One Year with ChatGPT: How Two ESL Students’ Writing Habits Evolved

The honeymoon phase with generative AI in education is officially over. For two years, we have been flooded with hot takes about whether ChatGPT will save or destroy academic writing, yet remarkably few studies have bothered to ask the most important question: what happens to a student's writing after a full year of sustained, daily interaction with these tools?

A groundbreaking longitudinal study published in System finally provides the nuanced answer we have been waiting for, tracking two advanced ESL students through twelve months of ChatGPT-assisted academic writing.

This research matters because it shifts the conversation from binary judgments to developmental reality. The novelty of AI writing tools has worn off, and educators now face the genuine challenge of understanding how long-term use shapes a student's independent voice, cognitive engagement, and sense of authorial agency.

The findings challenge both the techno-optimists who promised effortless fluency and the doom-sayers who predicted wholesale plagiarism, revealing instead a complex, evolving relationship between human writers and their machine collaborators.

What emerges from this year-long observation is a portrait of adaptation, dependency, and eventual reclamation of authorial control. The two students did not remain static in their usage patterns; they cycled through distinct phases of exploration, over-reliance, critical evaluation, and strategic integration.

Their journey offers invaluable lessons for educators, curriculum designers, and learners navigating the new reality of AI-mediated academic writing.

TL;DR A year-long study of two advanced ESL students reveals that ChatGPT use evolves through distinct phases, from initial exploration to over-reliance and finally strategic integration. The research challenges simplistic narratives about AI harming or helping writing, showing instead that sustained use reshapes student agency in complex ways. Students who develop critical evaluation skills can reclaim their authorial voice, while passive dependence leads to diminished writing confidence and weaker independent performance.

The Longitudinal Research Design and Its Significance

Longitudinal studies in educational technology remain frustratingly rare, and this research fills a critical gap in our understanding of AI writing tools. Most existing research captures snapshots, measuring student performance at a single moment, which fails to capture the dynamic evolution of human-AI interaction patterns.

The study followed two advanced ESL students at a university setting, collecting writing samples, reflective journals, and interview data across four academic semesters. This methodological richness allowed researchers to trace not just output quality but also the cognitive processes behind each writing decision, revealing how students thought about their own authorship.

What makes this design particularly powerful is its focus on agency, the sense of control and ownership that writers feel over their texts. Agency is the invisible currency of academic writing, and the study tracks how ChatGPT usage either strengthens or erodes this fundamental writerly quality over time.

The findings carry significant implications for ESL pedagogy specifically, where students already navigate the double challenge of language acquisition and academic genre mastery. Adding AI tools into this equation creates a tripartite challenge that demands careful instructional scaffolding.

Phase One: Initial Exploration and Uncensored Enthusiasm

The first weeks of ChatGPT use were characterized by what researchers describe as exploratory enthusiasm. Both students treated the tool as an oracle, feeding it prompts and marveling at the instant fluency it produced, often without critically examining the output quality.

During this phase, writing quality appeared to improve dramatically, at least superficially. Grammar errors vanished, vocabulary became more sophisticated, and sentence structures grew more complex, creating an illusion of rapid language development that masked the students' actual productive capabilities.

Neither student initially recognized the potential dangers of this dependency. They viewed ChatGPT as a tireless tutor, available at any hour, offering corrections without judgment, a stark contrast to the limited feedback available from human instructors in crowded ESL classrooms.

This honeymoon period, however, contained the seeds of future problems. The students began defaulting to ChatGPT for brainstorming, outlining, and even sentence-level decisions, gradually ceding cognitive territory that they would later struggle to reclaim.

Phase Two: The Dependency Trap and Eroding Confidence

By the second semester, a troubling pattern emerged as both students exhibited signs of what researchers term "automation complacency." Their reliance on ChatGPT shifted from assistance to substitution, with the tool becoming the primary generator of ideas rather than a collaborator in thinking.

Writing confidence plummeted during this phase. When asked to write without ChatGPT access, both students reported feeling "naked" and "lost," unable to trust their own linguistic instincts or generate ideas without machine prompting, a classic symptom of cognitive offloading.

The quality gap between AI-assisted and independent writing widened dramatically. Essays produced with ChatGPT earned high marks, while unassisted writing reverted to patterns the students thought they had overcome, revealing that much of their perceived progress was actually borrowed competence.

This dependency phase represents the greatest danger of AI writing tools in education. Students can complete entire courses with excellent grades while their underlying skills stagnate or even regress, creating a credentialing illusion that collapses when they face independent assessments.

Phase Three: Critical Evaluation and Strategic Reclamation

Remarkably, both students eventually emerged from the dependency trap, though through different pathways. One student began cross-checking ChatGPT outputs against academic sources, developing a skeptical stance that transformed the tool from authority to assistant.

The second student's reclamation came through frustration. Repeated encounters with hallucinated citations and generic, formulaic prose led her to reject ChatGPT's first drafts and instead use the tool primarily for localized editing and language polishing rather than content generation.

This critical phase marked the beginning of genuine skill development. Students learned to identify ChatGPT's weaknesses, including its tendency toward bland, hedged language and its inability to capture personal voice or cultural specificity, knowledge that proved invaluable for strategic tool use.

Most importantly, both students reported a restoration of authorial agency during this phase. They described feeling like "directors" rather than "executors," using ChatGPT as one resource among many rather than the primary source of their academic identity.

Phase Four: Strategic Integration and Sustainable Practice

By the final semester, both students had developed personalized frameworks for ChatGPT integration that maximized benefits while minimizing dependency risks. These frameworks were remarkably sophisticated, reflecting genuine metacognitive awareness of their own writing processes.

Their strategic usage patterns shared common elements: ChatGPT was used for brainstorming divergent ideas, checking grammar in final drafts, and translating complex thoughts into clearer English, but never for generating complete first drafts or making substantive argumentative decisions.

Writing confidence rebounded during this phase, though it differed qualitatively from the naive confidence of the first semester. Students now possessed calibrated self-assessment, understanding precisely what they could do independently and where they genuinely needed machine assistance.

The sustainability of these practices remains an open question, but the study's final assessments showed that both students could produce high-quality independent writing while also leveraging ChatGPT effectively, suggesting that strategic integration is a teachable, learnable skill.

Implications for ESL Pedagogy and Writing Instruction

The study's findings demand a fundamental reconsideration of how writing instructors approach AI tools in their classrooms. The current binary framing, either banning ChatGPT outright or encouraging unlimited use, fails to account for the developmental trajectory that this research has revealed.

Curriculum designers must now build explicit instruction around AI literacy, teaching students not just how to use these tools but when, why, and when not to use them. This metacognitive training proved essential in the study, separating students who developed healthy usage patterns from those who remained trapped in dependency.

Assessment practices also require urgent revision. If students can complete writing assignments with AI assistance that masks their true capabilities, then traditional take-home essays become unreliable measures of student learning, necessitating in-class writing components and process-based assessment.

The study's findings also challenge the assumption that AI tools level the playing field for ESL students. While ChatGPT can indeed help non-native speakers produce more fluent text, the dependency risks may be amplified for learners who lack confidence in their linguistic abilities and thus cling more tightly to machine support.

Teaching Agency as an Explicit Learning Outcome

Agency emerged from this study as the critical variable distinguishing healthy from unhealthy AI usage patterns. Students who maintained a strong sense of authorship used ChatGPT productively, while those who ceded agency experienced the most significant skill regression.

Writing instructors must therefore treat agency not as an abstract concept but as a concrete, teachable skill. Students need explicit instruction in maintaining authorial control, including strategies for evaluating AI suggestions, rejecting inappropriate outputs, and asserting their own voice in AI-mediated writing processes.

Reflective journaling proved particularly effective in the study for developing this agency. Students who regularly articulated their writing decisions, including their reasons for accepting or rejecting ChatGPT suggestions, developed stronger metacognitive awareness and more independent writing practices.

Peer review also emerged as a valuable counterweight to AI dependency. When students shared their AI-assisted drafts with classmates, they received feedback that ChatGPT could not provide, including cultural authenticity, personal voice, and argumentative coherence, reinforcing the value of human collaboration.

Designing AI-Integrated Writing Curricula

The phased trajectory revealed in this study suggests that AI integration should be scaffolded across the curriculum rather than introduced all at once. Early courses might restrict ChatGPT use to specific tasks, while advanced courses could allow more flexible integration as students develop critical evaluation skills.

Task design matters enormously in this context. Open-ended, personally meaningful writing assignments appear to reduce problematic AI dependency, while formulaic, template-driven assignments invite students to outsource their thinking entirely to ChatGPT.

Process documentation requirements, such as submitting brainstorming notes, multiple drafts, and revision rationales, can help instructors monitor how students are actually using AI tools and intervene when dependency patterns emerge.

Assessment rubrics must also evolve to reward the skills that matter in an AI-mediated world, including prompt engineering, output evaluation, and strategic revision, rather than merely evaluating the final written product.

Addressing the Equity Dimensions of AI Writing Tools

The study's findings raise uncomfortable questions about educational equity in the age of AI. Students with strong metacognitive skills and academic confidence may benefit enormously from ChatGPT, while struggling learners may fall deeper into dependency, widening existing achievement gaps.

Institutional access policies also create inequities. Students who can afford premium AI tools with advanced features may gain advantages over peers limited to free versions, and students from technology-rich backgrounds may adapt more quickly than those encountering AI tools for the first time.

ESL students face particular challenges, as the study demonstrates, because their language anxiety makes them especially vulnerable to AI dependency. These students need targeted support to develop confidence in their own linguistic capabilities before they can use AI tools productively.

Educators must advocate for equitable AI access while simultaneously building the critical literacy skills that allow all students to use these tools effectively, recognizing that access without agency merely creates new forms of educational disadvantage.

Preparing Students for AI-Mediated Professional Writing

Beyond the classroom, the study's findings carry implications for professional writing contexts where AI tools are becoming ubiquitous. Students who develop strategic integration skills will be better prepared for workplaces that expect AI collaboration as a standard practice.

The ability to evaluate AI outputs critically, maintain authorial voice while leveraging machine assistance, and know when human judgment must override algorithmic suggestions, these skills will define professional writing competence in the coming decade.

However, the study also sounds a cautionary note about workplace AI dependency. Professionals who outsource their writing entirely to AI tools may find their communication skills atrophy, leaving them unable to handle situations where AI assistance is unavailable or inappropriate.

Educational institutions have a responsibility to prepare students for this complex landscape, teaching not just tool proficiency but also the judgment and discernment that separate effective AI collaborators from passive AI dependents.

Beyond the Study: Broader Implications for AI in Education

While this research focused on two ESL students, its findings resonate far beyond that specific context. The phased trajectory of exploration, dependency, and strategic integration likely characterizes how many learners, across disciplines and proficiency levels, will interact with AI writing tools.

The study's emphasis on agency provides a valuable framework for evaluating AI tools across educational domains. Whether students are learning to code, solve mathematical problems, or analyze historical documents, the question of who maintains cognitive control remains paramount.

Educational technology researchers must now build on this foundation, conducting larger-scale longitudinal studies that can identify patterns across diverse student populations and institutional contexts. The field needs more than anecdotal evidence and cross-sectional snapshots.

Policymakers and accreditation bodies also need to grapple with these findings, developing guidelines for AI integration that protect educational quality while acknowledging the inevitability of AI tools in academic and professional life.

The Cognitive Science of Human-AI Writing Collaboration

The study's findings align with broader cognitive science research on distributed cognition and cognitive offloading. When students externalize their thinking to AI tools, they may be short-circuiting the very cognitive processes that build writing expertise.

Writing is not merely transcription but a mode of thinking, and the recursive process of generating, evaluating, and revising text builds neural pathways that cannot be developed by passively accepting machine-generated prose.

However, the study also suggests that AI tools can serve as cognitive scaffolds, supporting students as they develop skills that eventually become internalized. The key variable appears to be whether students actively engage with AI outputs or passively accept them.

Future research should investigate the neural and cognitive correlates of different AI usage patterns, potentially using neuroimaging or detailed process-tracing methodologies to understand what happens in writers' minds during AI-assisted composition.

Redefining Academic Integrity in the AI Era

The study's findings complicate simplistic notions of academic integrity that treat any AI use as cheating. The two students in this study used ChatGPT extensively, yet their final writing demonstrated genuine skill development and authentic authorial voice.

Academic integrity policies must evolve to distinguish between productive AI collaboration and problematic AI substitution. This requires transparent disclosure norms, clear guidelines for acceptable use, and assessment designs that can verify authentic student learning.

The concept of "AI literacy" must become as fundamental to academic integrity education as citation practices and plagiarism avoidance. Students need to understand not just the rules but the underlying values of academic honesty in an AI-mediated world.

Institutions that simply ban AI tools will find themselves fighting a losing battle, while those that thoughtfully integrate AI education into their integrity frameworks will better prepare students for ethical professional practice.

Future Research Directions and Open Questions

This longitudinal study opens more questions than it answers, and the research agenda for AI writing tools remains rich with possibilities. Replication studies with larger samples, diverse linguistic backgrounds, and varied institutional contexts are urgently needed.

Research must also investigate how different AI tools, from general-purpose chatbots to specialized writing assistants, affect student development differently. The specific affordances and limitations of each tool likely shape usage patterns in significant ways.

Longer-term follow-up studies would reveal whether the strategic integration patterns observed in this study persist after graduation, and whether students maintain healthy AI usage habits in professional contexts where the stakes are higher.

Finally, research must examine pedagogical interventions designed to promote healthy AI usage patterns. Which instructional approaches most effectively build the critical evaluation skills and authorial agency that this study identified as crucial?

Practical Recommendations for Educators and Students

For educators, the study's findings translate into concrete classroom practices. Begin by assessing students' current AI usage patterns, then scaffold instruction to move them from naive acceptance toward critical evaluation and strategic integration.

Design assignments that make thinking visible, requiring students to document their AI interactions and reflect on their decisions to accept, reject, or modify machine suggestions. This process documentation builds metacognitive awareness and supports agency development.

For students, the message is clear: use AI tools actively, not passively. Treat ChatGPT as a collaborator whose suggestions require evaluation, not as an authority whose outputs deserve automatic acceptance. Maintain ownership of your writing decisions.

Develop personal guidelines for AI use that specify which tasks you will delegate and which you will handle independently. Review these guidelines regularly, adjusting them as your skills develop and as AI tools evolve.

The year-long journey of these two ESL students offers a powerful corrective to oversimplified narratives about AI in education. Their experience demonstrates that ChatGPT is neither savior nor saboteur but a tool whose educational value depends entirely on how it is wielded.

The developmental trajectory they followed, from naive enthusiasm through problematic dependency to strategic integration, likely mirrors what many students will experience without proper guidance. Educators who understand this trajectory can intervene at critical moments, helping students navigate toward healthy usage patterns.

What ultimately distinguished the successful outcome in this study was not the students' initial proficiency or technological sophistication but their willingness to reflect critically on their own AI usage. This metacognitive capacity proved more important than any technical skill.

As AI tools become increasingly embedded in academic and professional writing, the lessons from this study will only grow in relevance. The students who thrive will be those who maintain their authorial agency, using AI as a powerful collaborator rather than a cognitive crutch.

The future of academic writing is neither purely human nor purely machine but a complex hybrid that demands new skills, new pedagogies, and new frameworks for understanding authorship. This study provides an invaluable roadmap for navigating that uncertain terrain.