Research Article
Changes in specific dimensions of English writing anxiety during AI chatbot-supported blended instruction: evidence from Chinese vocational college students
Synthesis: Bao and Zhang (2026) asked which dimensions of second-language writing anxiety an AI chatbot actually moves. In a quasi-experiment, 120 Chinese vocational college students in two intact classes spent eight weeks on English writing, one class with blended instruction built around the Doubao chatbot and the other with conventional instruction, and completed a five-dimension writing anxiety scale before and after; ten experimental students were also interviewed. Somatic anxiety fell further in the chatbot class (F(1, 118) = 14.682, p < 0.001, partial η² = 0.111) and avoidance behavior fell sharply (F(1, 118) = 23.866, p < 0.001, partial η² = 0.168), while test anxiety, self-doubt, and time anxiety showed no between-group difference. Interviews described the chatbot as a non-judgmental companion that students checked and revised against rather than simply copied.
Key Findings
- Two of five anxiety dimensions moved. Somatic anxiety and avoidance behavior fell further in the chatbot class than in the control, while test anxiety, self-doubt, and time anxiety did not differ.
- Test anxiety, self-doubt, and time anxiety did not differ. All three Group × Time interactions were non-significant, with partial eta-squared below 0.001.
- Baselines were equivalent. Pre-test means differed by 0.03 to 0.15 across the five dimensions, with p values from 0.355 to 0.864, so the interaction effects carry the comparison.
- Students described verification, not substitution. Interviewees reported seeking help from the chatbot, checking its suggestions, and revising their drafts during writing rather than after it.
- Writing performance was never measured. The study tracks anxiety only, so it cannot say whether the reductions came with better or worse writing.
Which anxieties a chatbot can reach
The five-dimension split is what makes this study useful, because it shows that "writing anxiety" is not one thing and a chatbot does not reach all of it. Somatic anxiety — the bodily symptoms, racing heart, tension — and avoidance behavior, the tendency to put writing off, are the two dimensions most tied to the experience of sitting down to write. Test anxiety, self-doubt, and time anxiety are tied to evaluation, self-comparison, and deadlines, and those are institutional features a chatbot cannot change. Somatic anxiety fell from 3.59 to 3.04 in the chatbot class against 3.47 to 3.42 in the control (F(1, 118) = 14.682, p < 0.001, partial η² = 0.111), and avoidance fell from 3.60 to 2.83 against 3.57 to 3.53 (F(1, 118) = 23.866, p < 0.001, partial η² = 0.168). Test anxiety, self-doubt, and time anxiety showed no Group × Time interaction, with F values of 0.020, 0.023, and 0.031. The pattern fits the conversational AI literature on low-stakes practice: a tool that never judges removes the social risk of being seen to struggle, which lowers the physiological response and the urge to avoid, but it does not touch the exam that is still coming.
What the interviews add
The qualitative strand is short — ten students from the experimental group — but it names a mechanism the scale cannot. Participants described the chatbot as supportive and non-judgmental, reported reduced emotional pressure, and said they actively checked and revised its suggestions. That last detail matters for the offloading debate: the behavioral description is verification, not delegation, which is the pattern associated with adaptive help-seeking rather than with dependence. It is a self-report and the sample is small, so it establishes plausibility rather than prevalence, but it is consistent with the quantitative finding that the anxiety that fell was the anxiety about being evaluated by a person.
The unmeasured outcome
The study is explicit that writing performance and writing quality were not assessed. That gap bounds what the result can mean. A calmer writer is not automatically a better one, and it is equally possible that lower somatic anxiety and less avoidance came with more writing practice — which would predict improvement — or with more reliance on generated text, which would predict the opposite. The authors also note that the comparison bundles chatbot support with the whole blended condition, including regular classroom instruction and writing activities, so the chatbot's independent contribution cannot be isolated from the redesign around it. Both caveats are stated plainly in the paper and both point the same way: this is a study of affect under a specific instructional arrangement, not of learning.
What this means for practice
- Instructors. Expect a chatbot to reduce the bodily and avoidance symptoms of writing anxiety before it touches evaluative worry; test anxiety and self-doubt need separate instructional attention.
- Language teachers. Teach the checking behavior explicitly, because the students who benefited described verifying and revising against suggestions rather than accepting them.
- Researchers. Pair an anxiety measure with a writing-quality measure; without it, the field cannot tell whether lower anxiety is a route to better writing or a route around it.
Limitations
- Two intact classes were assigned whole to condition, so class, teacher, and schedule are confounded with the intervention.
- Writing performance and writing quality were not assessed, leaving the learning consequence of the anxiety reduction unknown.
- The comparison bundles chatbot support with the whole blended instructional condition, so the chatbot's independent contribution cannot be isolated.
- Eight weeks with 120 students in one Chinese vocational college, and the interviews covered only ten experimental-group students.
Citation
Bao, X., & Zhang, J. (2026). Changes in specific dimensions of English writing anxiety during AI chatbot-supported blended instruction: Evidence from Chinese vocational college students. Frontiers in Psychology, 17, 1962253.