AI evidence extraction

Complementary AI in higher education: behavioral, cognitive, and ethical implications of ChatGPT and DeepSeek

Frontiers in Psychology · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

7/10

Relevance

3/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.3389/fpsyg.2026.1699114

Methodology & findings

Study design

Sequential mixed-methods approach with qualitative dominance.

Sample

N = 8, 2 groups

Primary method

Stage One: Qualitative content analysis using ATLAS.ti version 25.0.1 with primary and secondary coding; comparative thematic analysis with repeated comparative analysis of coded segments. Stage Two: Braun and Clarke's (2021) six-stage thematic analysis approach; independent coding by two research assistants with consensus-building for reliability; no statistical hypothesis testing or quantitative pooling performed.

Main result

The study found that "ChatGPT enhances creativity, language learning, and motivational participation, while DeepSeek excels in analytical accuracy and domain-specific performance, particularly in STEM and medical education." Additionally, "Evidence from both phases indicates that their integration underpins a twin-function pedagogy that couples generative fluency and systematic reasoning." However, "threats still exist, such as hallucination-induced errors, academic fraud, and a decline in critical thinking in the absence of metacognitive scaffolding and ethical regulation."

Reports effect sizes.

Research paradigm

Mixed methods (qualitative-dominant); pragmatist/positivist for systematic review phase, interpretive for qualitative phase

Author conclusions

"This mixed-methods overview demonstrates that the integration of ChatGPT and DeepSeek at the tertiary level is more than a technological advance—it is a paradigm-busting change in cognitive, ethical, and pedagogical frameworks for contemporary learning." The authors conclude that "both these generative AI models have complementary strengths: ChatGPT facilitates creativity, responsiveness, and generative fluency, while DeepSeek offers analytical rigor, domain-specific accuracy, and formative reasoning." Finally, "ChatGPT and DeepSeek are not competitive technologies but synergistic cognitive co-agents that can enhance creativity, precision, ethical reasoning, and reflective learning. With responsible application, they will catalyze a wiser, more inclusive, and transfigurative higher education culture."

Risk of bias

Selection bias: Purposive sampling of 8 participants from author's academic network may not represent diverse perspectives; Language bias: Restriction to English-language articles in Western databases excludes Global South scholarship; Indexing bias: Exclusion of non-indexed studies and conference proceedings; Recency bias: Temporal limitation to 2024-2025 excludes earlier foundational research; DeepSeek research immaturity: Limited empirical base for DeepSeek compared to ChatGPT; Author positionality: Researcher is insider within Saudi higher education system, may influence participant selection and interpretation; Selection bias in Stage Two: purposive sampling from author's academic network may not represent broader stakeholder perspectives; Researcher positionality bias: author is insider within Saudi higher education system; Language bias: English-only inclusion criteria excludes non-English scholarship; Database bias: Western-dominated databases (WOS, EBSCO, ProQuest) may under-represent Global South research; Temporal bias: 18-month window (Jan 2024-Jun 2025) may exclude recent developments; Publication bias: systematic review limited to peer-reviewed journal articles and conference papers; Emergent technology bias: DeepSeek literature is nascent, creating asymmetric evidence base between models; Sample demographics bias in Stage Two: all participants are AI experts with 10+ years experience; no undergraduate or non-expert student perspectives; Selection bias: Qualitative sample (n=8) purposively selected from author's academic network, limiting generalizability; Language bias: Only English-language articles included, excluding non-English scholarship particularly from Global South; Database bias: Limited to Western-dominated databases (WOS, EBSCO, ProQuest), excluding region-specific research; Temporal bias: Limited to January 2024-June 2025, potentially missing recent LLM breakthroughs; Emergent status bias: DeepSeek research base less developed than ChatGPT, limiting comparative evidence; Researcher positionality bias: Author is insider to Saudi higher education context, may influence interpretation; Publication bias: Review restricted to peer-reviewed journal articles and conference papers, excluding grey literature

Limitations

The authors state: "(1) Language and Indexing Bias: Only English-language articles indexed in Western-dominated databases (WOS, EBSCO, ProQuest) were included, which may limit cultural representativeness and exclude region-specific or non-English scholarship, particularly from the Global South
(2) Emergent Status of DeepSeek: The availability of longitudinal data is also scarce, and the volume of classroom-based research on DeepSeek remains in development
As such, the findings related to the comparison between DeepSeek should be viewed with caution, pending further empirical substantiation..
(3) Temporal Scope: Research was limited to January 2024–June 2025, potentially excluding more recent breakthroughs in LLM ability (e.g., GPT-4.5, DeepSeek-V3)
(4) Although the qualitative sample under Stage Two (n = 8) was purposively selected by the author from their academic network, this stage was used for depth of analysis rather than for generalizability." Additionally, "the perspectives of younger users, including undergraduate students, should be explored further, as their patterns of reliance, perceived risk, and metacognition may differ from those of professionals."

Open questions raised

Long-term, longitudinal research investigating AI effects on metacognition, critical thinking, and learning achievement
AI systems with embedded self-regulated learning (SRL) competency for real-time goal setting, monitoring, and reflection
Teacher-AI collaboration models for sustained feedback, grading, and one-on-one learning
Extension of research to multilingual and non-Western environments for culturally sensitive AI adoption
Interdisciplinary approaches combining education, cognitive science, behavioral science, data science, and policy design
Research on perspectives of younger users (undergraduate students) regarding AI reliance patterns and metacognition

Data: The authors state: "The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation." (Data Availability Statement); Raw data supporting conclusions stated as available: "The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation." However, no specific URL or repository is provided.; The authors state: "The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation." However, no URL or repository location is specified.Code: Not mentionedExtracted from: pdfAgreement 60%

Explore related topics

AI Literacy Training Doctoral Training Mixed Methods