📄 Full text: Stanford SCALE · local
The central open question in AI education research: Do performance gains with AI tools persist when students must demonstrate understanding independently?^stanford-evidence-base-ai-k12-2026
The Transfer Paradox
Students using AI typically show immediate, measurable gains on the tasks where AI is available. Yet when AI is removed:
- Effects become mixed or negative
- Gains often fail to transfer to unassessed settings
- Students may become dependent on the tool at the expense of independent reasoning
This pattern is the defining risk of integrating AI into learning environments.
Key Studies
| Study | Context | Immediate Effect | Transfer Effect | Mechanism |
|---|---|---|---|---|
| Bastani et al. (2025) | High school math | Higher practice grades | ~17% worse on closed-book finals | General-purpose chatbot did the work |
| Chen et al. (2025) | Programming homework | Higher homework scores | No improvement on unassisted exams | LLM-Tutor solved problems for students |
| Lehmann et al. (2025) | Programming | More topics covered | Harmed understanding; widened gaps | General AI for low-prior learners |
| Stadler et al. (2024) | Academic research | Faster task completion | Lower-quality reasoning vs. search | Reduced cognitive engagement |
| Kosmyna et al. (2025) | Essay writing | Higher essay quality | 83% failed to recall their own quotes | Outsourced authorship |
All five studies show a negative or null transfer pattern when general-purpose AI is the intervention.
Misalignment Amplification
Hardy & Kim (2026) demonstrate that shared pretraining biases across LLMs produce behaviors poorly aligned with teaching quality and negatively aligned with student outcomes. Multi-model ensembles even exacerbate misalignment. See educational-llm-alignment.
Mechanisms Undermining Transfer
1. Metacognitive Displacement
AI completing reasoning reduces opportunities for students to monitor their own understanding and select strategies. Students who used AI were less able to explain their answers when queried.^stanford-evidence-base-ai-k12-20262. Germane Load Suppression
General-purpose AI reduces not just extraneous (distracting) cognitive load but also germane load — the productive mental effort that encodes durable knowledge. Easier practice feels better but stores weaker traces. See tutoring-specific-vs-general-ai.3. Over-Reliance / Expertise Reversal
Novices given answers do not build schemas. General AI provides answers; effective tutoring provides structured guidance. When novices are given expert-level shortcuts, learning is disrupted.^stanford-evidence-base-ai-k12-20264. Tool-Dependent Performance
Students may optimize for the specific affordances of the AI tool (e.g., prompt engineering, reliance on generated code structure) rather than building domain generalization.Conditions Supporting Positive Transfer
The limited evidence suggests transfer is possible when:
- Pedagogical guardrails are present — step-by-step hints, misconception targeting, Socratic questioning (Bastani et al., 2025 tutoring variant)
- Traditional strategies are preserved — note-taking paired with AI use improved retention (Kreijkes et al., 2026)
- AI is used for formative, not summative, practice — scaffolding during learning, not during assessment
- Learner expertise is calibrated — the tool adapts support to readiness rather than defaulting to full assistance
Unanswered Questions
1. Time scale: Does transfer improve over weeks/months of use, or does dependence deepen? 2. Domain differences: Is transfer better in well-structured domains (math) vs. ill-structured domains (writing)? 3. Individual differences: Do high-prior-knowledge students suffer less transfer loss than novices? 4. Skill remediation: Can explicit "AI-off" practice sessions reverse tool dependence?
Implications
For educators and tool designers, the provisional rule is:
If students can complete a task with AI, design expectations so they must also demonstrate the task without it.
This may mean:
- Closed-book assessments after AI-assisted practice
- Reflective prompts requiring students to explain AI-generated content in their own words
- Strict separation between "practice with AI" and "demonstrate mastery" phases
Related Pages
- ai-assistance-reduces-persistence: Causal evidence (N=1,222) that brief AI assistance reduces persistence and impairs unassisted performance — rapid emergence of over-reliance effects
- cognitive-offloading-speedup-illusion — Speedup illusion may reduce deliberate practice and learning transfer
- structured-llm-feedback-programming — The Effects of Structured LLM-Generated Feedback on Programming Assignment Performance
- academiclaw-student-agent-benchmark — AcademiClaw: 55% ceiling on frontier models raises transfer questions for AI tutoring systems
- genai-meta-analysis-programming-learning — Highest-level evidence on whether AI gains persist when tools removed
- sequenced-ai-feedback-learning — Cao et al. RCT: learning outcome disparity between feedback conditions raises transfer and retention questions
- awareness-technological-isomorphism — Technological Isomorphism bridges elementary math and AI comprehension
- instructional-guidance-genai-learning — Structured prompting improves AI-assisted learning outcomes- ai-partner-science-epistemic-vigilance -- Epistemic vigilance determines whether AI augments or undermines learning; uniform AI integration risks widening achievement gaps
- self-efficacy-tutoring-learning — Self-Efficacy and Favorability Shape Learning from Tutoring
- adaptive-pretesting-retention -- Adaptive AI pretesting boosts initial learning, but long-term retention over 7 weeks depends on subsequent retrieval practice — learner-directed AI study underperforms structured retrieval.
- llm-item-difficulty-prediction — Model reasoning as proxy for human cognitive processes
- prompt-based-programming-lesson — prompt-based programming teaching intervention
- ai-availability-student-motivation — AI availability and student motivation study
- critical-engagement-code-completion — measuring critical engagement with AI code completion
- hybrid-human-ai-tutoring-differentiated — MAP growth as learning transfer measure
- genai-performance-vs-learning — Distinguishing performance gains from actual learning