From Answer Production to Thought Formation

Review scope and method

This article is a targeted narrative synthesis, not a systematic review. Literature was identified in September 2026 through structured scholarly web searches, Crossref and publisher records, and backward citation chaining. Search concepts combined terms including generative AI, learning, transfer, AI tutor, self-explanation, worked examples, cognitive offloading, and assessment.

Selection prioritized directly relevant human-subject experiments, recent reviews, foundational learning theory, and policy guidance. Peer-reviewed evidence was preferred; a working paper was retained where it directly examined human tutors using AI and is identified as such. Opinion pieces and claims without inspectable research support were excluded. The search was not exhaustive, no formal risk-of-bias assessment was performed, and relevant studies may have been omitted. The synthesis therefore supports hypothesis generation rather than field-wide causal conclusions.

Better output is not the same as more learning.

Research on skill acquisition has long distinguished performance during practice from durable learning revealed after delay or under changed conditions.¹ Generative AI makes this distinction unusually visible because the quality of a jointly produced artifact may increase even when the learner’s independent capability does not.

Selected evidence indicates conditional—not uniform—effects.

Finding: the reviewed studies include benefits, null effects, and a performance–learning dissociation. Synthesis: estimated effects belong to an intervention bundle—model, interface, prompts, sequencing, curriculum, teacher role, learner characteristics, and assessment—not to “AI access” in the abstract. Proposal: studies and schools should report assisted performance separately from later capability and agency transfer.

Useful difficulty must be calibrated, not romanticized.

Self-explanation can support conceptual learning and transfer,⁸ and constructive activity can produce more learning than passive reception.⁹ Yet requiring every learner to solve a full problem before receiving help would ignore the assistance dilemma.¹⁰ Novices facing high-complexity material may learn efficiently from worked examples,¹¹ whereas prior generation can help under conditions that make failure productive rather than merely frustrating.¹²

The appropriate initial contribution may therefore be a prediction, subgoal, partial representation, confidence judgment, or explanation of one worked step. Support should become more explicit after diagnostic evidence of a prerequisite gap, then fade after the learner reconstructs the repaired reasoning. Cognitive offloading is likewise neither inherently harmful nor beneficial: it changes the operations performed internally and should be evaluated against the learning target.¹³

Relation to existing frameworks

EFFORT-AI already proposes an effort-preserving six-phase architecture—elicit, formulate, feedback, organize, reflect, and transfer—with load-adaptive support and accountability.¹⁴ The present paper does not claim to rediscover that sequence. Its narrower contribution is an allocation lens that can be applied within EFFORT-AI or other pedagogies.

Cognitive Responsibility and Dual Transfer

1. 01
   Specify target cognition

   Name the exact operation the learner should acquire: representation, inference, explanation, verification, judgment, or another defined capability.

2. 02
   Allocate responsibility

   State what remains learner-owned, what AI may support or execute, and what requires teacher judgment.

3. 03
   Calibrate assistance

   Use prior knowledge, task complexity, confidence, and observed errors to choose a micro-attempt, hint, worked fragment, or explicit explanation.

4. 04
   Reconstruct and verify

   Ask the learner to rebuild the reasoning, diagnose the original error, and verify consequential claims with appropriate sources or ground truth.

5. 05
   Fade support

   Reduce directiveness once the learner shows a usable representation and can explain the repaired step.

6. 06
   Test dual transfer

   Measure both independent application and the capacity to use or challenge AI responsibly in a novel, ill-structured task.

A prompt history is not proof of cognition or authorship. Process data should be minimized and used formatively, then triangulated with explanation, delayed performance, and transfer. Learners should have clear retention rules and, where feasible, a non-AI route.

Operational case: proving π > 3.05

This illustration is a proposed instructional pathway, not evidence that the model works. The learning target is to connect the definition of π with geometric bounding and a justified inequality.

1. 01
   Diagnostic contribution

   The learner defines π and predicts how an inscribed regular polygon could produce a lower bound. A novice may label a diagram or explain one worked step rather than attempt a full proof.

2. 02
   Adaptive hint ladder

   The AI moves from a question about perimeter, to a diagrammatic cue, to a partially worked inequality only when the learner’s response shows the need.

3. 03
   Reconstruction

   The learner restates why the polygon’s perimeter is less than the circumference and identifies which inequality establishes the bound.

4. 04
   Verification

   Each geometric and numerical step is checked; the AI is also given one deliberately flawed step for the learner to diagnose.

5. 05
   Capability transfer

   After delay and without AI, the learner proves a different lower bound or adapts the method using another polygon.

6. 06
   Agency transfer

   With AI available, the learner compares two proposed proof strategies, rejects an invalid one, and justifies how the tool was used.

Four falsifiable hypotheses

H1 — Assistance timing. With instructional content and time held constant, contingent assistance following a bounded learner-generated response is hypothesized to produce higher delayed unaided retention and transfer than answer-first assistance, even if immediate assisted performance is equal or lower.

H2 — Expertise × complexity. Prior knowledge and task complexity are hypothesized to moderate the effect: novices facing high-complexity tasks should benefit more from a micro-attempt followed by worked fragments, while more knowledgeable learners should benefit more from sparse, progressively faded hints.

H3 — Reconstruction mechanism. Randomly requiring learners to reconstruct reasoning and diagnose an initial error after AI feedback is hypothesized to improve delayed transfer relative to feedback alone; reconstruction quality is a proposed mediator.

H4 — Dual transfer and calibration. Verification prompts plus faded support are hypothesized to improve both delayed unaided transfer and novel AI-augmented performance, including discrimination between correct and deliberately flawed AI responses.

A credible first test would preregister a factorial experiment crossing answer-first versus contingent assistance with reconstruction present versus absent, stratify by prior knowledge, fix model version and prompts, and measure immediate performance, delayed transfer, AI-augmented transfer, confidence calibration, cognitive load, motivation, and time.

Assessment must not become surveillance.

Restricting AI remains defensible in high-stakes assessment, privacy-sensitive work, some developmental contexts, and tasks where the tool would perform the precise operation being assessed. Elsewhere, responsible use requires attention to unequal access, language and cultural bias, hallucinated sources, age-appropriate consent, vendor retention, accessibility, teacher workload, and the availability of an equivalent non-AI route. UNESCO similarly frames educational AI as a human-centred design and governance problem.¹⁵

Reasoning traces should be concise, purpose-limited, and retained only as long as necessary. They are candidate indicators, not direct windows into thought. Oral explanation, multimodal response, and delayed transfer may provide fairer evidence for learners whose language, disability, or interaction style makes verbose prompting a poor proxy for understanding.

Limitations and conclusion

This selective review may omit relevant evidence and includes studies that differ in subject, age, geography, duration, AI role, and outcome horizon. The field changes rapidly, some results depend on specific model versions, and one central human-tutor study remains a working paper. The proposed model has not been tested as a unified intervention. Its terms and measures may require substantial adaptation for early-primary, multilingual, low-resource, accessibility, creative, and workplace-learning settings.

The resulting claim is deliberately conditional: generative AI may support durable learning when an instructional system makes cognitive responsibility explicit, calibrates assistance rather than maximizing it, returns consequential reasoning to the learner, and evaluates both unaided capability and responsible AI-augmented agency. Whether that claim holds is an empirical question.

Selected literature

1. Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10, 176–199.
2. Barnett, S. M., & Ceci, S. J. (2002). When and where do we apply what we learn? A taxonomy for far transfer. Psychological Bulletin, 128, 612–637.
3. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, O., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122, e2422633122.
4. Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (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.
5. Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. (2024). Tutor CoPilot: A human–AI approach for scaling real-time expertise. EdWorkingPaper 24-1054 / arXiv:2410.03017.
6. Bassner, P., Lenk-Ostendorf, B., Beinstingel, R., Wasner, T., & Krusche, S. (2026). Less stress, better scores, same learning: The dissociation of performance and learning in AI-supported programming education. Computers and Education: Artificial Intelligence, 10, 100537.
7. Deng, R., Jiang, M., Yu, X., Lu, Y., & Liu, S. (2025). Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies. Computers & Education, 227, 105224.
8. Chi, M. T. H., Bassok, M., Lewis, M. W., Reimann, P., & Glaser, R. (1989). Self-explanations: How students study and use examples in learning to solve problems. Cognitive Science, 13, 145–182.
9. Chi, M. T. H., & Wylie, R. (2014). The ICAP framework: Linking cognitive engagement to active learning outcomes. Educational Psychologist, 49, 219–243.
10. Koedinger, K. R., & Aleven, V. (2007). Exploring the assistance dilemma in experiments with cognitive tutors. Educational Psychology Review, 19, 239–264.
11. Atkinson, R. K., Derry, S. J., Renkl, A., & Wortham, D. (2000). Learning from examples: Instructional principles from the worked examples research. Review of Educational Research, 70, 181–214.
12. Kapur, M. (2008). Productive failure. Cognition and Instruction, 26, 379–424.
13. Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20, 676–688.
14. Liubertaitė, A. D. (2026). EFFORT-AI: An effort-preserving module for cognitively aligned human–AI learning. Frontiers in Education, 11, 1849821.
15. Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO.

Research integrity and disclosure

Status: AOHA Working Paper v0.2. Not externally peer reviewed. No original human-participant experiment is reported. Funding: No funding declaration has been made for this version. Potential conflict: The author is the founder of AOHA LAB, which has a commercial interest in AI-related research, development, and education services. Data and code: Not applicable. Ethics approval: Not applicable to this literature-based paper.

AI-use disclosure: The concept and direction were provided by Shinichi Asai. Literature discovery, drafting, and editing were conducted with OpenAI Codex under the author’s direction. The manuscript underwent an AI-assisted internal critical review using three separately instructed reviewer perspectives covering learning science, human–AI interaction, and cross-disciplinary methodology and academic editing. The author remains responsible for final verification, interpretation, and publication. This process does not constitute independent human peer review.