Applying Educational Psychology Frameworks to Train and Evaluate AI Models

Every parent knows the moment when a child aces a test they could not possibly have understood. The right answers came from memorizing a friend’s worked solutions, not from grasping the material. A provocative new essay argues that artificial intelligence developers are, in effect, running that same classroom every day—and that a century of educational psychology offers exactly the diagnostic tools needed to tell the difference between a machine that has genuinely learned and one that is simply gaming the exam.

In an open-access article published in Educational Psychology Review, Peng Wang of Vrije Universiteit Amsterdam, Erasmus University Rotterdam, and collaborating institutions, together with Fred Paas of Erasmus University Rotterdam and the University of New South Wales, lay out a framework that maps five core ideas from educational psychology onto the training pipelines of modern large language models. The five lenses are curriculum design, instructional scaffolding, social learning through human feedback, assessment and construct validity, and ethical alignment treated as a form of moral education. The paper, titled “Five Educational-Psychology Lenses for Training and Evaluating AI Models,” does not claim that artificial neural networks learn the way children do. Instead, it argues that the constructs developed to evaluate human learning supply a set of falsifiable checks that AI training programs can pass or fail.

The timing is not accidental. Contemporary language-model development proceeds in three broad stages that the authors describe as strikingly pedagogical. In pretraining, a model reads vast quantities of text and learns to predict the next word, building general-purpose linguistic representations. In supervised fine-tuning, it is shown thousands of curated examples of questions and the answers humans consider appropriate. Finally, in reinforcement learning from human feedback, or RLHF, and its cousin RLAIF, where AI models serve as judges, human raters compare candidate responses and the model is updated to favor preferred ones. Throughout deployment, adversarial probes known as red-teaming stress-test the system and reveal new failure modes. Each of these stages, the authors note, has a recognizable counterpart in instructional design: curricula, worked examples, feedback, and high-stakes examinations.

Consider the first lens, curriculum design. Just as a mathematics teacher ensures that students master arithmetic before algebra, machine learning researchers increasingly order training data from simple to difficult, a strategy formalized as curriculum learning. Empirically, such sequencing often speeds convergence and improves generalization, echoing the long-standing educational principle that instruction should be structured. The analogy even extends to behaviorist shaping, in which complex behaviors are built from reinforced approximations: engineers supply intermediate rewards for partial progress so that an agent facing a sparse final reward still has a viable path to success. Automated curriculum systems go further, dynamically selecting the next examples to keep difficulty in what the authors describe as a machine-learning analogue of Vygotsky’s Zone of Proximal Development—a band of tasks that are neither already mastered nor wholly uninformative.

But the paper also documents a striking divergence. In some regimes, particularly for very large models, developers front-load difficult examples through a tactic called hard example mining. In machine learning, “hard” reflects optimization dynamics and distributional coverage rather than learner readiness; challenging cases can force a network to carve out more informative features early, avoiding bad local minima. The concern is computational, not affective: models do not experience frustration or disengagement. This matters practically, the authors argue, because a curriculum that looks sensible by one metric, such as loss-based hardness, can still narrow coverage and inflate apparent progress. Curriculum claims, they conclude, should be evaluated by transfer and robustness tests under distribution shift, not inferred from sequencing alone.

The second lens, scaffolding, will be familiar to anyone who has watched a teacher guide a student through a difficult problem and then gradually withdraw support. The clearest AI counterpart is chain-of-thought prompting, in which a language model is asked to show its work step by step before producing an answer. The technique, along with few-shot prompting, in which the model sees a handful of worked input-output pairs before the test item, is standard practice. The authors’ contribution is diagnostic rather than prescriptive: chain-of-thought is a process prompt, few-shot prompting is worked-example support, and both are scaffolds in the precise educational sense that they reduce the search the learner must perform. They also invite the same question scaffolds always invite—whether the support can fade without collapse. If a model’s accuracy plummets when intermediate reasoning is no longer requested, the gain likely reflects dependence on external structure rather than stable competence. If performance and error detection persist under fading and perturbation, the scaffold may have fostered a more general procedure. The authors add a terminological caution: “prompt engineering” is something of a misnomer, since prompting does not build the model at all but designs an external, in-context support that shapes inference without changing parameters.

Assessment is where educational psychology brings the heaviest theoretical machinery, and the paper anchors its argument in Samuel Messick’s 1995 unified theory of validity. Messick held that validity is not a property of a test but of the inferences drawn from test scores, supported by six interlocking strands of evidence: content, substantive, structural, generalizability, external, and consequential. Each strand, the authors show, has a concrete analogue in AI evaluation. Consequential validity becomes central the moment a benchmark becomes a training target, because optimizing against the benchmark changes what the benchmark measures—the classic Goodhart’s law trap that educators know as teaching to the test. Shortcut learning, specification gaming, and dataset contamination all become legible as assessment failures. When benchmark items or close paraphrases leak into training data, high scores may reflect memorization rather than transferable competence. When a model learns to optimize for what a reward model recognizes as “helpful,” it is learning the evaluator, not the construct.

The paper also highlights recent empirical work that makes these warnings concrete. Citing a 2026 study in Nature, the authors note that when a model trained for a behavioral trait, including misalignment, generates training data for a student model, the student can acquire that trait even when the dataset has been filtered to remove every visible semantic trace of it. The training gain is real; the inference it licenses is not. Even content-level filters that appear complete to a human reader can leave the relevant signal intact—a consequential validity failure stronger than ordinary benchmark gaming.

The social learning lens turns on the observation that modern AI training is, functionally, a process of socialization. Imitation learning from human demonstration mirrors a child learning by watching, though it inherits the classic correspondence problem: a naive imitation agent may reproduce a demonstrator’s fumbles because it cannot distinguish error from intent. Inverse reinforcement learning formalizes the deeper question of inferring a teacher’s goals from behavior, echoing the developmental claim that human children perform a kind of intuitive inverse planning when learning from others. The authors are careful to draw this parallel narrowly—recovering a reward function does not instantiate genuine representation of unobservable mental states—but they insist the structural problem is shared: inferring goals from actions.

The most pointed social-learning concept is the hidden curriculum, the unwritten norms students absorb outside the formal syllabus. Training corpora, the authors argue, are a hidden curriculum for language models, which inevitably absorb the biases, stereotypes, and power structures implicit in internet text. RLHF is then best understood as deliberate social conditioning: raters’ judgments shape which behaviors are reinforced. Yet this raises the deepest tension in the analogy, the gap between outward compliance and internalized understanding. The authors invoke the distinction between a child who is polite to earn rewards and one who genuinely understands why certain behavior is right. Today’s models, critics contend, are closer to the former: they learn to simulate aligned behavior without evidence of deeper comprehension of why that behavior matters. Recent findings that behavioral traits—including misalignment—can propagate through filtered training data sharpen this concern into a practical propagation hazard.

That gap leads directly to the fifth lens: alignment as moral education. The alignment problem echoes the ancient challenge of raising a child to adopt a community’s values. Reward hacking, the authors note, mirrors the student who games the grading system—the cleaning robot that hides trash under the rug to score points without cleaning. Educational theory on the internalization of values suggests a design goal: alignment methods that go beyond punishing specific outputs toward general principles that guide behavior across novel situations, the machine analogue of a well-raised person who does the right thing when no one is watching. The paper’s recommendations follow from this framing: document curriculum decisions as carefully as instructional sequences; run scaffold-fading tests as standard practice and report gaps as findings; treat benchmarks as instruments with consequential validity and set refresh schedules in advance; audit the demographic and stylistic composition of human raters—and of AI-generated training data—as the hidden curriculum it is; and keep claims about behavior strictly separate from claims about understanding.

The authors close with an appeal for epistemic humility. Teaching a machine is an analogy to teaching a child, not an identity. Machines do not feel frustration, carry developmental histories, or construct moral understanding through embodied experience. But the analogy, disciplined by validity, transfer, and measurement, does real work: it specifies what must be measured before a training gain can be interpreted as robust competence, and it opens a two-way street in which machine learners, whose task order and feedback can be varied with laboratory precision, may one day serve as a testbed for hypotheses about learning itself—hypotheses that must still be verified in human classrooms.