INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4408
www.rsisinternational.org

ChatGPT in Undergraduate English Language Majors: Benefits and
Challenges for Writing and Speaking Proficiency

Nguyen Song Thao Anh1*, Nguyen Thi Kim Ngoc1, Tran Khanh Bang1, Huynh Thi Kim Ngan1, Phan
Nguyen Yen Quyen1, Nguyen Thi Phuong Hong2

1Student of the High-Quality English Language Program, Cohort 48, School of Foreign Languages, Can
Tho University, Vietnam

2School of Foreign Languages, Can Tho University, Vietnam

* Corresponding Author

DOI: https://dx.doi.org/10.47772/IJRISS.2025.910000364

Received: 20 October 2025; Accepted: 27 October 2025; Published: 12 November 2025

ABSTRACT

Large language models such as ChatGPT are reshaping undergraduate English language education. This
systematic review synthesizes benefits and challenges for developing writing and speaking proficiency among
English language majors. Following PRISMA-2020/PRISMA-S and SPAR-4-SLR, we searched Scopus, Web
of Science, ERIC, EBSCO and SSRN, and analyzed 31 of the 708 empirical and review papers published.
Findings converge on writing gains when ChatGPT is embedded in scaffolded, human-in-the-loop workflows
that emphasize pre-writing, drafting, and revision. Effects are strongest for accuracy, coherence, and argument
quality, and when teacher or peer moderation, prompt scaffolds, and transparent rubrics are present. For
speaking, learners benefit through low-stakes practice, rehearsal and anxiety reduction; however, robust
measurement lags behind, with scarce CEFR-aligned rubrics, limited voice-mode instrumentation, and few
validated acoustic indicators. Integrity and equity remain central. Text-only AI detectors are brittle and
sometimes unfair to non-native writers. Institutions should pivot to process-anchored assessment - combining
prompt logs, version histories, and brief viva voce - to evidence authorship while preserving learning value. A
forward research agenda is proposed, including a minimum reporting toolkit for speaking measurement (CEFR,
ASR features, and ICC/κ) and multi-site randomized trials.

Keywords: assessment integrity, ChatGPT, human-in-the-loop pedagogy, L2 writing, oral proficiency

INTRODUCTION

Large language models (LLMs) such as ChatGPT are accelerating their penetration into higher education,
offering scalable feedback, ubiquitous practice, and personalized scaffolding for undergraduate English language
majors. The applied linguistics literature describes this domain as emergent yet rapidly developing: a systematic
review synthesizing 70 empirical studies in English as a Second/Foreign Language (ESL/EFL) appeared within
18 months of ChatGPT’s release, mapping applications across writing, speaking, assessment, and policy (Lo et
al., 2024). Status-of-the-field essays consistently emphasize the nascency of the domain and call for clearer links
between mechanisms and outcomes (Fang & Han, 2025; Han, 2024).

Writing: Classroom and blended-methods studies indicate that ChatGPT can supplement teachers by providing
formative feedback, aiding ideation, and revision structuring, particularly in human-in-the-loop workflows
(Teng, 2024). In undergraduate writing assessment, ChatGPT’s scoring and qualitative comments show
moderate-to-good consistency with human rater evaluation while scaling up the volume and type of feedback
(Lu et al., 2024). Learner-centric studies report positive perceptions when ChatGPT is positioned as a
"companion," with improvements in accuracy and coherence during drafting/revising (Tram et al., 2024).
However, these benefits are uneven: without prompt scaffolds and rubric-based guidance, students may over-
rely on surface-level edits, with limited transfer to higher-order reasoning skills (Han, 2024; Lo et al., 2024).

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4409
www.rsisinternational.org

Speaking: The (emergent) voice and chat modes expand opportunities for low-stakes practice and can reduce
anxiety (Tram et al., 2024). However, robust measurement of speaking proficiency gains - phonology/intonation,
interactional competence - lags behind writing metrics. Validated CEFR-aligned rubrics and instrumentation
(e.g., Automatic Speech Recognition (ASR) features for speech rate or pause ratio) remain underdeveloped
(Fang & Han, 2025; Lo et al., 2024).

Challenges:

- Integrity and Equity: Academic integrity and the reliability/fairness of AI detection tools remain unresolved.
Emerging evidence suggests that common detectors systematically misclassify non-native writers’ (NNS)
text as AI-generated (Liang et al., 2023).

- LLM Quality and Safety: LLMs can "hallucinate" references, homogenize writing styles, and produce
feedback of variable specificity, reinforcing the need for teacher moderation and transparent AI usage
policies (Lo et al., 2024; Teng, 2024).

- Methodological Gaps: Methodological gaps limit strong inference: many studies are short-term, single-site,
or quasi-experimental, with sparse process data (e.g., keystroke logs, prompt histories) that could illuminate
learning mechanisms (Fang & Han, 2025; Han, 2024).

Review Framework and Research Agenda: Framing the review within the TCCM Framework (Theory - Context
- Characteristics - Methodology) clarifies contributions and gaps (Han, 2024). Contextually, the evidence is
concentrated in Asian higher education and first-year writing, with fewer multi-site trials and speaking
interventions (Fang & Han, 2025; Lo et al., 2024). At the characteristic level, task type, modality, human
coordination, and initial proficiency are likely to moderate effects. Methodologically, longer interventions (~ 12
- 16 weeks), multi-institutional Randomized Controlled Trials (RCTs), and open reporting standards are
priorities (Fang & Han, 2025; Han, 2024).

Against this background, our review synthesizes the benefits and challenges of ChatGPT for writing and
speaking in undergraduate English programs and outlines a forward research agenda. We ask:

What measurable advantages and risks emerge when ChatGPT is integrated into writing and speaking curricula?

Under what pedagogical and assessment conditions are effects enhanced or mitigated?

What design principles and research infrastructure - validated rubrics, process analytics, pre-registration - are
necessary to mature the evidence base?

METHODOLOGY

Design and reporting standards

We conducted a systematic review aligned with the Preferred Reporting Items for Systematic Reviews and Meta-
Analyses (PRISMA-2020) guidelines. To enhance review craft and transparency, we also followed the Scientific
Procedures and Rationales for Systematic Literature Reviews (SPAR-4-SLR) framework and structured data
extraction using the TCCM framework. This approach ensured comprehensive documentation of search
strategies and systematic coding of the evidence base (Page et al., 2021; Paul et al., 2021; Paul & Rosado-
Serrano, 2019; Rethlefsen et al., 2021).

Eligibility criteria

We included empirical studies that met the following criteria:

- Population and intervention: Involved ChatGPT or compatible LLM-based chatbots used by undergraduate
English language majors (or EFL/ESL undergraduates) for writing and/or speaking development.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4410
www.rsisinternational.org

- Outcomes: Reported learning outcomes (e.g., accuracy, fluency, complexity, coherence for writing;
pronunciation/prosody, fluency, interactional competence for speaking), perceptions linked to learning, or
assessment quality.

- Publication type: Were peer-reviewed journal articles, conference papers, or accepted “advance online”
articles published in English.

We excluded kindergarten through 12th grade (K-12) samples, purely conceptual or policy pieces without
empirical data, and studies centered on translation only.

Information sources and search strategy

We searched five core academic databases: Scopus, Web of Science Core Collection (WoS), ERIC
(EBSCOhost), EBSCO Academic Search, and SSRN. Search construction and documentation adhered to the
PRISMA-S guidance (Rethlefsen et al., 2021).

- Scopus (TITLE-ABS-KEY): (chatgpt OR "large language model*" OR "generative ai") AND (EFL OR ESL
OR "english language major*" OR TESOL OR "applied linguistics") AND (writing OR "academic writing"
OR "L2 writing" OR speaking OR pronunciation OR "oral proficiency") AND ("higher education" OR
undergraduate).

- WoS (Topic/TS): TS = (chatgpt OR "large language model*" OR "generative ai") AND TS=(EFL OR ESL
OR "english language major*" OR TESOL OR "applied linguistics") AND TS=(writing OR "academic
writing" OR "L2 writing" OR speaking OR pronunciation OR "oral proficiency") AND TS=("higher
education" OR undergraduate)

- ERIC/EBSCO (TI, AB): (TI chatgpt OR AB chatgpt OR TI "large language model*" OR AB "large language
model*" OR TI "generative ai" OR AB "generative ai") AND (TI EFL OR AB EFL OR TI ESL OR AB ESL
OR TI "english language" OR AB "english language") AND (TI writing OR AB writing OR TI speaking OR
AB speaking OR TI pronunciation OR AB pronunciation) AND (TI undergraduate OR AB undergraduate
OR TI "higher education" OR AB "higher education")

- SSRN (All fields): chatgpt AND (EFL OR ESL OR "english language") AND (writing OR speaking) .

A total of 708 records were identified (676 extracted from search queries and 32 from citation strings), from
which 31 studies were ultimately included for analysis.

RESULTS

Study corpus and designs

Across the screened literature, empirical work on ChatGPT in ESL/EFL higher education has expanded rapidly,
with a field-level review cataloguing 70 empirical studies within 18 months of release and noting concentration
in Asia, first-year writing courses, and mixed small-N classroom designs (Lo et al., 2024). Calls from applied
linguistics and second language acquisition (SLA) emphasize that this growth outpaces rigorous causal designs
and theory-driven mechanism testing, urging systematic research connecting pedagogical mechanisms to
outcomes (Fang & Han, 2025; Han, 2024).

The included corpus spans multiple regions and designs (Table 1). Notably, East Asia and Pacific dominates the
empirical subset, which constrains global generalizability. Therefore, we interpret effects as context-bound and
highlight transferability conditions rather than universal claims.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4411
www.rsisinternational.org

Table 1. Region and context distribution

Source Region Discipline Level Sample size Duration Design

Tsai et al.
(2024)

Taiwan (East
Asia and
Pacific)

Applied
Linguistics /
EFL writing

Undergraduate
English majors

(EFL)

44 students; 44
original + 44

revised essays

Single session
(paired pre-post,

Feb 2023)

Prospective double-
blinded paired-

comparison;
randomized cross-

over grading

Teng
(2024)

Macao SAR,
China (East

Asia and
Pacific)

EFL writing /
perceptions

Undergraduate
EFL learners

≈ 90 survey + 8
interviews

Cross-sectional
survey +

interviews
(course-based)

Mixed methods
(quantitative survey;

qualitative
interviews)

Tram et al.
(2024)

Viet Nam (East
Asia and
Pacific)

EFL self-
directed

learning /
acceptance

Undergraduate
EFL learners

344 survey + 19
interviews; +

SLR (40 studies)

Cross-sectional
survey +

interviews

Multi-methods
(systematic review

+ survey +
interviews)

Tarchi et
al. (2025)

Italy (Europe
and Central

Asia)

Educational
psychology /
source-based

writing

First-year
psychology

undergraduates

27 participants
(M_age≈20.4)

Timed single
session (2023)

Exploratory
experiment +
retrospective
think‑aloud

Uchida
(2024)

Japan (East
Asia and

Pacific) [dataset
multi-Asia]

Assessment /
writing and

speaking
scoring

Not applicable
(secondary

corpus;
ICNALE GRA)

ICNALE GRA:
140 writing +
140 speaking
samples; 80

raters

Not applicable
(secondary
analysis)

Verification study
(correlation/validity

)

Üstünbaş
(2024)

Türkiye (Europe
and Central

Asia)

CALL /
speaking
practice

Pre‑intermediat
e EFL

undergraduates

4 8 weeks Qualitative case
study (screen
recordings +
interviews +

stimulated recall)

Li et al.
(2024)

China (East
Asia and
Pacific)

Assessment /
reliability and

feedback
relevance

Undergraduate
non‑English

majors (CET-4
essays)

30 essays; 4
teacher raters;

ChatGPT-3.5/4

Single class
session (30 min)

G‑theory reliability
+ qualitative

feedback analysis

Lu et al.
(2024)

China (East
Asia and
Pacific)

Assessment /
AI‑assisted

teacher
assessment

Undergraduates
(Chinese
academic
writing)

46
undergraduates

Course‑embedde
d tasks (NR)

Mixed‑methods
(ICC consistency +

interviews)

Cumming
s et al.
(2024)

United States
(North

America)

Composition /
first‑year
writing

Higher
education

(FYW)

Not reported
(conceptual/earl

y classroom
analyses)

Not applicable Analytical/position
paper with early

classroom analyses

Jiang et al.
(2024)

China (dataset
from a

large‑scale
assessment)

Assessment
integrity /

detector bias

Large‑scale
writing

assessment
(operational)

NR (large-scale
operational

dataset)

Not applicable Observational
fairness analysis
(detector bias vs

NNS)

Jiang et al.
(2024)

China (East
Asia and
Pacific)

Assessment
integrity /

keystroke and
fairness

Writing
assessment

(lab/operational
)

NR Not applicable Predictive
modelling of

non‑authentic texts;
subgroup fairness

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4412
www.rsisinternational.org

Liang et
al. (2023)

United States
(North

America)

AI detection /
bias toward

NNS

Not applicable Benchmark
datasets (NR)

Not applicable Algorithmic audit /
bias analysis

Zare et al.
(2025)

Iran/UAE/Oma
n (Middle East)

EFL
argumentativ

e writing /
motivation

Undergraduates
(L2 learners)

≈ 69
(experimental +

control;
mixed‑methods)

Pre‑test →
immediate

post‑test →
1‑month delayed

post‑test

Randomized
pre‑/post‑test

control group +
mixed‑methods

Writing outcomes: performance and assessment

Across undergraduate contexts, writing performance improves when ChatGPT is integrated into drafting and
revision cycles, particularly under human-in-the-loop orchestration. A double-blinded paired-comparison with
EFL English majors found that ChatGPT-assisted revisions significantly increased scores on four dimensions
(largest gains in vocabulary), while also raising fairness concerns because benefits were disproportionately large
for initially weaker writers (Tsai et al., 2024). Reliability studies converge: ChatGPT shows moderate-to-good
consistency with teacher scoring (intraclass correlations), expands feedback volume/types (Lu et al., 2024), and
with G-theory ChatGPT-4 achieved higher reliability coefficients than teachers in CET-4 essay scoring and
provided more relevant qualitative feedback than teachers on language/content/organization (Li et al., 2024).

Learner-centered work complements these findings. Students who position ChatGPT as a “companion” reported
increased accuracy and coherence during drafting/revision, alongside positive perceptions of iterative feedback
(Teng, 2024). For task motivation, experimental evidence indicates that interacting with ChatGPT can raise
motivation to write argumentative essays, though effects may attenuate at delay, signaling novelty and self-
regulation issues (Zare et al., 2024). At the same time, in source-based writing, undergraduates often use
ChatGPT non-strategically and show weaker integration of literal source content; ethical concerns and low
prompting expertise suppress usage (Tarchi et al., 2024). These patterns reaffirm that prompting scaffolds,
rubrics, and teacher moderation are central for higher-order writing development beyond surface correctness
(Cummings et al., 2024; Lo et al., 2024).

A broader meta-analytic lens suggests that carefully designed ChatGPT interventions improve academic
performance (g⁺ ≈ 0.71), higher-order thinking (g⁺ ≈ 0.70), and affective-motivational states (g⁺ ≈ 0.88), while
reducing mental effort (Deng et al., 2025). Heterogeneity remains large, implying sensitivity to subject area,
duration, and application mode, and highlighting a need for moderator-aware designs in L2 writing (Deng et al.,
2025).

Speaking outcomes: practice, affect, and measurement gaps

Evidence for speaking is emergent but promising. Case-based and small-N studies report that students use
ChatGPT to rehearse dialogues, receive corrective feedback, and lower anxiety in low-stakes practice (Üstünbaş,
2024). Reviews and state-of-the-field papers nonetheless stress that oral proficiency measurement lags behind
writing, calling for CEFR-aligned rubrics, ASR-derived features (e.g., speech rate, pause ratio), and stronger
validation (Fang & Han, 2025; Lo et al., 2024). Initial verification work using the International Corpus Network
of Asian Learners of English - Global Rating Archive (ICNALE GRA) corpus finds high correlations between
ChatGPT and human ratings for writing and exploratory alignment for speaking assessment, useful but not a
substitute for validated oral scoring protocols (Uchida, 2024). Overall, practice opportunities scale readily, but
pronunciation/prosody and interactional competence remain under-measured in current trials.

Table 2 synthesizes speaking-assessment practices across the included studies, detailing the speaking task type,
the extent of CEFR alignment, ASR-derived acoustic indicators (e.g., speech rate, articulation rate, pause ratio,
mean length of run, pitch variability), inter-rater reliability (ICC/κ), intervention duration, sample size, and study
design. The evidence indicates substantial fragmentation: CEFR-based rubrics are not widely adopted, ASR
metrics and reliability estimates are reported only in a minority of studies, and designs are frequently short and
under-powered - limitations that reduce cross-study comparability and weaken inference. We therefore propose
a minimum reporting toolkit for future research: (i) standardized CEFR-aligned rubrics; (ii) a core set of ASR

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4413
www.rsisinternational.org

indices capturing fluency and prosody; and (iii) mandatory reporting of ICC/κ to enhance reproducibility and
enable more rigorous evidence synthesis.

Table 2. Speaking measurement practices in the included corpus

Study Region Level Speaking task Rubric Inter‑rater
reliability

Design Duration Notes

Üstünbaş
(2024)

Türkiye
(Europe

and Central
Asia)

Pre‑intermediate
EFL

undergraduates

ChatGPT‑mediated
speaking practice

(guided
conversation)

Not
reported

Qualitative
case study

8 weeks Screen
recordings +
stimulated

recall;
perceived

gains;
limited

measurement
rigor

Uchida
(2024)

Japan
(dataset

multi‑Asia)

N/A (secondary
dataset;

speaking and
writing samples)

Standardized short
responses
(ICNALE)

ICNALE
scoring;
CEFR

mapping
possible

(not
explicit)

80 human
raters;

validity/
consistency

checks

Verification
study

(correlation
between

human and
model

ratings)

N/A Writing–
speaking

rating
comparison;

speaking
correlations
lower than

writing

Table 3. Integrity approaches & fairness considerations

Approach Representative
studies

Evidence
summary

Key risks
(fairness)

Data/infra
requirements

Classroom
feasibility

Implementation
notes

Text‑only AI
detectors

Liang et al. (2023);
Jiang et al. (2024)

Detectors
exhibit
brittleness and
false positives,
especially for
non‑native
writers; high
variance across
prompts/models.

Penalizing
NNS;
domain/topic
bias; version
drift; opacity

Detector access;
reference corpora;
threshold tuning

Low–moderate
(licensing and
false positives
limit adoption)

Use only as triage;
never sole evidence
for misconduct

Keystroke/
behavioral
analytics

Jiang et al. (2024) Temporal/pausa
l dynamics can
flag
non‑authentic
authorship with
better specificity
when validated.

Privacy;
accessibility
accommodat
ions; device
heterogeneit
y

Logging tools;
consent; secure
storage;
calibration by
cohort

Moderate (needs
LMS/editor
integration)

Combine with
reflective artefacts;
clear consent & ethics

Process‑anch
ored
integrity
(recommend
ed)

Cummings et al.
(2024); Lo et al.
(2024)

In‑class drafting
+ version
histories +
prompt logs +
short viva
provides
transparent
authorship
evidence while
maintaining
learning value.

Workload;
training for
staff;
consistency
across
sections

Versioned docs;
prompt logging;
viva rubric

High
(policy‑aligned;
pedagogy‑friend
ly)

Adopt ‘human‑in‑the‑
loop’; publish rubric
and due‑process steps

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4414
www.rsisinternational.org

Integrity, fairness, and process-based assessment

The integrity landscape is complex. Large-scale evidence in computers and education suggests detector behavior
can misclassify texts and raises questions of bias against non-native English speakers in operational settings
(Jiang et al., 2024). Foundational work in Patterns shows that popular detectors systematically flag non-native
speakers (NNS) writing as AI-generated while showing near-perfect accuracy for native samples, underscoring
risks in evaluative settings (Liang et al., 2023). A complementary strand uses keystroke dynamics to detect
nonauthentic texts with attention to fairness across subgroups, suggesting a shift from product-only to process-
based integrity (Jiang et al., 2024). Taken together, institutions are moving towards in-class drafting, oral
verification, and prompt logs/version histories rather than relying on brittle text detectors (Cummings et al.,
2024; Lo et al., 2024).

Table 3 contrasts three families of integrity safeguards for generative-AI use in undergraduate EFL writing/
speaking: (i) text-only AI detectors; (ii) keystroke/behavioral analytics; and (iii) process-anchored integrity (in-
class drafting, version histories/prompt logs, brief viva). For each approach, the table synthesizes the evidence
base, typical failure modes and fairness risks (e.g., false positives for non-native writers), data/infra prerequisites,
classroom feasibility, and implementation notes. The comparison indicates that detector-only solutions remain
brittle and potentially inequitable, behavioral analytics can improve specificity but raise privacy/accessibility
concerns, whereas process-anchored designs provide the strongest, pedagogically aligned evidence of
authorship. Table 3 motivates our stance: use detectors only as triage signals, employ behavioral traces with
explicit consent and safeguards, and adopt process-anchored workflows as the default integrity regime.

Figure 1 illustrates a process-anchored integrity workflow for undergraduate EFL writing/speaking creating a
chain of evidence that authenticates authorship while enhancing equity and learning value.

Figure 1. Process-anchored integrity workflow for undergraduate EFL writing/speaking

Moderators and patterns

Across the included studies, student adoption of ChatGPT was generally positive, with perceived gains in
efficiency and reduced speaking anxiety. Acceptance tended to be higher when instructor-provided prompts and
checkpoints were present, and lower in large-cohort contexts without clear guidance. Instructors reported
benefits for formative feedback and drafting support but raised concerns about over-reliance and authorship
evidence.

Following SWiM guidance for narrative syntheses, we group studies by pre‑specified moderators and report
directionality of patterns (↑ improvement; ↔ mixed/no clear change; ↓ deterioration) rather than pooled effects.
The patterns below are indicative given design heterogeneity and short intervention durations and should guide
future moderator‑aware RCTs and meta‑analyses. These descriptive patterns are summarized here to characterize
the learning contexts; their moderating roles are synthesized in Table 4.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4415
www.rsisinternational.org

Table 4. Moderators and observed patterns across the included studies

Moderator Operationalization Observed pattern Illustrative evidence Implications for
practice

Task type Argumentative/expository;
source‑based synthesis;

narrative/personal
response; speaking
rehearsal/role‑play

Writing: ↑ for
argumentative/

source‑based under
scaffolded prompting;

Narrative/free‑response
↔. Speaking: affect

(anxiety) ↑; proficiency
↔ without rigorous

measurement.

Tarchi et al. (2025)
source‑based writing; Tsai
et al. (2024) argumentative
writing; Üstünbaş (2024)

speaking practice.

Prioritize structured
tasks with explicit

criteria; design
speaking tasks with

measurable sub‑skills
and standardized

rubrics.

Modality and
toolchain

Text‑only chat;
multimodal planning +

drafting; speech practice
with ASR feedback

Text writing ↑
(quality/accuracy);

Multimodal planning ↑ if
human moderation
present; Speaking

proficiency ↔
(measurement limits),

affect ↑.

Li et al. (2024) writing
reliability; Lo et al. (2024)

ESL/EFL overview; Üstünbaş
(2024) speaking.

Integrate ASR metrics
for speaking; couple
chat with versioned

documents for
traceability.

Human
oversight

(human‑in‑the‑
loop)

Unsupervised use; prompt
scaffolds; iterative

feedback with teacher
moderation

With
scaffolds/moderation ↑

(writing quality, process
transparency);

unsupervised ↔/↓
(surface‑level paraphrase,

over‑reliance).

Cummings et al. (2024)
classroom orchestration; Lo

et al. (2024) practice guidance.

Make teacher prompts
and checkpoints
explicit; require

process artefacts to
anchor authorship.

Baseline
proficiency

A2 - B1
(lower‑intermediate); B1 -

B2 (intermediate); C1
(advanced)

B1–B2 learners: ↑ (largest
practical gains). A2: ↔

without heavy
scaffolding. C1: ↔

(diminishing marginal
returns).

Teng (2024) perceptions;
Tram et al. (2024) multi‑

methods; Tsai et al. (2024).

Tailor scaffolds by
proficiency;

pre‑register moderator
analyses by CEFR

bands.

Feedback/
prompting

literacy

No training; brief prompt
hygiene; structured

prompting + feedback
literacy curriculum

Structured prompting +
feedback literacy ↑;
minimal training ↔.

Cummings et al. (2024); Lo
et al. (2024).

Teach prompt patterns
and feedback uptake;

capture logs for
formative analytics.

Dose and
duration

Single session; 4 - 8
weeks; ≥ 12 - 16 weeks

(multi‑site)

Single session ↔
(transient gains). 4 - 8

weeks ↑ modest. ≥ 12 - 16
weeks expected ↑ and
more durable (rare in

current corpus).

Deng et al. (2024)
meta‑analytic lens; Üstünbaş

(2024) 8‑week case.

Plan multi‑week
designs; pre‑specify
retention/ delayed

post‑tests.

Assessment
regime

Detector‑led (text‑only);
keystroke/behavioral

traces; process‑anchored
(versions + logs + viva)

Process‑anchored ↑
(learning + integrity).

Detector‑only ↔/↓
(brittleness, equity risks).

Cummings et al. (2024); Jiang
et al. (2024); Liang et al.

(2023).

Adopt
process‑anchored

integrity by default;
use detectors only as

triage with due
process.

Class size/
instructor
bandwidth

(contextual)

Small seminar; medium
lecture‑lab; large lecture

Smaller classes → ↑ with
closer moderation; large

cohorts ↔ unless
workflows are
streamlined.

Programmatic observations
across classroom studies.

Standardize workflows
(checkpoints, rubrics,

templates) to scale
human‑in‑the‑loop.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4416
www.rsisinternational.org

Taken together, Table 4 suggests that the largest and most defensible gains arise when (i) tasks are structured
and criteria‑rich, (ii) human‑in‑the‑loop scaffolds are explicit, (iii) learners are at B1 - B2 proficiency with
prompt/feedback literacy training, and (iv) integrity is anchored in documented process rather than detector
scores. Conversely, single‑session, unsupervised uses in open‑ended tasks produce at best transient
improvements and weaker authorship evidence. These patterns imply concrete design choices for future trials:
moderator‑stratified sampling, ≥ 12 - 16‑week interventions, delayed post‑tests for retention, and pre‑registered
moderator analyses (e.g., task type × human oversight × proficiency).

Moderators and boundary conditions

Patterns across studies indicate several moderators:

- Task type: benefits are clearest for iterative revision and argumentative writing; source-based synthesis
demands explicit prompting and source-use scaffolds (Tarchi et al., 2024).

- Modality and orchestration: text-only chat supports writing mechanics; voice modes scale rehearsal but
require latency-aware designs and validated oral rubrics (Fang & Han, 2025; Üstünbaş, 2024).

- Human oversight: teacher/peer mediation consistently strengthens learning and mitigates over-reliance (Li
et al., 2024; Lu et al., 2024; Teng, 2024).

- Learner factors: initial proficiency and feedback literacy condition transfer from surface accuracy to
coherence/argument quality (Lo et al., 2024; Zare et al., 2024).

In aggregate, undergraduate EFL writing benefits most when ChatGPT is embedded in scaffolded, monitored
processes that emphasize revision and rhetorical development; speaking benefits are visible in practice and
affect, but robust measurement remains a gap. Integrity and fairness concerns caution against detector-led
policing and favor process-based assessment. Heterogeneity in effects calls for moderator-aware designs and
validated outcome measures moving forward (Deng et al., 2025; Han, 2024; Lo et al., 2024).

DISCUSSION AND LIMITATIONS

This discussion integrates the study corpus to explain when and why ChatGPT assists undergraduate English
majors, and what infrastructural and pedagogical conditions are required for robust, equitable gains. Across the
evidence, writing improvements are most defensible when human moderation and structured task design are
present, whereas speaking benefits are presently concentrated in affect (e.g., reduced anxiety) with limited
proficiency gains due to measurement gaps. Integrity safeguards based solely on text‑only detectors remain
brittle and raise fairness concerns for non‑native writers; a process‑anchored regime is preferable.
Generalizability is constrained by an Asia‑leaning empirical base and the prevalence of short, small‑sample
studies. We therefore frame recommendations on two axes: (i) technical infrastructure and measurement, and
(ii) pedagogical orchestration.

Technical infrastructure and measurement

Standardizing outcome measurement is the most immediate lever for cumulative knowledge. For writing, recent
studies indicate that model‑assisted ratings can approach or exceed the reliability of human raters when anchored
to explicit rubrics and used with human moderation (e.g., Li et al., 2024; Lu et al., 2024). For speaking, however,
few studies report CEFR‑aligned rubrics, acoustic indicators derived from automatic speech recognition (ASR)
or inter‑rater reliability (ICC/κ). We propose a minimum reporting toolkit: CEFR‑aligned rubrics (B1 - C1), a
core ASR bundle for fluency/prosody (speech rate, articulation rate, pause ratio, mean length of run, pitch
variability), and mandatory reporting of inter‑rater reliability. Table 2 summarizes current practice gaps and
motivates this toolkit.

Second, integrity evidence should be grounded in production processes rather than post‑hoc text classification.
Detector‑only approaches are prone to false positives and subgroup disparities for non‑native writers (Liang et

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4417
www.rsisinternational.org

al., 2023; Jiang et al., 2024). By contrast, process‑anchored integrity triangulates in‑class drafting, version
histories with timestamps, prompt/response logs, and a brief viva for a sampled subset; this yields clearer
authorship evidence while preserving learning value (Cummings et al., 2024; Lo et al., 2024). Table 3 and Figure
1 operationalize this workflow.

Third, study design and reporting should reflect contemporary evidence standards. Where meta‑analysis is
infeasible, narrative synthesis should follow SWiM guidance, with pre‑specified groupings (e.g., task type ×
human oversight × proficiency) and clear rules for directionality coding. Risk‑of‑bias appraisals (RoB 2;
ROBINS‑I) should be reported, and screening/reporting should adhere to PRISMA‑2020/PRISMA‑S. Power
analysis, multi‑site trials of ≥ 12 - 16 weeks, and delayed post‑tests are necessary to move beyond transient,
under‑powered findings (see Deng et al., 2024/2025 for a cross‑domain meta‑analytic lens).

Finally, data governance and ethics: behavioral traces (e.g., keystrokes) require informed consent, clear retention
policies, and accessibility accommodations; detectors - if used at all - serve only as triage signals and never as
sole evidence. These safeguards should be codified in course‑level integrity statements and rubric‑linked
due‑process procedures.

Pedagogical orchestration

The largest and most defensible gains arise when ChatGPT is orchestrated as part of a transparent, scaffolded
writing‑and‑speaking workflow. For writing, structured prompts and human‑in‑the‑loop moderation shift usage
from surface paraphrase to higher‑order planning, argumentation, and revision. Evidence from classroom
implementations suggests improvements in accuracy, coherence, and organization under scaffolded prompting,
with teacher checkpoints providing both formative feedback and integrity anchors (Cummings et al., 2024; Tsai
et al., 2024). Source‑based tasks benefit particularly from explicit criteria and staged drafting (Tarchi et al.,
2025).

For speaking, current benefits are strongest for affective outcomes (confidence, reduced anxiety) and for
rehearsal/role‑play contexts, but proficiency effects remain under‑measured. Pedagogical design should
therefore combine low‑stakes practice with measurable sub‑skills (interaction, fluency, pronunciation), integrate
ASR‑supported feedback where feasible, and adopt CEFR‑aligned rubrics with rater calibration. Where class
sizes are large, standardized checkpoints (brief in‑class drafting, versioned submissions, short viva sampling)
help scale human oversight without resorting to punitive detectors (Uchida, 2024; Üstünbaş, 2024).

Learner characteristics and enabling conditions matter. B1 - B2 learners appear to realize the largest practical
gains; A2 learners require heavier scaffolding; advanced learners exhibit diminishing returns without
higher‑order task demands (Teng, 2024; Tram et al., 2024). A brief curriculum in prompting and feedback
literacy improves uptake and mitigates over‑reliance. At the programmed level, departments should standardize
artifacts (prompt templates, version‑control norms, integrity rubrics) to ensure parity of experience across
sections and to support moderator‑aware evaluation (see Table 4).

In short, the pedagogical stance is not “AI or human,” but human‑in‑the‑loop with transparent processes. When
aligned with rigorous measurement and integrity infrastructure, this orchestration can deliver durable writing
gains and a more equitable environment for speaking practice, while enabling cumulative research that is
transferable beyond the current Asia‑leaning evidence base (Cummings et al., 2024; Lo et al., 2024).

Limitations

In this review, we used a comprehensive strategy consistent with PRISMA-2020 and PRISMA-S to document
sources and search strings; however, study-level heterogeneity precluded a pooled effect size, so we adopted a
structured narrative synthesis. Screening and extraction procedures were standardized (e.g., Rayyan) and
quality/risk-of-bias tools were consulted (ROBINS-I; RoB 2), yet residual selection and reporting biases are
possible -particularly where grey literature and non-English outlets may not be fully captured. As in prior reviews
of AI-assisted language learning, construct alignment across studies (task type, proficiency, orchestration,
assessment criteria) remained uneven.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4418
www.rsisinternational.org

For the current evidence base, there are some limits:

- First, the geographical concentration of empirical studies in East Asia and Pacific and allied higher-education
settings constrains external validity; transfer to other regions, program structures, and assessment regimes
should be made cautiously (e.g., China, Japan, Viet Nam).

- Second, many interventions are short-duration (single-session to < 4 weeks) and/or small-sample designs
(case studies, pilot RCTs), which limits statistical power and precision and increases sensitivity to context
and task idiosyncrasies.

- Third, speaking proficiency is under-measured: relatively few studies report CEFR-aligned rubrics, ASR-
based acoustic indicators (e.g., speech rate, articulation rate, pause ratio, mean length of run), or inter-rater
reliability (ICC/κ). Evidence that does triangulate human and model ratings suggests lower correspondence
for speaking than for writing.

- Fourth, the integrity/fairness corpus is still emergent. Text-only AI detectors show brittleness and false-
positive risk for non-native writers; keystroke/behavioral analytics improve specificity but raise privacy and
implementation concerns.

- Finally, the literature remains heterogeneous in task design, “dose” of AI use, and orchestration (human-in-
the-loop vs. fully autonomous), making meta-analytic synthesis challenging; narrative synthesis following
SWiM guidance mitigates but does not remove subjectivity.

CONCLUSION & RECOMMENDATIONS

Across the corpus, the most reliable gains occur in L2 writing when ChatGPT is embedded in iterative
drafting/revision with teacher or peer mediation; effects are weaker or inconsistent for un-scaffolded, single-shot
tasks. Benefits for speaking are visible in increased practice and reduced anxiety, yet the evidence base is
constrained by fragile measurement and limited validation of CEFR-aligned oral rubrics and instrumentation.
System-level concerns about integrity and equity - especially detector misclassification of non-native writing -
argue for process-anchored assessment rather than detector-led policing. Meta-analytic results across disciplines
are encouraging but heterogeneous, underscoring the need for moderator-aware designs. Overall, the field is
promising but under-theorized and under-measured in speaking, with over-representation of short, single-site
studies.

Pedagogical recommendations

Scaffolded prompting & feedback literacy. Treat prompting and feedback use as learnable skills; require brief
rationales for accepting or rejecting AI suggestions, anchored to analytic rubrics.

Human-in-the-loop orchestration. Combine ChatGPT with peer/teacher conferencing; allow convergence of
human and AI judgements while keeping teachers as final arbiters to maintain rhetorical development and
authorial voice.

Process-based assessment for integrity. Replace detector-only policing with in-class drafting, version histories,
prompt logs, and brief oral verification; these approaches both protect equity and preserve learning value.

Research recommendations.

Speaking measurement. Co-develop validated CEFR-aligned oral rubrics and ASR-derived features (speech rate,
pause ratio, prosody) and report rater reliability alongside outcomes.

Causal strength and sustainability. Conduct multi-site RCTs (≥ 12 - 16 weeks), report follow-ups, and publish
process data to connect mechanisms to outcomes.

Transparency and equity. Pre-register protocols, share materials/data, and audit fairness for non-native writers

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4419
www.rsisinternational.org

whenever detection or automation is used.

Table 5 summarizes actionable recommendations by stakeholder, the evidentiary basis, KPIs and timelines for
implementation.

Table 5. Practical recommendations by stakeholder

Stakeholder Practical recommendations Key risks and
mitigation

KPIs / success
indicators

Timeframe Evidence

Instructors
(course level)

• Adopt human‑in‑the‑loop use
of ChatGPT; implement

process‑anchored integrity
(in‑class drafting,

version/prompt logs, short
viva)

• Use scaffolded prompting
templates and require ‘prompt

portfolios’ (students submit
prompts + model outputs +

reflections)
• Writing: grade with

CEFR‑aligned rubrics;
double‑mark 10 - 20% scripts
to report inter‑rater reliability

(target ICC ≥ .75)
• Speaking: add ASR indicators
(e.g., speech rate, articulation
rate, pause ratio, mean length

of run) alongside CEFR
descriptors

• Design AI‑resilient tasks
(local data, personal evidence,

oral defense)

Workload, privacy,
detector false

positives → Use
templates/checklists;

obtain consent;
detectors not sole
evidence; viva as

safety net

• % assignments with
process‑anchored

integrity
• ICC value reported

per course/assessment
• % courses using

CEFR‑aligned rubrics
(writing/speaking)

• # speaking tasks with
ASR indicators;

change in
anxiety/self‑efficacy

Now (pilot
this term)
→ Next

term (scale)

Lo et al. (2024);
Li et al. (2024);
Lu et al. (2024);

Üstünbaş
(2024); Uchida
(2024); Tsai et

al. (2024);
Tarchi et al.

(2025); Liang et
al. (2023); Jiang

et al. (2024)

Assessment and
QA units

(department/
faculty)

• Publish policy: detectors not
used as sole evidence;

due‑process anchored in
process artefacts

• Standardize measurement:
CEFR rubrics

(writing/speaking), ASR
indicator set, mandatory

reliability (ICC/κ) reporting
• Run fairness audits for
subgroup disparities and

document false‑positive rates

Policy acceptance;
staff capacity →

Provide exemplars,
training, phased

roll‑out

• Policy published and
communicated

• % assessments with
CEFR+ASR+ICC

reporting
• Fairness audit
completed per

semester; FP rate
tracked

Now →
Next term

Liang et al.
(2023); Jiang et
al. (2024 JEM;
2024 C&E); Lo

et al. (2024)

Program
administrators

(curriculum
leaders)

• Embed AI‑literacy and
feedback‑literacy outcomes

across writing/speaking
modules

• Provide micro‑PD for
instructors (prompting,

assessment with CEFR/ASR,
fairness)

• Ensure cross‑section
consistency (common rubrics,
shared task bank, moderation)

Fragmentation across
sections → Common

assessment
frameworks;

moderation panels

• % modules with
explicit

AI/feedback‑literacy
LOs

• PD coverage (%
instructors trained)

• Moderation records;
variance of grades

across sections

Next term
→ Next

academic
year

Lo et al. (2024);
Deng et al.

(2025)
(cross‑domain

gains under
structured

orchestration)

Students
(undergraduate

• Complete orientation on
ethical use, prompt design, and

reflective practice

Over‑reliance on AI
→ Reflection
prompts; oral

• Orientation
completion rate;

quality of process

Now Üstünbaş
(2024); Lo et al.
(2024); Uchida

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4420
www.rsisinternational.org

English majors) • Maintain personal ‘process
dossiers’ (drafts, version
histories, prompt logs,

reflections)
• Use ChatGPT to plan/practice
speaking with self‑monitoring
(ASR‑based dashboards when

available)

defense; teacher
feedback cycles

dossiers
• Hours of speaking

practice; improvement
in fluency/prosody

proxies
• Self‑efficacy/anxiety

change (pre–post)

(2024)

EdTech / IT
(infrastructure
and privacy)

• Provide versioned editors and
prompt‑logging integrated with
LMS (opt‑in; consent‑based)

• Enable ASR plug‑ins for
speaking metrics; ensure secure
storage and minimal retention

• Implement consent
management and access

controls; run privacy impact
assessments

Privacy, security,
interoperability →

PIA, DPA
compliance, vendor
due diligence, pilot

before scale

• Tools deployed and
uptime

• % courses using
process logging;

privacy incidents = 0
• Student consent
records; ASR data

coverage

Next term
→ Next

academic
year

Jiang et al.
(2024); Lo et al.

(2024)

Researchers
(faculty/grad

labs)

• Run multi‑site RCTs (≥ 12 -
16 weeks) with pre‑registration;
adopt PRISMA‑2020/SWiM in

syntheses
• Publish measurement toolkits:
CEFR‑aligned rubrics; minimal

ASR indicator set; reliability
reporting templates
• Share data/code

(OSF/Zenodo); audit fairness
across subgroups

Resource intensity
→ Consortia, shared

protocols, seed
grants

• # pre‑registered
RCTs; study duration
• Open materials/data

availability
• Completeness of

measurement
(CEFR+ASR+ICC)

Next
academic

year

Page et al.
(2021);

Campbell et al.
(2020); Deng et

al. (2025)

ACKNOWLEDGMENT

This study was conducted as part of the student scientific research project code THS 2025-15. The authors would
like to express their sincere gratitude to Can Tho University for funding and supporting this study.

REFERENCES

1. Aromataris, E., Lockwood, C., Porritt, K., Pilla, B., & Jordan, Z. (Eds.). (2024). JBI manual for evidence
synthesis. JBI. https://jbi-global-wiki.refined.site/space/MANUAL

2. Borenstein, M., Hedges, L. V., Higgins, J. P. T., & Rothstein, H. R. (2009). Introduction to meta-
analysis. Wiley.

3. Campbell, M., McKenzie, J. E., Sowden, A., Katikireddi, S. V., Brennan, S. E., Ellis, S., Hartmann-
Boyce, J., Ryan, R., Shepperd, S., Thomas, J., & Thomson, H. (2020). Synthesis without meta-analysis
(SWiM) in systematic reviews: Reporting guideline. BMJ, 368, l6890.
https://doi.org/10.1136/bmj.l6890

4. Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological
Measurement, 20(1), 37–46. https://doi.org/10.1177/001316446002000104

5. Cummings, R. E., Monroe, S. M., Watkins, M. (2024). Generative AI in first-year writing: An early
analysis of affordances, limitations, and a framework for the future. Computers and Composition, 71,
102827. https://doi.org/10.1016/j.compcom.2024.102827

6. Deng, R., Jiang, M., Yu, X., Lu, Y., & Liu, S. (2025). Does ChatGPT enhance student learning? A
systematic review and meta-analysis of experimental studies. Computers & Education, 227, 105224.
https://doi.org/10.1016/j.compedu.2024.105224

7. Fang, S., & Han, Z. (2025). On the nascency of ChatGPT in foreign language teaching and learning.
Annual Review of Applied Linguistics, 45, 148–178. https://doi.org/10.1017/S026719052510010X

8. Han, Z. (2024). ChatGPT in and for second language acquisition: A call for systematic research. Studies
in Second Language Acquisition, 46(2), 301–306. https://doi.org/10.1017/S0272263124000111

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue X October 2025

Page 4421
www.rsisinternational.org

9. Jiang, Y., Hao, J., Fauss, M., & Li, C. (2024). Detecting ChatGPT-generated essays in a large-scale
writing assessment: Is there a bias against non-native English speakers? Computers & Education, 217,
105070. https://doi.org/10.1016/j.compedu.2024.105070

10. Jiang, Y., Zhang, M., Hao, J., Deane, P., & Li, C. (2024). Using keystroke behavior patterns to detect
nonauthentic texts in writing assessments: Evaluating the fairness of predictive models. Journal of
Educational Measurement, 61(4), 571–594. https://doi.org/10.1111/jedm.12431

11. Li, J., Huang, J., Wu, W., & Whipple, P. B. (2024). Evaluating the role of ChatGPT in enhancing EFL
writing assessments in classroom settings: A preliminary investigation. Humanities & Social Sciences
Communications, 11, 1268. https://doi.org/10.1057/s41599-024-03755-2

12. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-
native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779

13. Lo, C. K., Yu, P. L. H., Xu, S., Ng, D. T. K., & Jong, M. S. Y. (2024). Exploring the application of
ChatGPT in ESL/EFL education and related research issues: A systematic review of empirical studies.
Smart Learning Environments, 11, 50. https://doi.org/10.1186/s40561-024-00342-5

14. Lu, Q., Yao, Y., Xiao, L., & Yuan, M. (2024). Can ChatGPT effectively complement teacher assessment
of undergraduate students’ academic writing? Assessment & Evaluation in Higher Education. Advance
online publication. https://doi.org/10.1080/02602938.2024.2301722

15. McHugh, M. L. (2012). Interrater reliability: The kappa statistic. Biochemia Medica, 22(3), 276–282.
https://www.biochemia-medica.com/en/journal/22/3/10.11613/BM.2012.031/fullArticle

16. Ouzzani, M., Hammady, H., Fedorowicz, Z., & Elmagarmid, A. (2016). Rayyan—A web and mobile
app for systematic reviews. Systematic Reviews, 5, 210. https://doi.org/10.1186/s13643-016-0384-4

17. Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., … & Moher,
D. (2021). The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ,
372, n71. https://doi.org/10.1136/bmj.n71

18. Paul, J., Lim, W. M., & O’Cass, A. (2021). Scientific procedures and rationales for systematic literature
reviews (SPAR-4-SLR). International Journal of Consumer Studies, 45(6), 1513–1527.
https://doi.org/10.1111/ijcs.12695

19. Paul, J., & Rosado-Serrano, A. (2019). Gradual internationalization vs born-global/ international new
venture models: A review and research agenda. International Marketing Review, 36(6), 830–858.
https://doi.org/10.1108/IMR-10-2018-0280

20. Rethlefsen, M. L., Kirtley, S., Waffenschmidt, S., Ayala, A. P., Moher, D., Page, M. J., & Koffel, J. B.
(2021). PRISMA-S: An extension to the PRISMA statement for reporting literature searches in
systematic reviews. Systematic Reviews, 10, 39. https://doi.org/10.1186/s13643-020-01542-z

21. SCImago (n.d.). SJR—SCImago Journal & Country Rank [Portal]. Retrieved October 12, 2025, from
https://www.scimagojr.com

22. Sterne, J. A. C., Hernán, M. A., Reeves, B. C., Savović, J., Berkman, N. D., Viswanathan, M.,… &
Higgins, J. P. T. (2016). ROBINS-I: A tool for assessing risk of bias in non-randomised studies of
interventions. BMJ, 355, i4919. https://doi.org/10.1136/bmj.i4919

23. Sterne, J. A. C., Savović, J., Page, M. J., Elbers, R. G., Blencowe, N. S., Boutron, I., … & Higgins, J.
P. T. (2019). RoB 2: A revised tool for assessing risk of bias in randomised trials. BMJ, 366, l4898.
https://doi.org/10.1136/bmj.l4898

24. Tarchi, C., Zappoli, A., Casado Ledesma, L., & Wennås Brante, E. (2025). The use of ChatGPT in
source-based writing tasks. International Journal of Artificial Intelligence in Education, 35, 858–878.
https://doi.org/10.1007/s40593-024-00413-1

25. Teng, M. F. (2024). “ChatGPT is the companion, not enemies”: EFL learners’ perceptions and
experiences in using ChatGPT for feedback in writing. Computers & Education: Artificial Intelligence,
7, 100270. https://doi.org/10.1016/j.caeai.2024.100270

26. Tram, N. H. M., Nguyen, T. T., & Tran, C. D. (2024). ChatGPT as a tool for self-learning English among
EFL learners: A multi-methods study. System, 127, 103528.
https://doi.org/10.1016/j.system.2024.103528

27. Tsai, C. Y., Lin, Y. T., & Brown, I. K. (2024). Impacts of ChatGPT-assisted writing for EFL English
majors: Feasibility and challenges. Education and Information Technologies, 29(17), 22427–22445.
https://doi.org/10.1007/s10639-024-12722-y

28. Uchida, S. (2024). Evaluating the accuracy of ChatGPT in assessing writing and speaking: A verification