Technical and pedagogical limitations of artificial intelligence in undergraduate EFL higher education: a systematic review

Synthesis: Espinoza-Celi and Vargas-Saritama ran a PRISMA 2020 qualitative evidence synthesis of the technical and pedagogical limitations of AI-supported EFL instruction in undergraduate higher education, using Scopus as the sole database. The search identified 284 records, 104 were screened after database filtering, 75 full-text reports were assessed for eligibility, and 54 empirical studies published between 2021 and 2026 met the inclusion criteria. Coding of evidence excerpts yielded seven technical limitations in two higher-order categories and eight pedagogical limitations in five higher-order categories; unreliable AI-generated output (22 studies, 41%) and overreliance on AI (27 studies, 50%) were the most frequently reported, and the most common technical–pedagogical pairing — AI system performance limitations with learner cognitive and autonomy risks — appeared in 25 studies (46%). The authors stress that these figures describe patterns of reporting and co-occurrence, not causal relationships.

Key Findings

• Unreliable output and limited contextual understanding dominate the technical evidence. Unreliable AI-generated output was reported in 22 studies (41%) and limited contextual understanding in 16 studies (30%), ahead of prompt sensitivity (8 studies, 15%), speech-recognition limitations (6 studies, 11%), and hallucinations, misinformation, and feedback limitations (4 studies, 7% each). Infrastructure and access constraints were reported in 18 studies (33%).
• Overreliance and reduced critical thinking lead the pedagogical evidence. Overreliance on AI was reported in 27 studies (50%), reduced critical thinking in 25 studies (46%), academic integrity concerns in 20 studies (37%), insufficient teacher guidance in 18 studies (33%), reduced classroom interaction in 11 studies (20%), limited curriculum alignment in 10 studies (19%), cognitive overload and learner confusion in 9 studies (17%) and limited support for higher-order language development in 4 studies (7%).
• Technical and pedagogical limitations were reported together, most often around learner autonomy. AI system performance limitations co-occurred with learner cognitive and autonomy risks in 25 studies (46%), with human interaction and teacher mediation constraints in 16 studies (30%), and with academic integrity and authorship concerns in 12 studies (22%); infrastructure and access constraints co-occurred with learner cognitive and autonomy risks in 13 studies (24%). Three further pairings were reported in 4 to 9 studies.
• The corpus is recent, unevenly distributed, and writing-heavy. Publication years were 2023 (2 studies, 3.7%), 2024 (17, 31.5%), 2025 (28, 51.9%) and 2026 (7, 13.0%); research concentrated in Saudi Arabia (15 studies, 27.8%), Vietnam (8, 14.8%) and Indonesia (5, 9.3%). Undergraduate students were the participants in 39 studies (72.2%), teachers in 10 (18.5%) and mixed samples in 5 (9.3%); designs were mixed methods (28, 51.9%), qualitative (18, 33.3%) and quantitative (8, 14.8%). Studies examined multiple AI tools (27, 50.0%) or ChatGPT (23, 42.6%), and integrated language skills (27, 50.0%) or writing (21, 38.9%) were the most common instructional foci.

What the technical limitations were

The seven technical limitations formed two categories: AI system performance limitations and infrastructure and access constraints. Unreliable output covered inaccurate, inconsistent, or misleading responses that reduced confidence; the authors argue reliability in EFL means more than factual correctness, because a response must also be linguistically accurate, pedagogically relevant, and matched to the learner's proficiency. Limited contextual understanding described difficulty with pragmatic meaning, cultural references, emotional tone, and communicative intent, so output could be grammatically correct yet pragmatically inappropriate — evidence that current systems support controlled, form-focused practice better than culturally nuanced or highly interactive use. Prompt sensitivity shifts the burden onto learners to phrase instructions clearly, disadvantaging those with limited AI Literacy, and speech-recognition limitations constrain speaking instruction and accent variation. Infrastructure constraints — connectivity, hardware, software incompatibility, licensing, and technical support — were reported in 18 studies (33%) and, the authors argue, determine whether learners can benefit from AI-supported instruction at all.

What the pedagogical limitations were

The eight pedagogical limitations formed five higher-order categories. Learner cognitive and autonomy risks was the dominant category: overreliance was tied to reduced self-regulation, weaker independent Problem Solving, and less engagement in language production, and the authors frame it as an instructional-design outcome rather than a fixed learner trait — AI neither promotes nor reduces autonomy by itself. Reduced critical thinking reflected students accepting generated information without evaluating its accuracy, relevance, or appropriateness. Academic integrity concerns covered plagiarism, cheating, ghostwriting, and unclear authorship, and the review argues these are simultaneously pedagogical problems, since they track overreliance, limited AI literacy, and weak teacher guidance. Human interaction and teacher mediation constraints split into insufficient teacher guidance (18 studies, 33%) and reduced classroom interaction (11 studies, 20%), with teacher mediation mattering more precisely because generated responses can be inaccurate. Limited curriculum alignment (10 studies, 19%) covered activities that did not match learning outcomes, proficiency levels, or assessment criteria, and limited support for higher-order language development (4 studies, 7%) captured AI's comparative strength in grammar correction, vocabulary, and controlled practice against weaker support for pragmatic competence, intercultural communication, Creativity, and authentic production. The technical–pedagogical separation follows Technological Pedagogical Content Knowledge (TPACK), though the authors stress the categories emerged inductively from the evidence.

What the co-occurrence patterns suggest

Seven patterns were identified, and the authors read them as association rather than causation. The dominant pairing — AI system performance limitations with learner cognitive and autonomy risks in 25 studies (46%) — is described as paradoxical, since unreliable systems might be expected to reduce trust: students may keep depending on AI when they lack the proficiency, disciplinary knowledge, or AI Literacy needed to spot inaccurate responses. The pairing of system performance limitations with teacher mediation constraints (16 studies, 30%) is read as a continuing need for verification and Scaffolding rather than proof that technical flaws create that need, and infrastructure constraints with learner autonomy risks (13 studies, 24%) shows that connectivity and device access have pedagogical consequences, widening Digital Divide gaps. The remaining pairings ranged from 4 to 9 studies.

What this means for practice

• Teach verification inside the language task, not beside it. Because system performance limitations co-occurred with learner cognitive and autonomy risks in 25 studies (46%), the authors recommend requiring students to evaluate the accuracy and contextual appropriateness of generated content, compare responses, justify revisions, and reflect on their choices rather than accept outputs passively, building AI Literacy alongside language.
• Keep teacher mediation in the design. The pairing of system performance limitations with teacher mediation constraints in 16 studies (30%) supports structured human support: teachers should interpret AI-generated feedback, help learners identify inaccurate responses, and align activities with curricular objectives rather than hand over the feedback loop.
• Treat academic integrity as instruction. With academic integrity concerns co-occurring with system performance limitations in 12 studies (22%), the authors argue for combining policy with explicit guidance on acceptable use, attribution, verification, and responsibility instead of relying on detection or punitive measures.
• Fix access before mandating AI-dependent work. Infrastructure and access constraints co-occurred with learner cognitive and autonomy risks in 13 studies (24%); the review recommends securing reliable platforms, connectivity, and technical support before requiring AI-dependent activities, since unstable access can push students back toward dependence on teachers or peers.

Limitations

• Scopus was the sole database and only English-language open-access studies published between 2021 and 2026 were included, so relevant work indexed elsewhere or published in other languages may be missing; the authors describe the review as a synthesis of Scopus-indexed evidence meeting the eligibility criteria rather than an exhaustive account.
• Heterogeneous designs and AI tools limited direct comparison, so the frequencies are patterns of reporting in the reviewed literature, not estimates of population prevalence, effect size, or causal relationships.
• No subgroup analyses were run, so the review cannot say whether frequencies or co-occurrence patterns varied across countries or instructional contexts; the concentration in Saudi Arabia, Vietnam, and Indonesia is not evidence that limitation profiles are more prevalent there.
• Coding was done by the first author and validated by the second, with no independent duplicate coding and no inter-rater reliability coefficient, so agreement could not be quantified.

Citation

Espinoza-Celi, V., & Vargas-Saritama, A. (2026). Technical and pedagogical limitations of artificial intelligence in undergraduate EFL higher education: a systematic review. Frontiers in Education, 11, 1948091.