner's reasoning and verification record; independent performance without assistance; and, for Physical AI, the observed behavior of the complete system under normal and adverse conditions.
Useful evidence includes source annotations, model cards, assumption logs, prediction-before-assistance tasks, oral defense, delayed retrieval, transfer to a new problem, error measurements, uncertainty estimates, failure tests, safe-stop demonstrations, and a governance rationale. The choice must match the learning objective. Requiring every measure in every lesson would create unnecessary burden.
Teachers and Institutions Retain Authority
Teachers remain responsible for pedagogy, participation, safety, and care. They decide when AI assistance supports the objective and when it replaces the thinking to be learned. They provide accessible alternatives, protect learners from compelled data disclosure, supervise physical interaction, and intervene when a system exceeds its limits.
Institutions should define approved use cases, age and access rules, privacy and procurement criteria, human-override procedures, incident reporting, maintenance responsibilities, and evidence requirements. UNESCO's guidance calls for human-centered, age-appropriate validation and protection of data privacy.6 NIST's AI Risk Management Framework offers the complementary functions Govern, Map, Measure, and Manage for lifecycle risk work.7
A human-centered system does not remove technology from education. It places technology inside an accountable educational purpose.
Conclusion
Digital AI changes how learners produce and evaluate information. Physical AI changes how they connect information with bodies, environments, and consequences. Education must prepare students to understand both, model their assumptions, build within boundaries, evaluate evidence, and govern their use.
This progression supports research-driven STEAM education because it joins scientific reasoning, mathematical modeling, technological design, ethical judgment, and human responsibility. Through HERO Science and Technology and www.alberttls.us, this work contributes to the development of intelligible, testable, inclusive, and human-centered learning systems.
Which stage is currently weakest in your curriculum: understanding, modeling, building, evaluating, or governing?
References
1. Miao, F., Shiohira, K., & Lao, N. (2024). AI competency framework for students. UNESCO. Source
2. Miao, F., & Cukurova, M. (2024). AI competency framework for teachers. UNESCO. Source
3. Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. Source
4. Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. Source
5. OECD. (2026). OECD digital education outlook 2026: Exploring effective uses of generative AI in education. OECD Publishing. Source
6. Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. Source
7. Tabassi, E. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). National Institute of Standards and Technology. Source