Three Students, Three Emotional Paths Through AI-Assisted Writing

When a Chinese university embedded generative artificial intelligence into a 14-week English writing course, three students who sat through the same lectures, used the same tools, and completed the same assignments came out with radically different emotional experiences. One grew more motivated with every correction the machine delivered. Another began by outsourcing her thinking to the chatbot and ended as a skeptical, strategic user. The third, overwhelmed and under-supported, eventually walked out of the classroom rather than face an in-class writing task. A new longitudinal multiple-case study published in Frontiers in Psychology documents these divergent trajectories and argues that the emotional impact of AI-assisted writing depends less on the technology itself than on the fit between each learner and the support surrounding them.

The study, conducted by Keqing Wu, Miao Yang, and Changmin Lin of Shantou University Medical College, followed three first-year medical students—pseudonymized as Leo, Grace, and Sarah—through a general English writing course that marked the institution’s first integration of generative AI into writing instruction. The course covered three genres across the semester: an individual exposition, a group feature article, and an individual argumentation essay. Students used AI tools for brainstorming, outlining, and feedback during pre-writing and revision stages, while first drafts of the individual tasks were handwritten in class without electronic devices. The university had built its own institutional AI platform and a tool called the 24H Intelligent Assistant, ran training workshops for instructors, and required teaching teams to formulate course-specific AI use policies.

The researchers grounded their analysis in a sociocultural framework of second-language student engagement with written feedback, originally developed by Mao and Lee in 2022. The framework distinguishes individual mediating factors—lived experiences such as mindsets, beliefs, and personal values, plus developmental competencies like cognitive skills and emotion regulation—from environmental factors spanning the specific learning tasks, the classroom setting, and the broader school environment. Drawing on Vygotskian concepts including mediation, the zone of proximal development, and perezhivanie, the notion of lived experience as the prism through which environmental factors are refracted, the framework treats engagement as socially situated and evolving rather than as a fixed learner trait.

Data collection was extensive. Following a pre-course survey of 241 students across seven classes, 14 participants were purposively selected, and three focal cases were chosen for their varied proficiency levels, gender, prior AI experience, and unusually rich data. Each focal student completed three rounds of semi-structured interviews with stimulated recall, totaling seven hours and five minutes, alongside reflective reports, complete chatlogs of their interactions with AI tools, essay outlines, first and final drafts, informal chats with the researchers, and the course syllabus. Transcriptions were machine-translated and then cross-checked, with back-translation of a sample reviewed by an experienced English teacher to protect the emotional nuance of idiomatic speech. Two authors coded independently, actively seeking disconfirming evidence to guard against the first author’s prior knowledge of the students.

Leo, an advanced-level student who had used ChatGPT for English writing since the previous semester, followed what the authors call a positive-resilient trajectory. In the first stage, he was excited by the formal introduction of AI assistance, selected his own topic, and drafted his outline independently before turning to the tool for examples. But when the AI polished his draft, it acted, in the study’s phrase, like a merciless mirror, exposing numerous grammatical errors and unnatural expressions. For the first time, this confident learner felt frustration and self-doubt, admitting he was now afraid to write without the tool. A teacher-assigned score of 95 on his first draft, far above his expectations, restored his belief in his own writing competence. By the final argumentation task, he reported that AI had boosted his enthusiasm for learning English, creating what he described as a positive cycle: repeated corrections made him aware of his weaknesses, and recognizing them motivated him to address them.

Grace, an intermediate student with a history of generating most of her essays with AI, began on a sharply different path. The course’s multi-step requirements—drafts, reflections, chatlogs—disrupted her efficiency-driven routine and annoyed her. In the first task she used the AI to bypass cognitive effort entirely, prompting it to explain the assignment, generate definitions, supply vocabulary, and even produce two contrasting revisions she could compare for her reflective report. Although she earned a high teacher score of 92, reading the AI’s superior text triggered strong insecurity about her own competence. The turning point came in the third task, when she discovered grammatical errors inside AI-generated text in her own first draft. The critical incident shattered her belief in the tool’s infallibility. She shifted to verifying AI’s example sources, critically evaluating its output, and combining it with a translation tool—a transformation the authors characterize as a move from practical disengagement to critical engagement.

Sarah’s trajectory was the most troubling. A beginner-level student who had studied Japanese as her foreign language examination subject and set English aside for two years, she could not follow the English-medium AI training workshops and resorted to repeatedly pressing generate and revise buttons. In the first task she had the AI produce an outline and a full essay, then struggled to fill in content the outline lacked, leaving a digital note apologizing to her teacher and explaining her background—a plea the teacher missed because grading covered only the paper-based drafts. Her score of 76 deepened her disappointment. In the group feature article stage, peer mediation briefly changed the picture: conducting interviews herself and translating her own Chinese sentences with a familiar tool, she reported feeling relaxed and engaged, and earned an 84. But when the third task returned to individual writing in an unfamiliar genre, she experienced what the authors call a profound emotional collapse and physically fled the classroom, later describing a fear of foreseeable failure rooted in years of perceived failure in English.

Weighing the frequency, intensity, and contextual embeddedness of positive and negative emotion descriptors across interviews, reflective reports, and informal chats, the researchers identified three qualitatively distinct trajectories: positive-resilient for Leo, marked by sustained positive affect with temporary fluctuations; pragmatic-transformative for Grace, moving from practical disengagement through socially mediated improvement to critical engagement; and vulnerable-contingent for Sarah, sliding from moderate disengagement through tentative peer-mediated gains to breakdown and intensified disengagement. Crucially, the same environmental features produced opposite effects in different learners. Group writing eased anxiety for Grace and Sarah but generated frustration for Leo, whose leadership responsibilities and workload mounted. The AI workshops benefited the two more proficient students but effectively excluded Sarah, whose limited English functioned as an additional barrier. The authors conclude that environmental affordances are only as effective as learners’ capacity to access them.

Beyond confirming the framework’s original factors—mindsets, beliefs, expectancies, values, cognitive skills, and emotion regulation—the study proposes two additions specific to AI-assisted writing: English proficiency and AI literacy as critical developmental competencies. Leo’s advanced proficiency let him critically evaluate AI output and confine the tool to objective aspects such as language while reserving judgment on structure and content for himself. Grace evolved from an unguided heavy user into a critical, strategic one. Sarah’s literacy remained rudimentary throughout. The authors also emphasize that emotional engagement is a dynamic, fluid state that interacts with behavioral and cognitive engagement over time, and that negative emotions alone do not equal disengagement—Leo’s anxiety, for instance, fueled motivation rather than undermining it.

The findings challenge universalist claims that generative AI uniformly improves learners’ emotional experience, and they carry practical weight for educators rushing to integrate these tools. The authors suggest differentiated support: agency-preserving challenges for advanced learners, critical incidents and peer mediation to transform pragmatic users’ relationship with AI, and affective scaffolding combined with foundational language and AI literacy development for vulnerable students. Teacher mediation, they argue, remains irreplaceable, especially for learners who may not voice their needs directly—Sarah’s unacknowledged note being a cautionary example. The authors caution that their conclusions rest on three cases within a single institution, rely heavily on self-report, and span a 14-week window that cannot capture long-term sustainability. Genre, task format, and timing also varied together across the three assignments, limiting attribution of changes to development over time. Larger, mixed-methods studies across diverse contexts, they note, are needed to test whether these trajectories hold more broadly.

Subject of Research: Longitudinal case study of Chinese EFL learners' emotional engagement trajectories in a GenAI-assisted college writing course

Article Title: Engagement or disengagement: understanding the emotional dynamics of EFL learners in GenAI-assisted writing from the sociocultural perspective

Article References: Wu, K., Yang, M., & Lin, C. (2026). Engagement or disengagement: understanding the emotional dynamics of EFL learners in GenAI-assisted writing from the sociocultural perspective. Frontiers in Psychology, 17, Article 1934782. https://doi.org/10.3389/fpsyg.2026.1934782

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1934782

Keywords: generative AI, EFL writing, emotional engagement, sociocultural theory, English proficiency, AI literacy, peer mediation, teacher feedback, perezhivanie, multiple-case study, higher education, Frontiers in Psychology