Deceptive Overgeneralization: When Adaptive Learning Enables Systematic Misapplication

Synthesis: An, McLaren, and Stamper (2026) introduce and empirically test deceptive overgeneralization — a learning phenomenon in which learners acquire a skill with an incomplete subset of its necessary conditions, omitting a critical application constraint, yet still produce correct actions. Because observed correctness looks like mastery, Adaptive Learning systems that infer mastery from correct performance risk prematurely stopping practice before learners encounter cases where the action should be withheld, leaving the overgeneralization undetected. Across 11 experiments (N = 192) with Intelligent Tutoring Systems for Riichi Mahjong, learners systematically misapplied learned actions on first "do-not-act" detector items (61.5%–100% across skills and cohorts), far exceeding the error rates predicted by Bayesian Knowledge Tracing. Tailored "do-not-act" practice with feedback that names the missing constraint reduced misapplication to near-floor levels. A secondary analysis of K-12 Decimal Point data shows the framework also accounts for whole-number bias, suggesting it generalizes beyond adaptive systems to traditional instruction.

Key Findings

1. Correctness can mask incomplete conditional understanding. Learners can appear competent during practice while having compiled an overgeneralized production rule — performing the expected action but omitting the "when to withhold it" constraint. This "deceptive" pattern inflates learners' perceived competence and misleads any assessment that infers mastery from observed actions.
2. Deceptive overgeneralization is empirically prevalent. In 11 experiments using ITSs for composite-condition, single-action skills in Riichi Mahjong (score calculation, Pinfu, Kabe, Riichi judgment), learners misapplied the learned action on the first detector item at rates of 61.5%–100% across skills and cohorts — significantly exceeding both the BKT accurate-mastery baseline (12% expected error) and the new-KC baseline (40%) in nearly all comparisons.
3. Correctness-based mastery stopping rules can terminate practice too early. With BKT-based adaptivity (95% mastery threshold), the system stopped assigning practice before learners encountered any "do-not-act" case — so the overgeneralization went undetected in the adaptive round and only surfaced when learners were later given full exposure.
4. Targeted "do-not-act" remediation is highly effective. Short sequences of practice where the correct response is to refrain from the learned action, paired with feedback naming the missing application constraint, reduced misapplication from 78.6%–100% initially to 0.0%–23.1% post-remediation (Cohen's h 1.696–2.441).
5. The mechanism generalizes beyond adaptive learning. A secondary analysis of K-12 Decimal Point decimal-learning data (13 datasets, 2015–2025) showed whole-number bias — recast as a deceptive overgeneralization — persists even in an effective learning game, with 84%–88% of decimal-comparison errors consistent with the overgeneralized "longer-is-larger" rule.

Deceptive overgeneralization as a learning phenomenon

The authors ground the concept in ACT-R production systems and the Knowledge-Learning-Instruction (KLI) framework. A production is an IF(conditions)-THEN(action) pair; knowledge compilation merges chains of productions into efficient "macro-productions" whose condition sides grow, making it more likely some conditions are overlooked. Deceptive overgeneralization occurs when a learner acquires IF a subset of required conditions are met THEN perform action instead of the fully-constrained rule. It is "deceptive" because the learner still takes the correct action — until a scenario violates a missing constraint. The Crossair Flight 498 crash is offered as a domain-agnostic worked example: a commander's compiled attitude-display rule omitted the "Soviet-designed display only" constraint, and over 8,000 hours of experience did not correct it.

This distinguishes deceptive overgeneralization from negative transfer: while negative transfer typically corrects with experience, deceptive overgeneralization may persist because the overgeneralized rule competes with later knowledge during conflict resolution. It also differs from shallow-feature errors: here, all features the learner uses are part of the correct solution path, making the overgeneralization behaviorally indistinguishable from accurate understanding until a withheld-action context appears.

Cross-domain case study: whole-number bias in K-12 decimals

The framework is extended to a traditional, non-adaptive domain. Whole-number bias — students treating the longer decimal as larger because they apply whole-number comparison — is modeled as omitting the application constraint of the "longer-is-larger" production. Because positive whole-number comparisons provide no counterexamples, the rule is strengthened through repeated success; when decimals are introduced, the older overgeneralized production competes with the newer decimal rule during conflict resolution. In Decimal Point assessment data, error rates on the "do-not-act" decimal item fell only modestly after the intervention (53.11% → 47.47% delayed), and 84%–88% of errors were whole-number-bias aligned — suggesting simply teaching the correct procedure may be insufficient; the older, competing production must be explicitly refined. This has implications for K-12 mathematics instruction.

What this means for practice

• Instructional designers. Treat streaks of correct performance as insufficient evidence of conditional mastery: represent each target skill as a production rule, identify the composite-condition, single-action rules most susceptible to overgeneralization, and prioritize the ones where misapplication is most consequential.
• Instructional designers. Build "do-not-act" detector items — scenarios that violate one necessary condition, where refraining is the right answer — and place them before mastery stopping rules can terminate practice.
• Instructional designers. Pair each detector item with Feedback that names the missing application constraint explicitly, so learners refine the rule they already hold rather than only acquiring the new one.
• EdTech designers. Audit Adaptive Learning stopping rules that infer mastery from correctness alone: under BKT-based adaptivity at a 95% mastery threshold, the system ended practice before learners encountered any case where the learned action should not be applied.
• Researchers. Extend the diagnostic to non-adaptive instruction: in K-12 decimals, teaching the correct comparison procedure moved whole-number-bias errors only from 53.11% to 47.47% at the delayed post-test, so the older competing production must be targeted directly.

Limitations

• The empirical work ran in Riichi Mahjong only (11 experiments, N = 192), in a voluntary, interest-driven course with optional study participation, so the misapplication-rate estimates may not hold in other domains or under different recruitment conditions.
• Individual experiments had relatively small samples, which limits precision for estimating effect magnitudes and for testing moderators.
• The decimal evidence is a secondary analysis: the 13 datasets in which 84%–88% of comparison errors aligned with whole-number bias are a theory-grounded exploratory application, not a direct empirical test of deceptive overgeneralization in K–12 mathematics.
• Remediation showed immediate effects only — misapplication fell to 0.0%–23.1% within the ITS — with no delayed retention or transfer measured beyond the remediation sequence, and the fixed item order (detector items always after ample non-diagnostic cases) may differ from organic instructional sequencing.

Citation

An, M., McLaren, B. M., & Stamper, J. (2026). Deceptive overgeneralization: When adaptive learning enables systematic misapplication. Journal of Computer Assisted Learning, 42, e70311.