Emerging Educational
Technologies
Emerging
Educational Technologies,2026, 1, 260103 https://doi.org/10.65553/EET.260103
Emotions and
Engagement in Generative AI-Supported L2 Writing: A Systematic Review
Xiaoshu
Xua
aSchool of Foreign Studies, Wenzhou
University, China
Submission date: May 1, 2026; Revision date: October 15, 2026; Online: November 20, 2026
Abstract
Generative AI
increasingly supports L2 writers in planning, drafting, revising, and
responding to feedback, but its value depends partly on how learners feel and
engage while using it. This systematic review synthesized 87 empirical studies
of emotions and engagement in GenAI-supported L2 writing, appraised study
quality, and checked reporting consistency. GenAI often increased enjoyment and
reduced social-evaluative pressure, but did not reliably reduce anxiety and
also introduced concerns about authorship, integrity, and the tool itself.
Engagement dimensions frequently diverged: behavioral or affective engagement
could increase while cognitive engagement remained unchanged or shifted toward
offloading. Positive emotions and favorable appraisals did not ensure deep
engagement, and negative responses could prompt either withdrawal or
verification. Proficiency, AI-output quality and volume, system design, and
teacher mediation shaped these patterns. Across studies, more productive
engagement was associated with learners retaining responsibility for
interpreting, checking, judging, and revising AI output. GenAI may therefore
increase comfort and participation without preserving the evaluation required
for learning.
Keywords: Generative AI, L2 writing, emotions, engagement,
control-value theory, systematic review
Introduction
Generative
AI now takes part in the processes through which student writers plan, draft,
revise, and respond to feedback (Watson Todd, 2026), and
research agendas for L2 writing place it at the center of the field (Hyland, 2026; Kessler, 2026; Khanh-Quynh &
Cong-Lem, 2026). Whether this change supports
writing development depends not only on the texts that learners produce with AI
but also on how they feel while writing with it and how they engage with what
it offers. A learner who enjoys revising with an AI tool but accepts its
suggestions without understanding them may submit a better text while learning
little, and a learner frustrated by output beyond their level may stop using
feedback altogether. Emotions and engagement therefore stand between the use of
generative AI and its effects on writing development.
Research
on these questions has grown quickly, and its conclusions point in different
directions. Some studies report that generative AI raises enjoyment and
engagement (e.g., Jiang et al., 2026; Teng & Huang,
2026), whereas others report that learners copy
AI output, offload the work of writing to the tool, or feel that improved texts
are no longer their own (e.g., Nguyen et al., 2025; S. Wang et al., 2026; Yang et al., 2026a).
Existing reviews of generative AI and L2 writing have mapped the field broadly,
synthesized AI-generated feedback, or estimated effects on writing performance
(Crosthwaite
& Sun, 2026; Feng et al., 2025; Ho et al., 2026; Teng, 2026). With
few exceptions (Crosthwaite & Sun, 2026; Ho et al., 2026), they
have treated emotions and engagement as one strand among perceptions and
outcomes, and those exceptions were limited to feedback. No review has
synthesized studies in which emotions or engagement are focal constructs across
the uses of generative AI in L2 writing, examined how the two relate, or
weighed this evidence by its quality.
This gap
leaves a practical question unanswered: under what conditions does generative
AI leave L2 writers more engaged, and when does it leave them merely more
comfortable? The two outcomes can look alike on a questionnaire while having
different consequences for learning. This distinction is especially important
because GenAI may increase comfort and observable participation while reducing
the evaluative work—interpreting, checking, judging, and revising—that learners
themselves perform. The present review addresses this question with a
systematic review and narrative synthesis of 87 empirical studies. Guided by
control-value theory (Pekrun, 2006) and by models of engagement in
learning and with feedback (Ellis, 2010; Fredricks et al.,
2004), it examines the emotions reported, the
dimensions of engagement, the relations between them, and the conditions under
which they occur. It also appraises each study’s quality and checks the
internal consistency of its reported statistics, so that patterns resting on
sound evidence can be separated from those that do not.
Literature review
Emotions in L2 Writing with Generative AI
This
review draws on control-value theory to define and interpret emotions (Pekrun,
2006). The theory traces achievement emotions, such as enjoyment,
anxiety, and boredom, to two appraisals: the writer’s sense of control over an
activity and its outcomes, and the value the writer attaches to them. Emotions
in turn shape the attention, motivation, and strategies that learners bring to
a task (Pekrun et al., 2002). L2 research has added that enjoyment and
anxiety are distinct dimensions rather than opposite ends of one scale (Dewaele &
MacIntyre, 2014) and that classroom language learning
involves a wider range of emotions than anxiety alone (Shao et al., 2019).
In
generative AI-supported L2 writing, learners’ capacity to evaluate fluent but
sometimes inaccurate AI output (Barrot, 2023) and their sense of authorship over
AI-assisted text may shape appraisals of control and value, and AI use may
introduce new objects of emotion, including the tool and the rules governing
its use. The review therefore coded the object of each emotion and kept
emotions, which are states, distinct from appraisals and beliefs such as
self-efficacy, confidence, trust, perceived usefulness, and tool acceptance.
Foreign
Language Anxiety in EFL Classrooms
Engagement is defined in this review as the
quality of learners’ involvement in a task, with behavioral, cognitive, and
affective dimensions (Fredricks et al., 2004). Behavioral engagement refers to participation and
effort, cognitive engagement to depth of processing and strategy use, and
affective engagement to emotional reactions during involvement. Language
learning research has adopted this model and added a social dimension (Philp
& Duchesne, 2016), but reviews of
the field have found that definitions and measures vary widely and that
self-report dominates (Hiver et al., 2024;
Qian et al., 2026). The review
therefore coded the dimension each study measured and the method it used rather
than accepting the label “engagement” at face value.
L2 writing research has also specified
engagement with respect to its object. Engagement with corrective feedback has
been described in cognitive, behavioral, and affective terms (Ellis, 2010), and it differs
between learners and between teacher and automated feedback (Han
& Hyland, 2015; Zhang
& Hyland, 2018). Two findings from
automated writing evaluation research are especially relevant. Learners could
act on a grammar checker’s suggestions while processing them only shallowly (Koltovskaia,
2020), and trust in automated feedback shaped whether they accepted
or questioned it (Ranalli, 2021).
With generative AI, the object of engagement can be the tool, its feedback, or
a writing task that the tool has partly completed, so the review coded the
object of engagement as well as its dimension.
Generative AI also makes low cognitive
engagement easier to achieve. Cognitive offloading, the use of external
resources to reduce the demands of a task, can free resources for higher-level
work but can also replace the processing on which learning depends (Risko & Gilbert, 2016).
Outside L2 writing, students who used ChatGPT improved their essays without
gaining more knowledge (Fan et al., 2025),
and the quality of students’ peer feedback fell once AI assistance was
withdrawn (Darvishi et al., 2024).
Emotional engagement is also another important issue in human-GenAI
collaborative language learning (Liu & Wang, 2026). The review therefore coded offloading separately from
cognitive engagement and distinguished writing produced with AI from writing
produced without it.
The Relation Between
Emotions and Engagement
Emotions
and engagement are theoretically linked but analytically distinct.
Control-value theory proposes that emotions influence engagement through
attention, motivation, and strategy use and that the results of engagement feed
back into appraisals (Pekrun, 2006), and L2 research has argued that
cognition and emotion cannot be separated in learning (Swain, 2013).
The tripartite model, however, includes affective engagement as a dimension of
engagement, so an emotion measured during a task can be read as either an
emotion or a form of engagement (Fredricks et al., 2004).
The review coded emotions measured as affective engagement as engagement
findings, recorded the emotions they contained, and examined relations in both
directions as a separate question.
Prior Reviews and the Present Review
Reviews of
generative AI and L2 writing have mapped contexts, uses, instructional designs,
and language skills (Feng et al., 2025; Lee et al., 2026;
Luo et al., 2025;
Tang et al., 2026),
synthesized feedback generated by large language models (Crosthwaite &
Sun, 2026; Ho et al., 2026),
and estimated effects on writing performance that depended on studies’ risk of
bias (Teng,
2026). Affective and engagement outcomes
recur across these reviews but are not their central object.
How emotions and
engagement change, relate, and depend on conditions across the uses of
generative AI in L2 writing therefore remains unresolved. The present review
addresses three research questions:
RQ1: What emotions
have L2 writers reported in generative AI-supported writing, what are these
emotions directed at, and how do they compare with those under non-AI
conditions?
RQ2: How do L2
writers engage behaviorally, cognitively, and affectively with generative AI,
AI feedback, and AI-supported writing tasks?
RQ3: How are
emotions and engagement related, under what learner, tool, instructional, and
institutional conditions do these relationships vary, and what evidence links
them to writing outcomes?
Method
Design
This systematic
review with narrative synthesis is reported according to PRISMA 2020 (Page et al., 2021).
A narrative synthesis was chosen over a meta-analysis because the included
studies differed in design, constructs, comparison conditions, and outcomes,
and most were qualitative or mixed-methods studies whose findings could not be
expressed as effect sizes (Popay et al., 2006).
Eligibility Criteria
A
study was eligible if it (E1) was an empirical primary study with human
participants; (E2) involved a generative AI tool, that is, a chatbot or writing
assistant built on a large language model or feedback generated by one, rather
than earlier automated writing evaluation, grammar checkers, or machine
translation alone; (E3) was set in L2 writing, with learners or teachers of a
second or foreign language and writing, writing feedback, or writing
instruction as a focus; (E4) treated emotions or engagement as a focal
construct, named in the aims or research questions or forming a main reported
finding; and (E5) was an English-language journal article published online
between November 30, 2022, the public release of ChatGPT, and October 5, 2026
in one of the 98 journals included in the predefined sampling frame, which was
constructed from SSCI-indexed journals in Education & Educational Research,
Linguistics, or Psychology, Educational. Emotions included discrete emotions
and general affect, and engagement included behavioral, cognitive, and
affective engagement with an AI tool, AI feedback, or a writing task. Studies
that pooled AI and non-AI data, or mentioned emotions or engagement only in
passing, did not meet E4. E4 was met through the aims or research questions in
79 included studies and through a main reported finding in eight.
Information sources and search
The
search, run on October 6, 2026, combined database searches, a
journal-by-journal search within the predefined 98-journal sampling frame, and
the candidate pool of the companion review. Semantic Scholar and ERIC were
searched with four concept blocks combined with AND: generative AI (e.g.,
ChatGPT, large language model), writing (e.g., writing, essay), L2 (e.g., EFL,
second language), and emotion or engagement terms (e.g., engagement, anxiety,
enjoyment). Self-report terms were not required, so that qualitative studies
were retrieved. The journal-by-journal search covered 98 SSCI-indexed journals
in applied linguistics, L2 writing, educational technology, and related areas
and retrieved every article since November 2022 with a generative AI term in
its title or abstract. An OpenAlex search failed because of rate limiting, but
OpenAlex records were represented in the companion pool. Search strings appear
in Supplementary Material S1.
Selection Process
In
the database route, records were de-duplicated, and records outside the date
window or the sampling frame, or lacking a generative AI, writing, L2, or
emotion or engagement term in the title and abstract (the concept rule), were
removed by rule. The remaining records were screened on titles and abstracts,
and full texts of included and unsure records were assessed, with reasons for
exclusion recorded. Title-and-abstract screening and full-text eligibility
assessment were conducted by one reviewer and checked by a second reviewer. Any
disagreements were resolved through discussion until consensus was reached.
Because the concept rule removes any record whose abstract is missing from the
source database, a second route audited all such records that fell within the
date window and sampling frame and had a generative AI term in the title.
Abstracts were retrieved from Semantic Scholar, OpenAlex, Crossref, or
publisher websites, and records were then screened and assessed using the same
review procedure as in the database route.
Data Items
A
structured form developed from the frameworks in the Background recorded study
context, participants, design, AI tools and their roles, the writing task, and
duration. For RQ1, it recorded the emotions studied, their measurement, and
their object; for RQ2, engagement dimensions, objects, and data sources; and
for RQ3, relations between emotions and engagement, associated conditions, and
writing outcomes, including whether outcomes were measured on writing produced
with or without AI. Emotions measured as affective engagement were coded as
engagement findings, and appraisals such as perceived usefulness were coded as
antecedents rather than emotions. Findings were condensed into evidence units,
each checked against the full text. Data extraction and coding were completed
by one reviewer and checked against the full texts by a second reviewer. Coding
disagreements were resolved through discussion and consensus.
Study Quality and Reporting Consistency
Study
quality was appraised with the Mixed Methods Appraisal Tool (MMAT), which suits
reviews combining qualitative, quantitative, and mixed-methods studies (Hong et al., 2018).
Consistent with MMAT guidance, appraisal was interpreted at the criterion level
rather than collapsed into an overall quality score. The synthesis therefore
focused on design-specific methodological strengths and weaknesses rather than
ranking studies by the number of criteria met. Because syntheses in applied
linguistics have documented problems in statistical reporting (Plonsky, 2013),
test statistics were also recomputed from reported descriptives with scripts,
following the logic of GRIM and statcheck (Brown & Heathers, 2017;
Nuijten et al.,
2016). A study was flagged for a
reporting-consistency concern when statistics used in the synthesis could not
be reproduced or were internally impossible, when effect sizes and test
statistics disagreed widely, or when cited references could not be matched to
existing publications. Isolated errors that left conclusions unchanged were not
flagged, and flags do not imply misconduct. Scripts and outputs are provided in
Supplementary Material S3, and the full list of included studies appears in
Supplementary Material S1; included studies cited in the text are marked with
an asterisk in the references.
Synthesis Methods
Following
narrative synthesis guidance (Popay et al., 2006), study
characteristics were tabulated, and evidence units were grouped by research
question and sorted by direction and by the conditions under which findings
occurred. Control-value theory guided the interpretation of emotions, and the
engagement frameworks guided that of engagement. Each synthesis claim was
assigned its supporting studies, whose number, designs, and flags were computed
by script. Study counts were used to describe the distribution of reported
findings rather than as pooled estimates of effect or as a vote-counting rule.
Quality informed the weight given to findings rather than inclusion:
experimental and longitudinal findings were weighted above single-occasion
self-reports, and in a sensitivity analysis every claim was re-examined with
flagged studies removed.
Results
Study Selection
The searches and
the audit together yielded 87 eligible studies (Figure 1). The database and
journal searches identified 33,713 records, of which 176 remained after
rule-based removal and 72 were sought in full text after screening. One was a
duplicate report, and 61 of the 71 assessed met all criteria. The audit of the
concept rule examined 1,414 records that had been removed because their
abstracts were missing. Of these, 114 were screened on abstracts, 41 were
assessed in full text, and 26 met all criteria. Across both routes, 25 of 112
assessed full texts were excluded, most often because emotions or engagement
were not focal (16) or the study was not set in L2 writing (6).
Figure 1. PRISMA 2020 Flow Diagram of Study
Selection
Note. Records removed by
rule were outside the date window (694), lacked a concept term (30,621),
duplicated another record (978), or appeared in a journal outside the sampling
frame (1,244). Full texts were excluded because emotions or engagement were not
focal (database route 6, audit route 10), the study was not set in L2 writing
(1, 5), no generative AI was involved (2, 0), or no participant data were
reported (1, 0).
Characteristics and Quality of the
Included Studies
The included
studies were recent, concentrated in Chinese higher education, and dominated by
feedback uses of ChatGPT (Table 1). The two constructs were rarely studied
together: emotions were focal in 44 studies, engagement in 58, and both in only
15. Emotions were measured mainly through interviews and rating scales, and
engagement through questionnaires, interviews, revision analysis, and log or
screen-recording data.
Table 1. Characteristics of the 87 Included
Studies
|
Characteristic
|
Category
|
k
|
|
Publication
year
|
2023 / 2024 / 2025 / 2026
|
1 / 6 / 28 / 52
|
|
Region
|
Mainland China, Hong Kong, Macao
|
56
|
|
|
Iran
|
7
|
|
|
Other (13 countries or multi-country)
|
24
|
|
Target
language
|
English
|
80
|
|
|
German, Chinese, Japanese, French, English and Spanish
|
7
|
|
Participants
|
University students
|
80
|
|
|
School students
|
5
|
|
|
Teachers
|
2
|
|
Design
(MMAT category)
|
Mixed methods
|
45
|
|
|
Qualitative
|
17
|
|
|
Quantitative non-randomized
|
15
|
|
|
Quantitative randomized
|
8
|
|
|
Quantitative descriptive
|
2
|
|
Role of
generative AIa
|
Feedback
|
72
|
|
|
Revision or editing
|
36
|
|
|
Language support
|
30
|
|
|
Idea generation
|
27
|
|
|
Drafting or text generation
|
12
|
|
Engagement
dimensiona
|
Behavioral
|
61
|
|
|
Cognitive
|
57
|
|
|
Affective
|
39
|
|
Construct
focus
|
Emotions focal
|
44
|
|
|
Engagement focal
|
58
|
|
|
Both focal
|
15
|
Note. k = number of studies.
a Studies could be
coded in more than one category.
The methodological
quality of the evidence base was uneven. The most common concerns differed by
design. Quantitative and mixed-methods studies were frequently limited by
convenience sampling, intact-class comparisons without adequate control of
baseline differences, unvalidated measures, and weak integration of qualitative
and quantitative strands. Qualitative studies generally showed fewer concerns
on the MMAT criteria, although the depth of sampling and analytic reporting
varied across studies. Twenty-two studies, all quantitative or mixed-methods,
were flagged for reporting-consistency concerns.
Emotions (RQ1)
Comparative studies
more often found higher enjoyment with generative AI than no difference, but
the differences were small. Nine studies reported more enjoyment or positive
affect with generative AI than with peer, teacher, automated, or no-AI
conditions, including a randomized study in which enjoyment was higher across
four writing tasks (d = 0.51; Jiang et al., 2026). Five studies found no
difference or lower enjoyment, including a ten-week comparison with teacher
feedback (S.
Wang et al., 2026).
Comparative
evidence on anxiety was mixed and rested on weaker designs. Six studies
reported lower anxiety with generative AI, but three of them were flagged. Four
found no difference or higher anxiety; in one, anxiety rose over ten weeks with
AI feedback but stayed stable with teacher feedback (S. Wang et al., 2026).
Differences concentrated in demanding tasks: in one randomized study, the AI
group reported more enjoyment and less anxiety in argumentative but not
narrative writing (Jiang et al., 2026).
Positive and
negative emotions commonly co-occurred. Twenty-two studies reported both among
the same participants: relief when the AI fixed or explained problems and
frustration or confusion when its output was vague, inaccurate, or beyond
learners’ level. Favorable appraisals, such as greater confidence, sometimes
accompanied these positive emotional responses (e.g., Koltovskaia et al., 2024;
Zhao & Wang,
2026). The same tool could ease anxiety in
the moment and create it later, as when reliance on AI proofreading raised
anxiety about AI-free tests (W. Yang & Lin, 2025). Emotions also
differed across learners and changed over time. In a study of learners of
Chinese, high-proficiency learners moved from calm to happiness after ChatGPT
feedback, whereas low-proficiency learners’ anxiety persisted (Zhao & Wang,
2026), and in a nine-week study emotions
moved from positive to negative and back as students discovered fabricated
output and then used the tool selectively (Zhang & Liu, 2026).
Ten studies
reported emotions directed at the AI or the rules surrounding it rather than at
the writing task. These included worry about falling behind peers’ AI use (Wang & Zou,
2026), guilt and fear of detection (e.g., Du & Yang, 2025;
Yao & Liu, 2026),
and unease about authorship when the AI rewrote students’ paragraphs (Park
& Ryoo, 2026). Thesis writers described distress when their self-written
text was flagged by AI detectors (Ardana & Christiana, 2026). The two studies
of teachers found that teachers’ emotions followed their reading of students’
AI use, ranging from pride when students engaged critically (Li et al., 2026)
to exhaustion and anger when judging suspected AI-generated assignments (Khojasteh et al.,
2025).
Engagement (R2)
Behavioral
engagement with AI feedback was partial and favored local suggestions. In eight
of the 11 studies reporting uptake rates, learners acted on about half to
two-thirds of the AI suggestions they received. Eight studies found higher
uptake for local, form-focused, or specific feedback than for global or content
feedback; in one case study, 67.6% of local and 34.0% of global suggestions
were taken up (Ma et al., 2026). The design of the system
also mattered. A system that gave diagnostic feedback without rewriting
produced a median of six revision cycles, against two with free chatbot use (Park & Ryoo,
2026), although young learners' use of a
feedback chatbot declined across tasks (Wolf et al., 2026).
Cognitive
engagement ranged from wholesale acceptance to evaluative integration, and the
range was patterned by proficiency. In 14 studies, higher-proficiency learners
verified, adapted,
or rejected AI
output, whereas lower-proficiency learners accepted suggestions without
understanding them or pasted AI text verbatim (e.g., Lu & Jia, 2026;
Rong et al., 2025).
Advanced graduate writers rejected 30 of 31 inaccurate ChatGPT comments (Koltovskaia
et al., 2024), whereas learners without support understood fewer
than half of the suggestions they handled (Almutlaq & Alsaleh, 2026).
Twelve studies reported cognitive offloading. In one planning task, 55 of 56
learners asked ChatGPT to write the whole paragraph (Nguyen et al., 2025).
Five studies, however, reported a shift over weeks from wholesale acceptance
toward questioning and selective use (e.g., Mohamed, 2026;
Zhang & Liu,
2026).
The three
dimensions of engagement did not move together. In seven studies, generative AI
raised one dimension while leaving another unchanged or lower. In a
class-randomized study with primary school students, behavioral and emotional
engagement rose while cognitive engagement did not (Lu et al., 2026);
in a randomized study of group writing, analytical talk rose while behavioral
and emotional engagement did not (Hu et al., 2025). Only two studies
reported gains in all three dimensions (Qi et al., 2026;
Teng & Huang,
2026). Self-reported and observed engagement
also diverged in two studies (Wang et al., 2026; Yang et al., 2026b).
Relations and Conditions (R3)
Relations Between Emotions and Engagement
Few studies tested
the relation between emotions and engagement statistically. All four that did
found that anxiety was associated with lower engagement (e.g., r =
−.32 to −.15; Dong, 2025), and in one of them enjoyment predicted engagement
positively but more weakly than anxiety predicted it negatively (Wang
& Zou, 2026).
Positive emotions
and favorable appraisals did not ensure deep engagement. Five studies,
including three qualitative case studies with no major concerns on the
applicable MMAT criteria, reported enjoyment or satisfaction, together with
favorable appraisals such as trust, alongside copying or uncritical acceptance
of AI output (e.g., Cengiz et al., 2025; Yeung, 2025;
Zhang & Liu,
2026). Negative emotions and unfavorable
appraisals also led in two directions. Seven studies reported that frustration
with feedback learners could not understand led them to skip it or avoid the
tool (e.g., Takahashi & Kadowaki, 2026; Zhao & Wang,
2026), whereas six reported that distrust or
disappointment prompted verification and cross-checking (e.g., Ma et al., 2026;
Shi & Wang,
2026). In two studies, critical use of AI
output was itself a source of positive emotion (Du & Yang, 2025;
Zhong et al., 2026).
Learner, Tool, Instructional, and Institutional
Conditions
Proficiency and
literacy conditioned how learners felt about and used AI output. Seventeen
studies reported proficiency differences in the direction described under RQ2,
and two quantitative studies reported proficiency as a moderator (Abdelhadi, 2026;
Sun & Jordan,
2026). Ten studies associated AI literacy,
feedback literacy, or metacognition with more productive use; iterative,
specific prompting yielded actionable feedback, whereas single-turn prompts
yielded generic feedback and verbatim acceptance (e.g., Du & Yang, 2025;
Meng et al., 2026).
The volume and
accuracy of AI output conditioned negative emotions and disengagement. Fifteen
studies linked long, numerous, or advanced feedback to overload and unprocessed
suggestions (e.g., Liu et al., 2026), and 16 linked inaccurate
or generic output to lost trust, rejection, or verification. Ten studies
associated designs that restricted what the AI did or displayed its reasoning
with more effortful engagement. Such designs withheld rewrites (Park & Ryoo,
2026), required learners to propose repairs
first (Mohamed,
2026), or gave clues rather than corrections
(Lukešová &
Jennings, 2026), although most did not isolate the
design feature experimentally.
The social position
of the AI and the surrounding instruction also mattered. Fourteen studies
reported relief because learners could ask questions and receive criticism
without the evaluative pressure of a teacher or peer (e.g., Yuan & Dao,
2026; Zheng et al., 2026).
Seven studies associated teacher mediation with better understanding and more
positive emotions; in one randomized but flagged study, teacher explanation
raised understanding of ChatGPT feedback from 57.0% to 87.7% of suggestions (Almutlaq
& Alsaleh, 2026). Ten studies reported that institutional
integrity rules and detection practices generated anxiety, and detection-score
thresholds amplified it, whereas checking results together with supervisors
contained it (Ardana & Christiana, 2026).
Writing Outcomes
Evidence on writing
outcomes was limited and weaker than evidence on experience. Twelve studies
compared writing produced without AI across conditions and found gains mostly
in language rather than in content or organization (e.g., Lu et al., 2026;
Zheng et al., 2026),
but six were flagged, including the three reporting the largest gains. In five
further studies, the scored text had been produced with the AI. Six studies
related engagement to outcomes, two of them randomized. In the flagged one,
cognitive and behavioral engagement, but not affective engagement, carried part
of the treatment effect on writing growth (Sun & Jordan, 2026);
in the unflagged one, engagement did not predict writing gains once group
membership was modeled (Teng & Huang, 2026).
Sensitivity to Flagged Studies
Most findings
rested mainly on unflagged studies, but the quantitative evidence on anxiety
and on writing outcomes did not (Table 2). Patterns grounded in qualitative and
process data lost few or no studies when the 22 flagged studies were removed.
In contrast, removing flagged studies left the evidence on anxiety balanced
between lower, unchanged, and higher anxiety, and halved the evidence comparing
AI-free writing across conditions.
Table 2. Summary of Evidence for the Main
Synthesis Findings
|
RQ
|
Finding
|
k
|
RCT
|
Other
|
Qual.
|
Flagged
|
k
excl. flagged
|
|
RQ1
|
Higher enjoyment or positive affect than a comparison condition
|
9
|
3
|
6
|
0
|
1
|
8
|
|
RQ1
|
No difference or lower enjoyment than a comparison condition
|
5
|
1
|
4
|
0
|
0
|
5
|
|
RQ1
|
Lower anxiety than a comparison condition
|
6
|
1
|
5
|
0
|
3
|
3
|
|
RQ1
|
No difference or higher anxiety than a comparison condition
|
4
|
1
|
3
|
0
|
1
|
3
|
|
RQ1
|
Positive and negative emotions co-occurring
|
22
|
0
|
12
|
10
|
4
|
18
|
|
RQ1
|
Emotions directed at the AI or its rules
|
10
|
0
|
6
|
4
|
0
|
10
|
|
RQ2
|
Uptake favoring local, specific feedback
|
8
|
0
|
4
|
4
|
1
|
7
|
|
RQ2
|
Range from wholesale acceptance to evaluative use, patterned by
proficiency
|
14
|
0
|
9
|
5
|
3
|
11
|
|
RQ2
|
Cognitive offloading
|
12
|
0
|
9
|
3
|
2
|
10
|
|
RQ2
|
Dissociation among engagement dimensions
|
7
|
1
|
6
|
0
|
0
|
7
|
|
RQ3
|
Anxiety associated with lower engagement (quantitative)
|
4
|
0
|
4
|
0
|
2
|
2
|
|
RQ3
|
Positive emotions coexisting with shallow engagement
|
5
|
0
|
2
|
3
|
0
|
5
|
|
RQ3
|
Negative emotions leading to withdrawal
|
7
|
0
|
3
|
4
|
1
|
6
|
|
RQ3
|
Negative emotions leading to verification
|
6
|
0
|
3
|
3
|
2
|
4
|
|
RQ3
|
Proficiency differences
|
17
|
1
|
11
|
5
|
3
|
14
|
|
RQ3
|
Overload from volume or complexity of AI output
|
15
|
0
|
11
|
4
|
3
|
12
|
|
RQ3
|
Inaccurate or generic output undermining trust
|
16
|
0
|
9
|
7
|
3
|
13
|
|
RQ3
|
Relief from social-evaluative pressure
|
14
|
0
|
12
|
2
|
2
|
12
|
|
RQ3
|
Restrictive or explanatory AI designs with more effortful engagement
|
10
|
2
|
7
|
1
|
3
|
7
|
|
RQ3
|
Teacher mediation
|
7
|
1
|
5
|
1
|
1
|
6
|
|
RQ3
|
Integrity rules and detection as sources of anxiety
|
10
|
0
|
5
|
5
|
0
|
10
|
|
RQ3
|
AI-free writing compared across conditions
|
12
|
5
|
7
|
0
|
6
|
6
|
|
RQ3
|
Engagement related to writing outcomes
|
6
|
2
|
4
|
0
|
4
|
2
|
Note. k = number of studies; RCT =
randomized studies; Other = non-randomized, descriptive, and mixed-methods
studies; Qual. = qualitative studies; Flagged = studies with
reporting-consistency concerns (see Method). Study lists are in Supplementary
Material S3. Study counts indicate the distribution of reported findings and
should not be interpreted as pooled effect estimates or as a vote-counting
rule.
Discussion
This review
synthesized 87 studies of emotions and engagement in generative AI-supported L2
writing and reached three main conclusions. Generative AI changed not only how
much enjoyment or anxiety L2 writers reported but also what their emotions were
about. Engagement did not rise as a whole, and cognitive engagement was the
dimension most likely to stay flat or give way to offloading. More productive
forms of AI-supported engagement tended to involve learners, rather than the
tool, interpreting, checking, and revising AI output. The sections below
interpret these findings through the frameworks introduced in the Background.
Generative AI Extends the Objects of Emotion
in L2 Writing
The emotional
findings support the prediction of control-value theory that emotions follow
appraisals of control and value (Pekrun, 2006), and they show how
generative AI reshapes both appraisals in L2 writing. Control depended on the
match between the learner and the output rather than on the tool. Help on
demand and the AI’s position outside human evaluation raised learners’ sense of
control, whereas output that learners could not understand or check lowered it,
especially at lower proficiency. Value depended on ownership. Improvement
achieved with AI did not always feel like the learner’s own, so AI assistance
could raise the quality of a text while lowering the pride attached to it.
The theory’s
distinction between activity and outcome emotions does not, by itself, fully
capture the range of emotion objects identified in this synthesis. Guilt about
reliance, worry about detection, and unease about authorship have the AI and
its governance as their objects. One included study proposed adding appraisals
of the tool to the theory (Zhong et al., 2026); the present synthesis
supports this proposal and adds institutional rules as a third object, which
generated anxiety for students and teachers alike. The uneven evidence on
enjoyment and anxiety also fits the view that the two are distinct dimensions (Dewaele &
MacIntyre, 2014). A tool can make writing more pleasant
without making its outcome less uncertain, and uncertainty about outcomes is
what the theory places at the root of anxiety.
Engagement Dimensions Decouple
The dissociation
among engagement dimensions is the most consequential finding of this review.
In the two strongest designs in the corpus, behavioral and emotional engagement
rose without cognitive engagement, or cognitive talk rose without behavioral or
emotional engagement. Earlier research on automated writing evaluation found
that learners could act on a grammar checker’s suggestions while processing
them only shallowly (Koltovskaia, 2020). The present synthesis
shows how generative AI extends this concern from accepting individual
suggestions to offloading larger portions of the writing process, as when
nearly every learner in
one planning task
asked the AI to write the whole paragraph. In L2 writing, the evidence that AI
assistance can improve products without improving learning (Fan et al., 2025)
and can weaken performance once withdrawn (Darvishi et al., 2024)
thus takes the form of offloading during revision and planning (Risko &
Gilbert, 2016).
The decoupling has
a methodological consequence. Behavioral participation, the most easily
observed and most often measured dimension, may be a less informative indicator
of cognitive engagement in generative AI-supported writing, because accepting a
suggestion need not involve evaluating it. A questionnaire item about revising
after feedback cannot distinguish a learner who revised by reasoning from one
who pasted the AI’s revision. This sharpens concerns about inconsistent
definitions and reliance on self-report in engagement research (Hiver et al., 2024;
Qian et al., 2026),
and it is consistent with the divergence between self-reported and observed
engagement in the present corpus.
The relation
between emotions and engagement points the same way. Control-value theory
predicts that positive activating emotions support engagement (Pekrun,
2006), yet positive emotions and favorable appraisals, including
enjoyment, satisfaction, and trust, coexisted with copying and uncritical
acceptance. One possible explanation lies in what the emotion or appraisal
concerns. When enjoyment arose from the AI’s output rather than the learner’s
own problem solving, it may have sustained use of the tool rather than
processing of the task. Trust, as an appraisal of the tool or its feedback,
worked similarly. Earlier research found that trust shaped whether learners accepted
automated feedback (Ranalli, 2021); here, trust in fluent output went with
acceptance without verification, whereas distrust prompted scrutiny. Emotional
valence alone therefore provides limited insight into engagement unless it is
considered alongside learners’ appraisals and their capacity to evaluate and
act on AI output.
Productive Support Leaves the Evaluative Work
to the Learner
The conditions
identified for RQ3 suggest that AI support was more productive when learners
could do the evaluative work themselves. Proficiency, AI literacy, and
metacognition allowed learners to verify, reject, and adapt output, whereas low
proficiency led to acceptance, skipping, or confusion. Research on teacher and
automated feedback found that engagement depended on individual and contextual
factors (Han
& Hyland, 2015; Zhang & Hyland, 2018);
generative AI raises the threshold, because learners must judge not only
whether a suggestion is correct but whether the AI understood their intent.
Restrictive or explanatory AI designs and teacher mediation may have helped
learners meet this higher threshold by supporting rather than replacing their
evaluative work. This interpretation extends the view that feedback literacy
involves appreciating, judging, and acting on feedback and managing the
emotions it provokes (Carless & Boud, 2018).
Taken together,
these findings suggest that generative AI complicates the interpretation of
familiar indicators of affect, engagement, and writing development. Enjoyment
may reflect a more comfortable writing experience rather than deeper learning;
feedback uptake may reflect acceptance rather than reasoning; and high scores
for AI-assisted texts may reflect the tool's contribution rather than
independent writing development. One potentially important distinction is
whether learners retain responsibility for evaluating AI output. Although this
interpretation has not been tested as a causal mechanism, it shifts attention
from how much learners use AI to what cognitive work they continue to perform.
The evidence behind
these conclusions is uneven. Patterns grounded in qualitative and process data
were robust to the removal of flagged studies, whereas comparative evidence on
anxiety and evidence on writing outcomes were not. Claims that generative AI
reduces writing anxiety or improves AI-free writing should therefore be treated
as provisional.
Implications for Writing Instruction
Three principles for instruction follow
from the synthesis, although they are inferences from patterns across studies
rather than tested interventions. First, teachers can preserve the evaluative
work of writing by asking AI tools for diagnosis, questions, or clues rather
than rewrites, and by asking learners to propose their own revisions before
consulting the AI. Second, teachers can match AI output to learners’ processing
capacity by limiting the volume of feedback and explaining it for lower-proficiency
learners, since overload led to skipped feedback and frustration. Third,
teachers can assess engagement through learners’ decisions and their reasons,
such as which suggestions they accepted or rejected and why, rather than
through the amount of AI use or learners’ satisfaction with it, and can use
occasional writing without AI to gauge independent performance. Suspected
misuse is better handled through dialogue than through detection scores alone,
given the anxiety that detection practices generated.
Conclusion
This systematic
review examined emotions and engagement in 87 empirical studies of generative
AI-supported L2 writing. The evidence suggests that AI can make writing more
enjoyable and reduce social-evaluative pressure without necessarily increasing
cognitive engagement, while its effects on anxiety and independent writing
development remain less certain. By separating what L2 writers feel, which
dimensions of engagement change, and who performs the evaluative work, the
review shows that GenAI may increase comfort and observable participation
without preserving the evaluative work required for learning. AI-supported
writing should therefore be judged not only by learners’ satisfaction or the
amount of AI use, but also by whether learners continue to interpret, check,
judge, and revise AI output themselves.
This review has
several limitations. The restriction to English-language articles within a
predefined 98-journal SSCI sampling frame limits the comprehensiveness and
generalizability of the review, and the concentration of studies in Chinese
higher education further limits generalization. The use of single-session
designs in many studies leaves longer-term patterns unclear. Although an audit
recovered 26 studies that rule-based filtering had removed, eligible studies
without abstracts or without generative AI terms in their titles may still have
been missed. Heterogeneity in designs and measures precluded quantitative
pooling.
Future research
should combine self-reports with process data, examine emotions and engagement
longitudinally and in both directions, and assess writing development on texts
produced without AI. Teachers’ emotions, examined in only two studies, also
warrant attention. Most importantly, future studies should examine whether
learners’ evaluative work helps explain the relationship between AI support and
subsequent independent writing development.
Acknowledgement
We gratefully acknowledge the anonymous reviewers for their diligent
efforts in reviewing this study.
Declaration
of Interest Statement
There is no potential conflict
of interest.
References
References marked with an asterisk indicate studies
included in the review that are cited in the text. The full list of the 87
included studies is provided in Supplementary Material S1.
*Almutlaq, S. A., &
Alsaleh, Z. M. (2026). Does metalinguistic explanation of teacher and ChatGPT
feedback improve EFL learners’ writing quality and engagement? Findings from an
intervention study. Assessment &
Evaluation in Higher Education, 51(3),
591–609. https://doi.org/10.1080/02602938.2025.2586124
Author Bios
Xiaoshu Xu is professor at Wenzhou University, PhD supervisor at
Macao Polytechnic University and Stamford International University (Thailand),
researcher of Macau Global Education Research and Innovation Center at City
University of Macau. As Executive Editor-in-Chief of JETI, her research focuses
on personalized learning environments, teacher professional development, and
EFL teaching practices.
E-mail: Lisaxu@wzu.edu.cn
ORCiD: https://orcid.org/0000-0002-0667-4511