Custom Prompt AI_tutoring_cognitive_offloading_literature_map

Open in new tab ↗

Structured AI Tutoring vs. Cognitive
Offloading in Undergraduate STEM
A citation-lineage and evidence map centered on Kestin et al. (2025)
Research question
Research question: Does using AI tools for problem-solving improve undergraduate STEM students'
learning and long-term retention, or does it create a cognitive-offloading effect that reduces
skill-building - and does the outcome depend on how the AI is used, such as answer-giving versus
Socratic tutoring?
Conceptual illustration: the same AI interface can sit on a learning path that preserves cognitive work or a shortcut path that
displaces it; the empirical literature increasingly suggests the difference is pedagogical design and user behavior, not AI access
alone.
Seed paper 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.
doi:10.1038/s41598-025-97652-6
Scope 20 related papers selected for conceptual lineage and research-question relevance: 10 confirmed antecedents cited by
the seed + 10 recent descendants/adjacent studies. The seed itself is shown separately and is not counted in the 20-paper
limit.
Literature
check
Recent-source verification performed through 18 August 2026. Direct citation/extension relationships are labeled
separately from inferred thematic relationships.

AI tutoring, cognitive offloading, and undergraduate STEM learning
2Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
Executive synthesis
Bottom line
The evidence does not support a simple "AI helps" or "AI harms" conclusion. Structured AI that keeps
the learner cognitively active - prompting attempts, providing targeted explanations or hints, asking
Socratic questions, and withholding complete solutions when appropriate - can improve learning. By
contrast, open-ended answer-giving can raise assisted task performance while reducing independent
performance, persistence, study time, or retention once the tool is removed.
What the seed establishes. Kestin et al. randomized 194 undergraduates in introductory physics to a
carefully engineered GPT-4 tutor versus a strong active-learning classroom condition. The AI condition
produced substantially higher immediate post-test performance in less median time and higher
engagement/motivation, demonstrating that an AI tutor can outperform current best-practice instruction
when its interaction is tightly scaffolded.
What the seed does not establish. The intervention spans two lessons and measures immediate post-tests;
the authors explicitly identify retention, longer course integration, higher-order synthesis, and the decision
of when to give an answer versus guide reflection as open research directions. That limitation is decisive for
your question.
What subsequent evidence adds. A direct extension (Miller, 2026) shows that pedagogical system prompts
change engagement and cognitive load. Later/parallel experiments sharpen the mechanism: guarded or
Socratic tutoring preserves cognitive work, while unrestricted answer access can create dependence.
Contractor and Reyes (2026) add one-week undergraduate evidence in which explanation-seeking
"augmentation" users retain gains better than text-generating "automation" users.
Where cognitive offloading becomes visible. Bastani et al. (2025) observe a reversal from higher assisted
performance to lower unassisted performance under unrestricted GPT-4; Liu et al. (2026) find reduced
independent performance and persistence after brief AI exposure; Rismanchian et al. (2026) report less time
on AI-susceptible college-math problems and weaker later proctored retention in a large longitudinal
dataset.
Current answer to the research question. The most defensible conclusion is conditional: AI can improve
learning when it behaves like a tutor and is used as an augmentation; it can reduce skill-building when it
behaves like an answer engine and is used as a replacement for reasoning. What remains under-tested is
the exact combination you care most about: a long-duration, undergraduate STEM RCT that randomizes
answer-giving versus Socratic scaffolding and measures delayed, unaided transfer and retention.

AI tutoring, cognitive offloading, and undergraduate STEM learning
3Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
Relationship map
FOUNDATIONS SEED RECENT EVIDENCE
Tutoring
Bloom
Scaffolding
Wood et al.
Cognitive load
Sweller
Feedback
Shute
Active learning
Hake/Freeman
Early LLM
Kumar/Krupp/Forero
Kestin et al. (2025)
Structured GPT-4 tutor
Undergraduate physics RCT
Short-term learning: strong +
Scaffolded / Socratic / augmentation
Answer-giving / offloading risk
Cross-cutting evaluation / synthesis
Miller 2026 LearnLM 2025
Degen 2025 Sun & Liu 2026
Contractor & Reyes 2026
Bastani 2025 Liu et al. 2026
Rismanchian et al. 2026
SafeTutors 2026 Wang et al. 2026
Solid arrows indicate a verified citation or explicit extension relationship. Dashed arrows indicate thematic adjacency selected
for the research question; those papers were not confirmed as direct citations of the seed in the accessible full text searched.

AI tutoring, cognitive offloading, and undergraduate STEM learning
4Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
1. The seed paper as the pivot
Seed: 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. doi:10.1038/s41598-025-97652-6.
Dimension What the seed paper actually shows Implication for your question
Population /
context
194 eligible students in Harvard Physical Sciences 2, an
introductory physics course for life sciences; crossover design
across two lessons.
Strong undergraduate STEM relevance, but one
institution and a narrow learning window.
Comparator Research-based in-class active learning, not passive lecture. The effect is not merely "AI beats lecture"; the
tutor outperformed an already effective
pedagogy.
AI design GPT-4 tutor with instructor-crafted prompts, sequential
scaffolding, brief responses, step-by-step reference solutions,
targeted feedback, and self-pacing.
The intervention is closer to an engineered
Socratic/scaffolded tutor than to unrestricted
ChatGPT.
Outcome Immediate post-test learning; median AI post-score 4.5 versus
3.5 for active learning, with large adjusted effects; AI students
also reported greater engagement and motivation.
Strong evidence for immediate acquisition and
user experience, not durable retention.
Critical gap The authors call for studies of spacing/retention, extended
course use, higher-order synthesis, and decisions about when
the tutor should openly provide answers versus guide
reflection.
Your question is a direct continuation of the
seed paper's stated limitations and future-work
agenda.
Interpretive caution
Because the seed tutor explicitly discourages giving away the full solution and embeds scaffolding, it
should not be generalized to "ChatGPT use" in the abstract. The treatment is a particular pedagogy
delivered through an LLM. That distinction explains why later studies of unrestricted answer access
can produce opposite effects without actually contradicting Kestin et al.

AI tutoring, cognitive offloading, and undergraduate STEM learning
5Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
2. Foundational papers the seed builds on
All ten papers below are confirmed antecedents: they appear in the seed paper's reference list and are used
to motivate its tutoring design, active-learning comparator, learning/perception measures, or contrast with
unguided LLM use.
CONFIRMED
ANTECEDENT
1. Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group
instruction as effective as one-to-one tutoring. Educational Researcher, 13(6),
4-16. doi:10.3102/0013189X013006004
Relationship to seed: Directly cited by Kestin et al. as the benchmark for the effectiveness - and
scalability problem - of expert one-to-one tutoring.
One-sentence summary: Bloom framed individualized tutoring as a high-performing
instructional benchmark, motivating the seed paper's attempt to reproduce tutor-like
personalization at scale with AI.
Research-question fit: Foundational to the tutoring side of the question, but it predates
generative AI and does not address offloading.
CONFIRMED
ANTECEDENT
2. Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in
problem-solving. Journal of Child Psychology and Psychiatry, 17(2), 89-100.
Relationship to seed: Directly cited by the seed paper as the basis for scaffolding content and
sequencing support during problem solving.
One-sentence summary: This classic tutoring study formalized scaffolding as contingent
support that helps learners perform beyond current capability while preserving learner
participation in the solution process.
Research-question fit: Provides the clearest pre-AI rationale for Socratic or hint-based
assistance rather than simply supplying answers.
CONFIRMED
ANTECEDENT
3. Sweller, J. (2011). Cognitive load theory. In J. P. Mestre & B. H. Ross (Eds.),
The Psychology of Learning and Motivation: Cognition in Education (pp. 37-76).
Elsevier Academic Press.
Relationship to seed: Directly cited and explicitly implemented in the AI tutor design through
brief, targeted exchanges intended to manage working-memory demands.
One-sentence summary: Cognitive load theory argues that instruction should reduce
extraneous load and manage task complexity so limited working memory can be devoted to
schema construction.
Research-question fit: Central to distinguishing productive offloading of irrelevant burden from
harmful offloading of the reasoning practice itself.
CONFIRMED
ANTECEDENT
4. Shute, V. J. (2008). Focus on formative feedback. Review of Educational
Research, 78(1), 153-189. doi:10.3102/0034654307313795
Relationship to seed: Directly cited by Kestin et al. to justify targeted, timely feedback as a core
design principle of the tutor.
One-sentence summary: Shute synthesizes evidence that formative feedback is most useful
when it is timely, specific, non-evaluative, and oriented toward closing the gap between current
and desired performance.
Research-question fit: Supports AI as a feedback partner that responds to student thinking
rather than a solution generator.

AI tutoring, cognitive offloading, and undergraduate STEM learning
6Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
CONFIRMED
ANTECEDENT
5. Hake, R. R. (1998). Interactive-engagement versus traditional methods: A
six-thousand-student survey of mechanics test data for introductory physics
courses. American Journal of Physics, 66(1), 64-74. doi:10.1119/1.18809
Relationship to seed: Directly cited as evidence that interactive engagement outperforms
traditional lecture in introductory physics, helping define the strong control condition the AI
tutor must beat.
One-sentence summary: Across thousands of introductory physics students,
interactive-engagement methods produced substantially larger conceptual gains than
traditional instruction.
Research-question fit: Establishes active problem solving as the benchmark against which
AI-supported learning should be judged.
CONFIRMED
ANTECEDENT
6. Freeman, S., Eddy, S. L., McDonough, M., Smith, M. K., Okoroafor, N., Jordt,
H., & Wenderoth, M. P. (2014). Active learning increases student performance in
science, engineering, and mathematics. Proceedings of the National Academy of
Sciences, 111(23), 8410-8415. doi:10.1073/pnas.1319030111
Relationship to seed: Directly cited to establish active learning as an evidence-based STEM
standard rather than a weak lecture-only comparator.
One-sentence summary: This meta-analysis of 225 STEM studies found that active learning
improves exam performance and lowers failure rates relative to traditional lecturing.
Research-question fit: Makes the seed RCT more consequential: its comparison is against an
established effective pedagogy, not passive instruction.
CONFIRMED
ANTECEDENT
7. Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. (2019).
Measuring actual learning versus feeling of learning in response to being
actively engaged in the classroom. Proceedings of the National Academy of
Sciences, 116(39), 19251-19257. doi:10.1073/pnas.1821936116
Relationship to seed: Directly cited and authored partly by the same team; it motivates the
seed paper's separation of objective learning gains from students' subjective experience.
One-sentence summary: Students in active classrooms learned more even when they felt they
learned less, demonstrating that perceived fluency or enjoyment can diverge from actual
learning.
Research-question fit: Important for AI studies because a frictionless, confidence-inducing tool
can feel helpful while reducing independent skill formation.
CONFIRMED
ANTECEDENT
8. Kumar, H., Rothschild, D. M., Goldstein, D. G., & Hofman, J. M. (2023). Math
education with large language models: Peril or promise? SSRN Electronic
Journal. doi:10.2139/ssrn.4641653
Relationship to seed: Directly cited by Kestin et al. as early mixed evidence on LLM-supported
learning and as a precursor to testing how instructional structure changes outcomes.
One-sentence summary: In a preregistered experiment with about 1,200 participants, LLM
explanations improved later math performance over answers alone, with the largest gains
when learners attempted problems before consulting the explanation.
Research-question fit: One of the most direct antecedents for the answer-giving versus
explanation/attempt-first dimension in the research question.

AI tutoring, cognitive offloading, and undergraduate STEM learning
7Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
CONFIRMED
ANTECEDENT
9. Krupp, L., Steinert, S., Kiefer-Emmanouilidis, M., Avila, K. E., Lukowicz, P.,
Kuhn, J., Küchemann, S., & Karolus, J. (2024). Unreflected acceptance -
Investigating the negative consequences of ChatGPT-assisted problem solving in
physics education. In HHAI 2024: Hybrid Human AI Systems for the Social Good
(pp. 199-212). doi:10.3233/FAIA240195
Relationship to seed: Directly contrasted with the seed results as evidence that unrestricted
ChatGPT use can elicit shallow interaction and uncritical acceptance.
One-sentence summary: Physics-background participants frequently over-trusted incorrect
ChatGPT-supported solutions and used copy-and-paste prompting far more than search-engine
users, indicating limited reflection during problem solving.
Research-question fit: Directly supports the cognitive-offloading concern in higher-education
STEM problem solving.
CONFIRMED
ANTECEDENT
10. Forero, M. G., & Herrera-Suárez, H. J. (2023). ChatGPT in the classroom:
Boon or bane for physics students' academic performance? arXiv preprint
arXiv:2312.02422. arXiv:2312.02422
Relationship to seed: Directly contrasted by Kestin et al. as a prior physics study in which
relatively unstructured ChatGPT integration was associated with poorer academic performance.
One-sentence summary: A second-semester engineering-physics cohort encouraged to use
ChatGPT earned lower exam grades than a prior control cohort, while many students
simultaneously reported convenience and concerns about reduced critical or independent
thinking.
Research-question fit: A close pre-seed warning that unguided AI access may harm learning
even when students perceive it as useful.

AI tutoring, cognitive offloading, and undergraduate STEM learning
8Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
3. Recent papers that extend, cite, or directly test
the seed paper's thesis
The first five are verified citation-lineage papers. The final five are deliberately included as adjacent
evidence because they address the exact causal mechanisms in your question even though I did not confirm
a direct citation to Kestin et al. in the accessible full text.
CONFIRMED EXTENSION 1. Miller, K. (2026). Prompt Matters: How Pedagogical Engineering Shapes
Behavior and Engagement with AI Tutors. Technology, Knowledge and Learning.
doi:10.1007/s10758-026-09983-6
Relationship to seed: The paper explicitly states that it uses the same framework as Kestin et
al. and acknowledges the work as an extension of the earlier experiment.
One-sentence summary: In a randomized study (N=125), an enhanced tutor prompt
incorporating active learning, cognitive-load management, and growth-mindset support
produced more student-AI interaction, higher engagement and motivation, and less cognitive
overload than a minimal prompt, while both groups learned.
Research-question fit: Directly isolates prompt-level pedagogy as a mechanism, although it
does not establish long-term retention.
CONFIRMED
DESCENDANT
2. LearnLM Team, Google & Eedi. (2025). AI tutoring can safely and effectively
support students: An exploratory RCT in UK classrooms. arXiv preprint
arXiv:2512.23633. arXiv:2512.23633
Relationship to seed: The paper directly cites Kestin et al. in its review of evidence for effective
generative-AI tutoring.
One-sentence summary: In an exploratory RCT with 165 secondary-school students,
human-supervised LearnLM tutoring matched human tutoring and improved success on novel
subsequent-topic problems by 5.5 percentage points, with tutors specifically praising its
Socratic questions.
Research-question fit: Strong evidence for guided dialogue and transfer, but the population is
K-12 rather than undergraduates.
CONFIRMED
DESCENDANT
3. Hazra, R., Ghuku, B., Marchenko, I., Tokarieva, Y., Layek, S., Banerjee, S.,
Stoyanovich, J., & Pechenizkiy, M. (2026). SafeTutors: Benchmarking
Pedagogical Safety in AI Tutoring Systems. arXiv preprint arXiv:2603.17373.
arXiv:2603.17373
Relationship to seed: The benchmark paper directly cites Kestin et al. while reframing
educational AI safety around preserving learning effort and scaffolding, not merely avoiding
harmful content.
One-sentence summary: Across mathematics, physics, and chemistry, SafeTutors finds
widespread pedagogical harms such as answer over-disclosure and collapsed scaffolding, with
multi-turn failure rates rising sharply in its evaluation scenarios.
Research-question fit: Does not measure student learning directly, but operationalizes the
exact failure mode behind answer-giving and cognitive offloading.

AI tutoring, cognitive offloading, and undergraduate STEM learning
9Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
CONFIRMED
PREPRINT-LINEAGE
4. Degen, P.-B., & Asanov, I. (2025). Beyond Automation: Socratic AI, Epistemic
Agency, and the Implications of the Emergence of Orchestrated Multi-Agent
Learning Architectures. arXiv preprint arXiv:2508.05116. arXiv:2508.05116
Relationship to seed: The paper cites the 2024 Kestin et al. preprint version and extends the
design logic toward explicitly Socratic, dialogue-driven AI support.
One-sentence summary: In a controlled experiment with 65 pre-service teacher students, a
Socratic tutor was rated as providing more support for critical, independent, and reflective
thinking than an uninstructed chatbot.
Research-question fit: Directly tests Socratic versus generic chatbot interaction, but outcomes
are largely self-reported rather than delayed performance measures.
CONFIRMED
DESCENDANT
5. Sun, Y., & Liu, F. (2026). The impact of an AI Digital Teacher on human-AI
collaborative learning in higher education. Smart Learning Environments, 13,
30. doi:10.1186/s40561-026-00454-0
Relationship to seed: The authors directly cite Kestin et al. as evidence of strong gains in
structured physics and frame their study as testing whether such benefits generalize to a
complex non-STEM domain.
One-sentence summary: A semester-long RCT with 84 undergraduates found that adding a
bespoke AI teacher to traditional literature instruction improved objective tests and analytical
essays while reducing extraneous and increasing germane cognitive load.
Research-question fit: Useful evidence for duration and cognitive-load mechanisms, but not a
STEM problem-solving test.
INFERRED / PARALLEL 6. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025).
Generative AI without guardrails can harm learning: Evidence from high school
mathematics. Proceedings of the National Academy of Sciences, 122(26),
e2422633122. doi:10.1073/pnas.2422633122
Relationship to seed: I did not confirm a direct citation to the seed paper in accessible full text;
it is included because it is the closest causal test of unrestricted answer-giving versus a
learning-protective tutor design.
One-sentence summary: In a field experiment with nearly 1,000 high-school math students,
unrestricted GPT-4 raised assisted performance but led to 17% lower unassisted grades after
access was removed, whereas a guarded GPT Tutor largely mitigated that learning penalty.
Research-question fit: Probably the strongest current causal evidence that design and answer
access determine whether AI becomes a crutch.
INFERRED / ADJACENT 7. Contractor, Z., & Reyes, G. (2026). Experimental Evidence on the Learning
Impact of Generative AI. arXiv preprint arXiv:2607.08849. arXiv:2607.08849
Relationship to seed: No direct citation to Kestin et al. was confirmed in the accessible version;
it is selected because it directly fills the seed paper's retention and usage-mode gap.
One-sentence summary: In a randomized experiment with 211 undergraduates, AI access
improved unaided knowledge immediately and about one week later, but delayed gains were
concentrated among augmentation users who sought explanations, while automation users'
short-run gains vanished once AI was removed.
Research-question fit: The clearest undergraduate evidence so far that augmentation versus
automation predicts durable benefit, though the task is not a standard STEM course problem
set.

AI tutoring, cognitive offloading, and undergraduate STEM learning
10Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
INFERRED / ADJACENT 8. Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI
Assistance Reduces Persistence and Hurts Independent Performance. arXiv
preprint arXiv:2604.04721. arXiv:2604.04721
Relationship to seed: The accessible full text contains no Kestin citation; it is included as
mechanism-level causal evidence that immediate AI assistance can undermine independent
skill and persistence.
One-sentence summary: Across randomized experiments totaling 1,222 participants, AI
assistance improved performance while available but reduced later unassisted performance
and increased giving up after only brief exposure, especially when users obtained direct
answers.
Research-question fit: Strong causal evidence for a cognitive-offloading/persistence
mechanism, with less direct correspondence to authentic undergraduate STEM courses.
INFERRED / ADJACENT 9. Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026).
Faster Completion, Less Learning: Generative AI Reduced Study Time on Math
Problems and the Knowledge They Build. arXiv preprint arXiv:2605.21629.
arXiv:2605.21629
Relationship to seed: The paper does not cite Kestin et al. in the accessible version; it is
included because it tests long-run college-math behavior and delayed proctored retention at
scale.
One-sentence summary: Using 3.2 million ALEKS learning interactions over a decade, the
authors estimate a 26.9% post-ChatGPT decline in time spent on AI-susceptible college math
problems and a 25% cumulative decline in the odds of correct responses on later proctored
retention items.
Research-question fit: Highly relevant to durable STEM learning, but the quasi-experimental
design infers AI use from task susceptibility rather than observing prompts directly.
INFERRED / SYNTHESIS 10. Wang, G., Wang, W., Yang, D., & Ren, J. (2026). Generative AI, Cognitive
Offloading, and Learner Agency in Higher Education: A Scoping Review.
Behavioral Sciences, 16(7), 1150. doi:10.3390/bs16071150
Relationship to seed: I did not find a direct Kestin citation in the accessible article; it is included
as a current higher-education synthesis of the augmentation-versus-replacement mechanism
central to the seed paper's design logic.
One-sentence summary: Synthesizing 123 higher-education studies, the review concludes that
scaffolded, self-regulated, augmentation-oriented AI use is associated with learner agency,
whereas replacement-oriented use clusters with offloading, overreliance, dependence, and
weakened judgment.
Research-question fit: A useful map of the broader literature, but its conclusions are
configurative rather than causal and not STEM-specific.

AI tutoring, cognitive offloading, and undergraduate STEM learning
11Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
4. What the 20-paper map says about your research
question
Sub-question Best current reading Most probative papers Confidence / gap
Does AI improve
immediate
problem-solving
performance?
Often yes, especially when
explanations, feedback, or tutoring are
available; assisted performance is not
the same as learning.
Kestin 2025; Kumar 2023;
Bastani 2025; Miller 2026
High confidence for
immediate performance;
interpretation must separate
assisted output from unaided
knowledge.
Does AI improve
durable, unaided
learning?
Conditionally. One-week undergraduate
gains can persist under augmentation,
but unrestricted solution access can
reduce later unassisted performance or
retention.
Contractor & Reyes 2026;
Bastani 2025; Liu et al. 2026;
Rismanchian et al. 2026
Moderate confidence; direct
long-duration undergraduate
STEM RCTs remain scarce.
Is there a
cognitive-offloading /
deskilling effect?
Yes under some usage patterns.
Evidence appears as copying/over-trust,
reduced persistence, reduced
time-on-task, and weaker unassisted
performance.
Krupp 2024; Bastani 2025; Liu
et al. 2026; Rismanchian et al.
2026
Moderate-to-high for
mechanism; ecological and
causal strength varies by
study.
Does answer-giving
versus
Socratic/scaffolded
tutoring matter?
Very likely yes. Attempt-before-help,
pedagogical prompts, guardrails, and
Socratic questions consistently point
toward greater engagement, reflection,
transfer, or preserved independent
ability.
Kumar 2023; Miller 2026;
LearnLM 2025; Degen 2025;
Bastani 2025; SafeTutors 2026
Strong convergent evidence,
but few studies randomize the
exact same model into
answer-giving versus Socratic
modes with delayed STEM
tests.
Can we claim
long-term retention
benefits in
undergraduate
STEM?
Not yet. The seed is short-term; the
strongest delayed evidence either uses
non-STEM tasks, younger students,
broad participant samples, or
quasi-experimental college math data.
Kestin 2025; Contractor &
Reyes 2026; Rismanchian et al.
2026
Low-to-moderate; this is the
central research gap.
The emerging mechanism
A useful working model for a new study
AI as cognitive amplifier: student attempts first -> AI diagnoses or asks a question -> targeted
hint/explanation -> student produces the next reasoning step -> delayed unaided retrieval/transfer.
AI as cognitive substitute: student delegates the reasoning step -> AI returns a complete solution
-> student verifies superficially or copies -> immediate task success rises while practice
opportunities, persistence, and durable schema construction fall.
A rigorous next RCT would hold content, model, time, and interface constant while randomizing only the
assistance policy: (A) full answer on request, (B) explanation after an independent attempt, and (C)
Socratic/hint-based tutoring that withholds full solutions until predefined thresholds. Outcomes should
include immediate mastery, unaided transfer, delayed retention at one week and several weeks,
time-on-task, persistence, hint/answer-seeking behavior, and process traces that distinguish productive
self-explanation from cognitive surrender.
The most informative primary endpoint for your exact question is not homework completion or post-test
performance while AI is present; it is delayed unaided performance on novel or isomorphic problems after AI
removal. A secondary endpoint should capture whether students continue attempting difficult problems
before seeking an answer.

AI tutoring, cognitive offloading, and undergraduate STEM learning
12Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
5. Selection method, relationship audit, and
limitations
Selection rule. I prioritized papers that are closest to the target construct: undergraduate or STEM problem
solving; objective learning or retention; manipulation or observation of AI assistance style;
scaffolding/Socratic interaction; and evidence of cognitive offloading, overreliance, or independent
performance after AI removal. Canonical papers were retained only when they directly explain the seed
tutor's design.
Relationship verification. "Confirmed antecedent" means the work is cited by the uploaded seed paper.
"Confirmed extension/descendant" means the newer paper directly cites the published seed or its 2024
preprint, or explicitly calls itself an extension. "Inferred / adjacent" means the paper was selected because it
tests the same causal mechanism, but a direct citation relationship was not confirmed in the accessible full
text searched.
Publication-status caveat. Several of the most directly relevant 2026 studies are arXiv preprints. Their
results should be treated as provisional until peer review and replication. Conversely, peer-reviewed status
does not remove limitations in population, intervention duration, or causal identification.
Bibliometric caveat. This report is a targeted evidence map, not an exhaustive citation census. The seed
article has accumulated a fast-moving citing literature, and some recent papers may be indexed unevenly
across databases. The ten recent selections are therefore optimized for your research question, not for raw
citation count.
Recent paper Relationship status used
here
Verification basis
Miller, K et al. CONFIRMED EXTENSION Direct statement/citation located in accessible full text.
LearnLM Team, Google & Eedi
et al.
CONFIRMED DESCENDANT Direct statement/citation located in accessible full text.
Hazra, R et al. CONFIRMED DESCENDANT Direct statement/citation located in accessible full text.
Degen, P et al. CONFIRMED
PREPRINT-LINEAGE
Direct statement/citation located in accessible full text.
Sun, Y et al. CONFIRMED DESCENDANT Direct statement/citation located in accessible full text.
Bastani, H et al. INFERRED / PARALLEL No direct seed citation confirmed in accessible full text; thematic
relationship is an inference.
Contractor, Z et al. INFERRED / ADJACENT No direct seed citation confirmed in accessible full text; thematic
relationship is an inference.
Liu, G et al. INFERRED / ADJACENT No direct seed citation confirmed in accessible full text; thematic
relationship is an inference.
Rismanchian, S et al. INFERRED / ADJACENT No direct seed citation confirmed in accessible full text; thematic
relationship is an inference.
Wang, G et al. INFERRED / SYNTHESIS No direct seed citation confirmed in accessible full text; thematic
relationship is an inference.
Primary source trail for recent papers
• Miller (2026): Springer article and notes explicitly identify the same AI tutoring framework as Kestin et al. and call the
study an extension.
• LearnLM Team (2025): arXiv full text directly lists Kestin et al. (2025) in its references and discusses mixed evidence on
guarded versus unguarded AI tutoring.

AI tutoring, cognitive offloading, and undergraduate STEM learning
13Evidence map centered on Kestin et al. (2025) | literature checked through 18 Aug 2026
• Sun & Liu (2026): Springer/Smart Learning Environments article directly cites Kestin et al. when motivating the gap
beyond structured physics.
• SafeTutors (2026): arXiv full text directly cites Kestin et al. and defines answer over-disclosure and lost scaffolding as
pedagogical safety risks.
• Degen & Asanov (2025): arXiv PDF directly cites the 2024 Research Square preprint version of the Kestin study.
• Bastani et al. (2025), Contractor & Reyes (2026), Liu et al. (2026), Rismanchian et al. (2026), and Wang et al. (2026):
included for topical/causal relevance; no direct Kestin citation was confirmed in the accessible full text searched.
6. Compact bibliography: the 20 selected related
papers
Seed paper is listed on the cover and in Section 1; the 20 entries below are the selected antecedent and recent papers
requested.
1. Bloom, B. S. (1984). The 2 sigma problem: The search for methods of
group instruction as effective as one-to-one tutoring. Educational
Researcher, 13(6), 4-16. doi:10.3102/0013189X013006004
2. Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in
problem-solving. Journal of Child Psychology and Psychiatry, 17(2),
89-100.
3. Sweller, J. (2011). Cognitive load theory. In J. P. Mestre & B. H. Ross
(Eds.), The Psychology of Learning and Motivation: Cognition in Education
(pp. 37-76). Elsevier Academic Press.
4. Shute, V. J. (2008). Focus on formative feedback. Review of
Educational Research, 78(1), 153-189. doi:10.3102/0034654307313795
5. Hake, R. R. (1998). Interactive-engagement versus traditional
methods: A six-thousand-student survey of mechanics test data for
introductory physics courses. American Journal of Physics, 66(1), 64-74.
doi:10.1119/1.18809
6. Freeman, S., Eddy, S. L., McDonough, M., Smith, M. K., Okoroafor, N.,
Jordt, H., & Wenderoth, M. P. (2014). Active learning increases student
performance in science, engineering, and mathematics. Proceedings of
the National Academy of Sciences, 111(23), 8410-8415.
doi:10.1073/pnas.1319030111
7. Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G.
(2019). Measuring actual learning versus feeling of learning in response
to being actively engaged in the classroom. Proceedings of the National
Academy of Sciences, 116(39), 19251-19257.
doi:10.1073/pnas.1821936116
8. Kumar, H., Rothschild, D. M., Goldstein, D. G., & Hofman, J. M. (2023).
Math education with large language models: Peril or promise? SSRN
Electronic Journal. doi:10.2139/ssrn.4641653
9. Krupp, L., Steinert, S., Kiefer-Emmanouilidis, M., Avila, K. E., Lukowicz,
P., Kuhn, J., Küchemann, S., & Karolus, J. (2024). Unreflected acceptance -
Investigating the negative consequences of ChatGPT-assisted problem
solving in physics education. In HHAI 2024: Hybrid Human AI Systems for
the Social Good (pp. 199-212). doi:10.3233/FAIA240195
10. Forero, M. G., & Herrera-Suárez, H. J. (2023). ChatGPT in the
classroom: Boon or bane for physics students' academic performance?
arXiv preprint arXiv:2312.02422. arXiv:2312.02422
11. Miller, K. (2026). Prompt Matters: How Pedagogical Engineering
Shapes Behavior and Engagement with AI Tutors. Technology, Knowledge
and Learning. doi:10.1007/s10758-026-09983-6
12. LearnLM Team, Google & Eedi. (2025). AI tutoring can safely and
effectively support students: An exploratory RCT in UK classrooms. arXiv
preprint arXiv:2512.23633. arXiv:2512.23633
13. Hazra, R., Ghuku, B., Marchenko, I., Tokarieva, Y., Layek, S., Banerjee,
S., Stoyanovich, J., & Pechenizkiy, M. (2026). SafeTutors: Benchmarking
Pedagogical Safety in AI Tutoring Systems. arXiv preprint
arXiv:2603.17373. arXiv:2603.17373
14. Degen, P.-B., & Asanov, I. (2025). Beyond Automation: Socratic AI,
Epistemic Agency, and the Implications of the Emergence of
Orchestrated Multi-Agent Learning Architectures. arXiv preprint
arXiv:2508.05116. arXiv:2508.05116
15. Sun, Y., & Liu, F. (2026). The impact of an AI Digital Teacher on
human-AI collaborative learning in higher education. Smart Learning
Environments, 13, 30. doi:10.1186/s40561-026-00454-0
16. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R.
(2025). Generative AI without guardrails can harm learning: Evidence
from high school mathematics. Proceedings of the National Academy of
Sciences, 122(26), e2422633122. doi:10.1073/pnas.2422633122
17. Contractor, Z., & Reyes, G. (2026). Experimental Evidence on the
Learning Impact of Generative AI. arXiv preprint arXiv:2607.08849.
arXiv:2607.08849
18. Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R.
(2026). AI Assistance Reduces Persistence and Hurts Independent
Performance. arXiv preprint arXiv:2604.04721. arXiv:2604.04721
19. Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E.
(2026). Faster Completion, Less Learning: Generative AI Reduced Study
Time on Math Problems and the Knowledge They Build. arXiv preprint
arXiv:2605.21629. arXiv:2605.21629
20. Wang, G., Wang, W., Yang, D., & Ren, J. (2026). Generative AI,
Cognitive Offloading, and Learner Agency in Higher Education: A Scoping
Review. Behavioral Sciences, 16(7), 1150. doi:10.3390/bs16071150