CoChat

A Systematic Literature Review on the Pedagogical Implications and Impact of GenAI on Students' Critical Thinking

Open access

The DOI matches its Crossref record; 103 of the 103 references it cites were found in a registry or an index; 64 of the 206 sources its claims cite were checked for support.

Abstract

Critical Thinking (CT) is recognized as a foundational competency for professional readiness, innovation, and ethical reasoning in higher education, enabling students to analyze information, evaluate evidence, and make reasoned decisions in complex environments. The rapid integration of Generative Artificial Intelligence (GenAI) tools, such as large language models, presents new opportunities and risks for CT development. This study conducts a systematic literature review to synthesize empirical evidence on the pedagogical implications and cognitive impact of GenAI on students’ CT. Following PRISMA guidelines, and search terms around GenAI Tools, Critical Thinking And Higher Education, on five major education research databases—Web of Science; Scopus; EBSCOhost (Education Source, ERIC, and APA PsycInfo); and Compendex and Inspec (Elsevier)—63 empirical studies published between January 2023 and April 2025 were analyzed across higher education contexts, disciplines, and intervention designs. Results indicate that GenAI offers notable cognitive affordances, including scaffolding reflective reasoning, promoting self-regulation, and facilitating iterative dialogue and argument evaluation. Pedagogical strategies clustered into four primary integration typologies: AI-based feedback prompts, dialogue simulation and reflection, AI-supported peer review, and critical engagement with AI-generated content. Nearly half of the studies reported statistically significant CT improvements, particularly when GenAI use was guided by structured prompts, reflective activities, and performance-based assessment. However, multiple risks persist, including cognitive offloading, uncritical acceptance of AI outputs, and diminished intellectual autonomy, especially in unguided or surface-level usage. This review highlights the need for intentional pedagogical design, validated CT assessment tools, and longitudinal studies to ensure GenAI acts as a catalyst rather than a substitute for human reasoning. By identifying effective integration strategies and outlining potential pitfalls, this study provides evidence-informed guidance for educators and institutions aiming to responsibly leverage GenAI to strengthen students’ CT skills.

Overview

CoChat Summary

TL;DR

This PRISMA systematic review synthesizes 63 empirical studies (January 2023–April 2025) on GenAI and higher-education students’ critical thinking, finding both cognitive affordances and pedagogical risks. Nearly half of the studies reported statistically significant CT gains, but effects depended on structured integration and were constrained by heterogeneous, mostly short-term evidence.

Key findings

  • Of 63 included studies, 30 (47%) reported statistically significant quantitative CT gains. Examples include Shalong et al.’s ~13% improvement in self-directed learning and CT scores sustained at follow-up; Liu and Wang’s CTDI-CV gain of 14.57 points (20%, 70-item scale); Essel et al.’s CTDS effect (partial η²p = 0.229); and Fabio et al.’s MANOVA result (F(2, 123) = 16.43, p < 0.01, η²p = 0.29, d ≈ 0.82–0.89).
  • GenAI integration clustered into four typologies: feedback prompts/AI-based metacognitive scaffold, dialogue simulation and reflection, AI-based peer review, and critical engagement with AI-generated content. ChatGPT dominated (41 studies, 65%); most interventions were face-to-face synchronous (66%), while 16 studies (25%) occurred outside class.
  • Theoretical and assessment gaps were prominent: 27 studies (42.9%) explicitly used a recognized CT framework, while 36 (57.1%) did not. Facione’s Delphi Report guided 11%, Bloom’s Revised Taxonomy underpinned 10 studies (15.9%), and Paul and Elder’s framework and HOTS frameworks each informed four studies. Standardized instruments were underused; many studies relied on custom Likert scales or rubrics with limited validation and inconsistent inter-rater reliability.
  • Risks included cognitive offloading, uncritical acceptance of AI outputs, overreliance, and the illusion of understanding. Benefits were most evident when GenAI use was guided by structured prompts, reflective activities, and performance-based assessment.

Methods

  • PRISMA-guided systematic literature review registered on OSF (en3p5). Searches covered five major databases—Web of Science; Scopus; EBSCOhost (Education Source, ERIC, APA PsycInfo); and Compendex & Inspec—for English-language studies published January 2023–April 2025. From 1,788 records, 800 duplicates were removed, 988 underwent abstract screening, 170 moved to full-text screening, and 63 empirical studies were included.
  • Inclusion criteria required higher-education settings, empirical studies, English language, and measurable CT outcomes or observations. Exclusion criteria removed theoretical papers, K-12 or informal contexts, non-GenAI AI studies, and studies without student participants. Screening reliability checks yielded 92% agreement on an abstract sample, 86% initial full-text agreement, and a final 91% agreement with 100% agreement on the final 30 papers.
  • Analysis used a hybrid inductive-deductive approach with structured extraction in Covidence, including LLM-assisted extraction under human oversight. Extracted variables covered bibliographic details, context, participants, design, GenAI tools, integration strategies, CT instruments, and reported CT outcomes.

Limitations

  • Stated: the search was restricted to English-language studies from January 2023–April 2025 indexed in five databases, likely underrepresenting non-English research and biasing the evidence base toward contexts that publish in English. Heterogeneity of study designs, interventions, and measurement instruments precluded quantitative synthesis. Many studies used custom or adapted instruments with limited psychometric reporting, and rubric- or coding-based assessments often omitted inter-rater reliability. Many studies were exploratory, descriptive, or quasi-experimental, often with small samples, short follow-up, and minimal control for confounders; few used randomized controlled designs or longitudinal tracking. Publication bias may have favored positive findings. Geographic and disciplinary concentration further limits generalizability.
  • Evident from the text: over 57% of reviewed papers did not reference a formal CT framework, and validated CT instruments remained underutilized, which the authors note constrains explanatory power and comparability across studies.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/a19030179

Cites

103 of 103 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 CT skills identified by Facione [2,3].
Figure 1 CT skills identified by Facione [2,3].CC-BY · From the PDF
Image 2
CC-BY · From the PDF
Figure 3 Distribution of research paradigms.
Figure 3 Distribution of research paradigms.CC-BY · From the PDF

Verification

Claims the paper makes

46 Valid · 10 Not valid · 92 Not checked · 88 Own results and interpretations

125 more claims · 25 with their cited sources checked

Create a free account to see all 236 claims and to verify the 21 references still open.

See all claims — free account
  • It may state two separate claims as one. 85
  • Probably not a substantial claim. 41

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • ConsistentF(2, 123) = 16.43reported p < .01 · recomputed p < .001

    “F(2, 123) = 16.43,p < .01”

1 of 1

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

5 of 6 · the rest aren’t loaded

1 more finding

Create a free account to see every finding of this check.

See all findings — free account

Citations

Cites

Research works printed but not linked: 12

  1. Facione, P. (1990). Critical Thinking: A Statement of Expert Consensus for Purposes of Educational Assessment and Instruction (The Delphi Report). Educational Resources Information Center (ERIC), California Academic Press.
  2. Shanto, 2024, Enriching Learning Process with Generative AI: A Proposed Framework to Cultivate Critical Thinking in Higher Education Using Chat GPT, Tuijin Jishu J. Propuls. Technol., 45, 3019
  3. Husain, 2023, Educational Renaissance: Unleashing the Potential of Interdisciplinary AI Integration, J. Environ. Sci. Technol., 2, 159
  4. Facione, 2011, Critical Thinking: What It Is and Why It Counts, Insight Assess., 1, 1
  5. Paul, R., and Elder, L. (2013). Critical Thinking: Tools for Taking Charge of Your Professional and Personal Life, Pearson Education.
  6. Ennis, 1985, A Logical Basis for Measuring Critical Thinking Skills, Educ. Leadersh., 43, 44
  7. Jiang, 2024, Enhancing Critical Thinking Skills with ChatGPT-Powered Activities in Chinese Language Classrooms, J. Chin. Lang. Teach., 5, 47
  8. Wu, 2025, Exploring ChatGPT’s Potential to Enhance Problem-Solving and Critical Thinking in Education, Educ. Technol. Soc., 28, 310
  9. Chu, 2025, How GenAI-Supported Multi-Modal Presentations Benefit Students with Different Motivation Levels: Evidence from Digital Storytelling Performance, Critical Thinking Awareness, and Learning Attitude, Educ. Technol. Soc., 28, 250
  10. Chang, 2024, Facilitating Nursing and Health Education by Incorporating ChatGPT into Learning Designs, Educ. Technol. Soc., 27, 215
  11. Palupi, S., Hardi, R., Pribadi, A.S., Zulkarnain, R., Setiawan, M.N., Sari, N.W.W., and Utomo, D.T. Enhancing Computational Strategies to Decode ChatGPT’s Influence on the Critical Thinking Abilities of University Students. Proceedings of the 2024 4th International Conference of Science and Information Technology in Smart Administration (ICSINTESA), Balikpapan, Indonesia, 12 July 2024, IEEE.
  12. Balart, 2025, A Framework for Integrating Artificial General Intelligence into Engineering Education, Int. J. Eng. Educ., 41, 171

Cited by

See all 29

Metadata

Authors
Balart, T., Díaz, B., & Shryock, K.
Affiliations
  • Texas A&M University (US)
  • Utah State University (US)
Corresponding author
Balart, Trini
Venue
Algorithms
Publisher
MDPI AG
Published
2026-02-27
Volume
19
Issue
3
Pages
179
ISSN
1999-4893 (ISSN-L)
OpenAlex
W7131796316
Cited by
29
References
78
Citations by year
2026: 29
FWCI
28.39
Type
journal article
Language
en
Open access
Locations
  • Algorithms (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • OSF Preprints (OSF Preprints) (submittedVersion)
Funders
ANID/PIA/Basal Funds for Centers of Excellence (FB0003)
Topics
  • Social Sciences › Education › Artificial Intelligence in Education
  • Social Sciences › Education › Education and Critical Thinking Development
  • Social Sciences › Psychology › Developmental and Educational Psychology › Educational Strategies and Epistemologies
Keywords
generative AI, critical thinking, higher education, large language models, artificial intelligence in education, teaching methods, cognitive offloading, intellectual autonomy, reflective thinking, self-regulation, argument evaluation, AI-generated content, critical thinking assessment

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.