Evidence of a Cognitive Shift in AI Education: How Students Are Rethinking Human Intelligence?

Islem Rekik

arxiv: 2605.16292 · v1 · pith:NEY7LMTEnew · submitted 2026-04-14 · 💻 cs.CY · cs.AI

Evidence of a Cognitive Shift in AI Education: How Students Are Rethinking Human Intelligence?

Islem Rekik This is my paper

Pith reviewed 2026-05-21 01:27 UTC · model grok-4.3

classification 💻 cs.CY cs.AI

keywords AI educationhuman intelligencestudent perceptionscognitive shiftlongitudinal studylearner autonomyAI tools

0 comments

The pith

Students in AI courses increasingly prefer human intelligence over AI as tools become routine, with majorities favoring HI by 2026.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines six years of classroom poll data from AI-focused courses to show how student views on human intelligence evolve with greater AI exposure. Early years showed slight AI preference, followed by phases of hype, distrust, trust, and dependency. From 2024 a clear shift emerged toward valuing human intelligence more. By 2026 this reached 65 percent in technical courses and 90 percent in design courses. A sympathetic reader would care because it suggests routine AI use may restore appreciation for uniquely human thinking and affect how education builds learner independence.

Core claim

Longitudinal poll data from 471 students in technical AI courses such as Machine Learning and design-oriented courses such as Design Thinking for AI, collected 2020-2026, reveal four phases—hype, distrust, trust, and dependency—with a consistent move toward preferring human intelligence starting in 2024. By 2026 preference for HI stood at 65 percent in technical cohorts (12-point rise from 2025) and 90 percent in design cohorts (36-point rise), pointing to a gradual reappraisal of human intelligence once AI functions as an everyday tool, with direct implications for learner autonomy and epistemic agency.

What carries the argument

Repeated classroom poll questions tracking preference between human intelligence and AI across successive student cohorts in the same courses.

If this is right

Routine integration of AI tools prompts students to revalue human intelligence above AI capabilities.
This reappraisal supports greater learner autonomy and epistemic agency in AI-augmented settings.
The observed sequence of hype, distrust, trust, and dependency may appear in other disciplines adopting generative AI.
Course design should anticipate and work with this shift rather than assume steady AI preference.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The pattern may indicate that easy AI access highlights what remains distinctively human rather than diminishing it.
Similar longitudinal polls in non-AI fields could test whether the shift is specific to computer science students or more general.
If the preference stabilizes, curricula could emphasize human strengths such as judgment and creativity that complement AI output.
The four-phase cycle offers a possible model for predicting public attitudes toward new technologies beyond education.

Load-bearing premise

Classroom poll responses accurately and unbiasedly capture students' genuine valuations of human intelligence relative to AI, independent of course context, instructor influence, or self-reporting tendencies.

What would settle it

A new poll conducted in the same courses in 2027 or later that shows preference for HI dropping below 50 percent or remaining flat, or independent interviews revealing that students gave socially expected answers rather than private views.

Figures

Figures reproduced from arXiv: 2605.16292 by Islem Rekik.

read the original abstract

Perceptions of intelligence shape how learners evaluate and rely on artificial intelligence (AI) systems. Despite rapid advances in AI capabilities, the impact of sustained exposure to these tools on students' valuation of human intelligence (HI) relative to AI remains underexplored. This paper presents a longitudinal analysis of classroom poll responses collected between 2020 and 2026 in AI-focused undergraduate and MSc courses in computer science. Data from 471 students across technical courses (such as Machine Learning and Deep Graph Learning) and design-oriented courses (such as Design Thinking for AI) reveal four recurring phases: hype, distrust, trust, and dependency. While early responses in 2020 slightly favored AI, a consistent shift toward HI emerged from 2024 onward across all MSc cohorts. By 2026, preference for HI reached 65 percent in a technical course (a 12 percentage-point increase from 2025) and 90 percent in a design-oriented course (a 36 percentage-point increase). These findings suggest a gradual reappraisal of human intelligence as AI becomes a routine tool, with implications for learner autonomy and epistemic agency. We conclude by reflecting on this cognitive shift from favoring artificial intelligence toward prioritizing human intelligence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper tracks six years of student poll answers in AI courses and claims a shift toward preferring human intelligence by 2026, but the methods are too thin to support the claim.

read the letter

The main point is straightforward: over six years of polls in machine learning and design-thinking courses, student answers moved from slight AI preference early on to 65% favoring human intelligence in technical classes and 90% in design ones by 2026, with a pattern of hype, distrust, trust, and dependency phases. The longitudinal classroom data is the part that stands out, since most work on this topic is single snapshots rather than repeated measures across cohorts. That gives a concrete timeline that could help people thinking about how exposure to AI tools changes what students value in themselves. The numbers on the increases are new enough to be worth noting if they hold up. The weak part is the data itself. The abstract reports exact percentages and point changes but gives no poll wording, no yearly response counts, no rates, and no tests for whether the shifts are real or just sampling variation. The bigger jump in the design course lines up with the stress-test worry that answers may track course framing or what the instructor highlights rather than any independent reappraisal. Nothing in the description rules out social-desirability effects or selection into the classes. If the full paper adds raw data, anonymity details, or comparison groups, that would fix most of it; otherwise the link from poll to cognitive shift stays loose. This is mainly useful for people who teach AI and want to discuss learner autonomy with their own students. A reader looking for evidence that generalizes beyond one instructor's classes will need more before treating the percentages as settled. Send it for review so the authors can supply the missing method details and let referees check whether the pattern survives basic scrutiny.

Referee Report

2 major / 2 minor

Summary. The paper presents a longitudinal analysis of classroom poll responses collected from 471 students in AI-focused undergraduate and MSc courses (technical courses like Machine Learning and design-oriented courses like Design Thinking for AI) between 2020 and 2026. It identifies four recurring phases in student perceptions (hype, distrust, trust, and dependency) and reports a shift toward preference for human intelligence (HI) over AI, with HI preference reaching 65% in a technical course (12 percentage-point increase from 2025) and 90% in a design-oriented course (36 percentage-point increase) by 2026, suggesting a cognitive reappraisal as AI becomes routine.

Significance. If the poll data reliably capture genuine student valuations independent of context, the longitudinal trends across course types could provide valuable observational evidence on how sustained AI exposure in education influences perceptions of intelligence, with potential implications for curriculum design, learner autonomy, and epistemic agency. The multi-year, multi-cohort design offers a rare window into evolving attitudes that could inform broader discussions in AI education research.

major comments (2)

Data collection description: The reported preference percentages and shifts (e.g., 65% and 90% by 2026) are presented without the exact poll question wording, yearly sample sizes per course, response rates, or any mention of statistical tests, confidence intervals, or controls for selection bias, preventing assessment of whether the trends are statistically meaningful or robust.
Interpretation of results: The claim that the data demonstrate a 'gradual reappraisal of human intelligence' and 'cognitive shift' does not address or control for course-specific framing effects, such as emphasis on human-centered design in the design-oriented cohort (which shows the larger 36-point jump), instructor influence, or demand characteristics in AI-focused classroom settings.

minor comments (2)

The abstract and results would benefit from a table summarizing sample sizes and response rates by year and course type to support the longitudinal claims.
Clarify whether polls were anonymous and how multiple responses from the same students across years were handled, if at all.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We address each major comment below, indicating where we will revise the paper to improve transparency and strengthen the interpretation while maintaining the integrity of the observational data.

read point-by-point responses

Referee: Data collection description: The reported preference percentages and shifts (e.g., 65% and 90% by 2026) are presented without the exact poll question wording, yearly sample sizes per course, response rates, or any mention of statistical tests, confidence intervals, or controls for selection bias, preventing assessment of whether the trends are statistically meaningful or robust.

Authors: We agree that greater transparency in data collection is essential. In the revised manuscript, we will add the exact poll question wording used consistently across years, a table with yearly sample sizes broken down by course (technical vs. design-oriented), available response rates, and any post-hoc statistical tests or confidence intervals that can be computed from the aggregated responses. We will also explicitly discuss potential selection bias arising from the classroom setting and voluntary participation. These additions will allow readers to better evaluate the robustness of the reported trends. revision: yes
Referee: Interpretation of results: The claim that the data demonstrate a 'gradual reappraisal of human intelligence' and 'cognitive shift' does not address or control for course-specific framing effects, such as emphasis on human-centered design in the design-oriented cohort (which shows the larger 36-point jump), instructor influence, or demand characteristics in AI-focused classroom settings.

Authors: We acknowledge that course framing and contextual factors may contribute to the observed differences. The larger shift in the design-oriented course aligns with its human-centered focus, and we will revise the interpretation section to discuss this as a plausible contributing factor rather than attributing the trend solely to a broad cognitive shift. We will add a dedicated limitations paragraph addressing instructor influence and demand characteristics, noting that the study is observational and these variables were not experimentally controlled. At the same time, the consistency of the HI preference increase across both technical and design cohorts from 2024 onward provides supporting evidence for the overall pattern, which we will emphasize while qualifying the claims. revision: partial

Circularity Check

0 steps flagged

No circularity: purely observational poll reporting

full rationale

The paper presents descriptive longitudinal trends from classroom poll responses collected 2020-2026, reporting raw percentage shifts (e.g., HI preference rising to 65% and 90% by 2026) without any mathematical derivation, equations, fitted parameters, or predictions. No self-definitional loops, fitted-input-as-prediction, or load-bearing self-citations appear; the claims rest directly on the stated data collection rather than reducing to inputs by construction. This is a standard empirical report whose central findings are independent of any internal redefinition or ansatz smuggling.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The analysis rests on the assumption that self-reported poll answers validly reflect cognitive shifts; no free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption Poll responses from classroom settings provide an unbiased measure of students' relative valuation of human versus artificial intelligence.
All reported trends and percentage increases are derived directly from interpreting these responses as evidence of a cognitive shift.

pith-pipeline@v0.9.0 · 5741 in / 1193 out tokens · 37302 ms · 2026-05-21T01:27:19.535359+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 2 internal anchors

[1]

npj Science of Learning , year =

Behrendt, Mei Grace and Clark, Carrie and Elliott, McKenna and Dauer, Joseph , title =. npj Science of Learning , year =. doi:10.1038/s41539-024-00231-z , url =

work page doi:10.1038/s41539-024-00231-z
[2]

and others , title =

Fleur, David S. and others , title =. npj Science of Learning , year =. doi:10.1038/s41539-021-00089-5 , url =

work page doi:10.1038/s41539-021-00089-5
[3]

Strachan, James W. A. and Albergo, Dalila and Borghini, Giulia and others , title =. Nature Human Behaviour , year =. doi:10.1038/s41562-024-01882-z , url =

work page doi:10.1038/s41562-024-01882-z
[4]

When Combinations of Humans and

Vaccaro, Marija and others , title =. Nature Human Behaviour , year =. doi:10.1038/s41562-024-02024-1 , url =

work page doi:10.1038/s41562-024-02024-1
[5]

arXiv preprint arXiv:2503.10728 , year=

Darkbench: Benchmarking dark patterns in large language models , author=. arXiv preprint arXiv:2503.10728 , year=

work page arXiv
[6]

arXiv preprint arXiv:2504.18458 , year=

Fast-slow thinking for large vision-language model reasoning , author=. arXiv preprint arXiv:2504.18458 , year=

work page arXiv
[7]

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? , author=. arXiv preprint arXiv:2504.13837 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[8]

PhysBench: Benchmarking and enhancing vision-language models for physical world understanding.arXiv preprint arXiv:2501.16411, 2025

Physbench: Benchmarking and enhancing vision-language models for physical world understanding , author=. arXiv preprint arXiv:2501.16411 , year=

work page arXiv
[9]

arXiv preprint arXiv:2503.00564 , year=

Tooldial: Multi-turn dialogue generation method for tool-augmented language models , author=. arXiv preprint arXiv:2503.00564 , year=

work page arXiv
[10]

Evaluating Large Language Models Trained on Code

Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[11]

Science , volume=

Competition-level code generation with alphacode , author=. Science , volume=. 2022 , publisher=

work page 2022
[12]

Advances in neural information processing systems , volume=

Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

work page
[13]

International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

Multi-sensory Cognitive Computing for Learning Population-Level Brain Connectivity , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2025 , organization=

work page 2025
[14]

MICCAI Workshop on Deep Generative Models , pages=

CogGNN: Cognitive Graph Neural Networks in Generative Connectomics , author=. MICCAI Workshop on Deep Generative Models , pages=. 2025 , organization=

work page 2025
[15]

Computers and Education: Artificial Intelligence , volume=

Students’ use of large language models in engineering education: A case study on technology acceptance, perceptions, efficacy, and detection chances , author=. Computers and Education: Artificial Intelligence , volume=. 2023 , publisher=

work page 2023
[16]

Learning and individual differences , volume=

ChatGPT for good? On opportunities and challenges of large language models for education , author=. Learning and individual differences , volume=. 2023 , publisher=

work page 2023
[17]

arXiv , year=

Is ChatGPT Massively Used by Students Nowadays? A Survey on the Use of Large Language Models such as ChatGPT in Educational Settings , author=. arXiv , year=

work page
[18]

npj Science of Learning , volume=

Metacognition: ideas and insights from neuro-and educational sciences , author=. npj Science of Learning , volume=. 2021 , publisher=

work page 2021
[19]

Educational Psychologist , volume=

Why Minimal Guidance During Instruction Does Not Work: An Analysis of the Failure of Constructivist, Discovery, Problem-Based, Experiential, and Inquiry-Based Teaching , author=. Educational Psychologist , volume=

work page
[20]

Psychology and the Real World: Essays Illustrating Fundamental Contributions to Society , pages=

Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning , author=. Psychology and the Real World: Essays Illustrating Fundamental Contributions to Society , pages=. 2011 , publisher=

work page 2011
[21]

Nature Human Behaviour , volume=

When combinations of humans and AI are useful: A systematic review and meta-analysis , author=. Nature Human Behaviour , volume=. 2024 , publisher=

work page 2024
[22]

Scaling Learning Algorithms Towards

Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

work page
[23]

and Osindero, Simon and Teh, Yee Whye , journal =

Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

work page
[24]

2016 , publisher=

Deep learning , author=. 2016 , publisher=

work page 2016

[1] [1]

npj Science of Learning , year =

Behrendt, Mei Grace and Clark, Carrie and Elliott, McKenna and Dauer, Joseph , title =. npj Science of Learning , year =. doi:10.1038/s41539-024-00231-z , url =

work page doi:10.1038/s41539-024-00231-z

[2] [2]

and others , title =

Fleur, David S. and others , title =. npj Science of Learning , year =. doi:10.1038/s41539-021-00089-5 , url =

work page doi:10.1038/s41539-021-00089-5

[3] [3]

Strachan, James W. A. and Albergo, Dalila and Borghini, Giulia and others , title =. Nature Human Behaviour , year =. doi:10.1038/s41562-024-01882-z , url =

work page doi:10.1038/s41562-024-01882-z

[4] [4]

When Combinations of Humans and

Vaccaro, Marija and others , title =. Nature Human Behaviour , year =. doi:10.1038/s41562-024-02024-1 , url =

work page doi:10.1038/s41562-024-02024-1

[5] [5]

arXiv preprint arXiv:2503.10728 , year=

Darkbench: Benchmarking dark patterns in large language models , author=. arXiv preprint arXiv:2503.10728 , year=

work page arXiv

[6] [6]

arXiv preprint arXiv:2504.18458 , year=

Fast-slow thinking for large vision-language model reasoning , author=. arXiv preprint arXiv:2504.18458 , year=

work page arXiv

[7] [7]

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? , author=. arXiv preprint arXiv:2504.13837 , year=

work page internal anchor Pith review Pith/arXiv arXiv

[8] [8]

PhysBench: Benchmarking and enhancing vision-language models for physical world understanding.arXiv preprint arXiv:2501.16411, 2025

Physbench: Benchmarking and enhancing vision-language models for physical world understanding , author=. arXiv preprint arXiv:2501.16411 , year=

work page arXiv

[9] [9]

arXiv preprint arXiv:2503.00564 , year=

Tooldial: Multi-turn dialogue generation method for tool-augmented language models , author=. arXiv preprint arXiv:2503.00564 , year=

work page arXiv

[10] [10]

Evaluating Large Language Models Trained on Code

Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

work page internal anchor Pith review Pith/arXiv arXiv

[11] [11]

Science , volume=

Competition-level code generation with alphacode , author=. Science , volume=. 2022 , publisher=

work page 2022

[12] [12]

Advances in neural information processing systems , volume=

Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

work page

[13] [13]

International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

Multi-sensory Cognitive Computing for Learning Population-Level Brain Connectivity , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2025 , organization=

work page 2025

[14] [14]

MICCAI Workshop on Deep Generative Models , pages=

CogGNN: Cognitive Graph Neural Networks in Generative Connectomics , author=. MICCAI Workshop on Deep Generative Models , pages=. 2025 , organization=

work page 2025

[15] [15]

Computers and Education: Artificial Intelligence , volume=

Students’ use of large language models in engineering education: A case study on technology acceptance, perceptions, efficacy, and detection chances , author=. Computers and Education: Artificial Intelligence , volume=. 2023 , publisher=

work page 2023

[16] [16]

Learning and individual differences , volume=

ChatGPT for good? On opportunities and challenges of large language models for education , author=. Learning and individual differences , volume=. 2023 , publisher=

work page 2023

[17] [17]

arXiv , year=

Is ChatGPT Massively Used by Students Nowadays? A Survey on the Use of Large Language Models such as ChatGPT in Educational Settings , author=. arXiv , year=

work page

[18] [18]

npj Science of Learning , volume=

Metacognition: ideas and insights from neuro-and educational sciences , author=. npj Science of Learning , volume=. 2021 , publisher=

work page 2021

[19] [19]

Educational Psychologist , volume=

Why Minimal Guidance During Instruction Does Not Work: An Analysis of the Failure of Constructivist, Discovery, Problem-Based, Experiential, and Inquiry-Based Teaching , author=. Educational Psychologist , volume=

work page

[20] [20]

Psychology and the Real World: Essays Illustrating Fundamental Contributions to Society , pages=

Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning , author=. Psychology and the Real World: Essays Illustrating Fundamental Contributions to Society , pages=. 2011 , publisher=

work page 2011

[21] [21]

Nature Human Behaviour , volume=

When combinations of humans and AI are useful: A systematic review and meta-analysis , author=. Nature Human Behaviour , volume=. 2024 , publisher=

work page 2024

[22] [22]

Scaling Learning Algorithms Towards

Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

work page

[23] [23]

and Osindero, Simon and Teh, Yee Whye , journal =

Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

work page

[24] [24]

2016 , publisher=

Deep learning , author=. 2016 , publisher=

work page 2016