The Use of GenAI in Writing Classrooms

1. Introduction

Generative artificial intelligence (GenAI) tools such as ChatGPT are transforming how students approach writing tasks in higher education. This study investigates the role of ChatGPT in the writing process among undergraduate students in one Finnish and one Croatian university of applied sciences. With AI increasingly embedded in digital learning environments, educators must examine how students engage with these tools, what kind of output the students produce with them, and what implications arise for teaching writing, critical thinking, and academic integrity. This study is based on a comparative analysis between the Croatian and Finnish cohort, and on its part, it aims to further understanding on how AI-assisted learning could be used ethically, transparently and effectively in writing education.

2. Theoretical background

The integration of generative artificial intelligence tools into writing classrooms has sparked significant academic interest. With tools like ChatGPT now easily accessible, educators are examining the current writing practices among learners and ways to teach and evaluate writing differently, for example, Hikmawati and Mohammad (2025) recently conducted a literature review in the area of critical thinking and generative AI. Furthermore, Boillos and Idoiaga (2025) have discussed the undergraduate and graduate student perspectives (n-314) on artificial intelligence-based language tools (AILT) and their use in academic writing. Their results show three positive aspects of AILT use in providing “language support,” “help to develop and organize ideas,” and “easy, quick and simple information” – and three negative aspects where AI use leads to issues with “plagiarism and laziness,” “reliability of AI-generated information,” a situation where it “affects academic integrity, and development of critical thinking” (Boillos & Idoiaga, 2025, pp. 162–164). In their article, Boillos and Idoiaga discuss a range of uses associated with such tools, from passive to active use, and they claim that the students in their study were often aware of their linguistic difficulties and used AILTs actively to rectify their mistakes instead of letting the tool make the key decisions passively for them (p. 166). It is also interesting that the study shows that the students are aware of the negative impact of AILTs on their creativity and critical thinking (pp. 158, 165). Boillos and Idoiaga (2025) also highlight the need for teacher mediation to make the use of such tools beneficial for the development of critical thinking and key competencies of the learners.

The concept of AI literacy, defined as the ability to understand the functions of AI, to apply them in different scenarios, to evaluate these applications, and to consider AI ethics in the process, has emerged as crucial for digital education (Ng et al., 2021). For example, Kim et al. (2025) studied the relationship between the level of the students’ AI literacy, their writing performance, and student-generative AI interaction patterns. The study found that the low AI literacy group performed the writing tasks independently without interacting with generative AI while the high literacy group took a collaborative approach, actively accepting generative AI’s suggestions through active engagement, involving it in planning, and allowing the tool to evaluate their work process and task outcomes. These results build a foundation for embedding AI literacy into writing classrooms in the context of critical and active use – when such tools need to be used.

Su et al. (2023) explored the integration of ChatGPT into writing classrooms in a way where the chatbot acts as a student assistant in argumentative writing tasks. Their study found that the student interaction with the artificial intelligence chatbot could support structural, dialogical, and language aspects of the genre. However, they also note that the responses are “reactive in nature” (Su et al., 2023, p. 9), making the response quality dependent on the prompts and the information provided with them. Consequently, the ability of the AI tool to provide proactive and constructive feedback is limited. Hence, it is important for educators to train their learners towards critical use of generative artificial intelligence tools to avoid loss of core competencies and learning potential.

Kosmyna et al. (2025) discuss a study at the MIT lab on brain activity while using an AI assistant in an essay writing activity. This study observed the cognitive engagement and cognitive load of the participants that were divided into three groups based on the tool they were designated to use in the study: an LLM group, a search engine group, and brain-only group. Their results showed that AI use for writing or external support decreased the brain activity and the brain-only group’s brain activity was strong with “widest-ranging networks” (Kosmyna et al., 2025). Also, the sense of ownership that the LLM group’s participants felt over their writing after the writing sessions was low, while the brain-only group exhibited the strongest sense of ownership over the work they had produced. Noting that the group that used AI for their writing performed worse at neural, linguistic and scoring levels, Kosmyna et al. (2025) hence encourage critical perspectives about the impact of AI on learning.    

Overall, these studies, accumulating empirical understanding of GenAI integration in writing pedagogy, have highlighted the strengths and weaknesses of using AI in classrooms, but they also seek to respond to calls for practical strategies in building AI literacy.

3. Methodology

The study participants included 41 undergraduate students at the Rochester Institute of Technology (RIT), Croatia, enrolled in a Writing seminar course, and 56 students from International Business department in Häme University of Applied Sciences (HAMK), Finland. The Croatian participants study International Business, Web and Mobile Technology and New Media Design in English, and the Finnish participants study for a Bachelor’s degree in international business in English. The students’ participation was voluntary and not part of any course work, and all the students gave informed consent for their participation in the study. All the participating students were in their first year of undergraduate studies. At the time of data collection, all the Croatian students were enrolled in a writing seminar course that aims to develop their analytical reading, writing, and critical thinking skills. The course focuses on research and writing processes as well as self and peer assessment in professional contexts. The students in Finland were at the time of data collection completing a course on English in global contexts, which is part of their basic business competence module. This course introduces the students to multicultural teamwork and cross-cultural communication, aiming to enhance their language and presentation skills as well as critical thinking skills in professional contexts.

All of the participating students were instructed to use Microsoft Copilot or ChatGPT to write a well-structured paragraph answering the following question: “What is circular economy? Give an example of circular economy in practice from your experience.” The students submitted their prompt(s), the AI-generated text, a final edited paragraph, and a reflection on the AI tool’s usefulness. Analysing the materials submitted by the students included qualitatively coding the students’ prompt strategies, the writing quality of the submitted paragraphs, their citation practices, and reflective insights.

Our analysis of the writing quality of the submitted paragraphs is based on examining how well the students follow the Claim-Evidence-Analysis model (Cuff, 2022; Gough, 2021). The claim-evidence model was introduced to the students at the beginning of the session together with the paragraph structure and a corresponding example. The students were instructed to follow this model to answer the question they were given for their paragraph writing task on circular economy. They were also instructed to include an example from personal experience. This gave the researchers clear criteria to assess the students’ writing quality. They were also given instructions about citations, the use of ChatGPT, the need to also mention how ChatGPT was used. In addition to this, they were given examples of how these citations in APA7 format should look like. The students then worked on the task, while the screen showed the steps for their writing process, the draft versions they needed to submit, and the aspects their final submission should include.

Although the students could choose between two GenAI models, most of them used ChatGPT and only a very small number of them used Copilot. To ensure consistency, only the students who used ChatGPT are included in the analysed sample. In our analysis, the student responses were qualitatively analysed and the results marked in Excel sheets. This process involved reading the submitted final paragraph along with the submitted information on the different stages writing process to identify similarities between the generative AI output and the final paragraph written by the students. Attention was also paid to the prompt type, that is, whether they used a single prompt, a discussion, or multiple prompts in their interaction with the GenAI model. Since the students had to submit all of these materials that they produced during the process – the prompts, the AI outputs, their initial paragraph drafts, and their final paragraph – it was possible for us to analyse the writing process in detail.

4. Results and discussion

The students displayed a wide range of prompting. Many relied on generic queries such as “What is a circular economy?” which was used by 61% of students (25) from Croatia and 41% of students (23) from Finland, while others, about 24% of students (10) from Croatia and 1.7% of students (1) from Finland demonstrated more complex prompting. This more complex approach included asking for examples, academic tone, or APA citations (e.g., “Give examples with references in APA style” or “Exclude Wikipedia and focus on sustainable development”). The largest difference between the participants from the two countries was in iterative prompting behaviour, as only 15% of students (6) from Croatia but 57% of students (32) from Finland refined or revised their prompts after evaluating the first output from the GenAI model.

For the purposes of this study, as generic queries we classified all prompts that were simple factual questions which gave the AI chatbot no additional context and instead asked an isolated question, such as “what is circular economy?” As complex prompts, on the other hand, we classified prompts that included details such as context, why the student needs the information, the role of the generative AI tool in the exchange, the expected outcome, the desired length and mode of the outcome, and so on. Finally, as iterative prompting we classified prompts that were a combination of complex prompting and further clarifications through additional prompts that continued the same query. This could include, for example, providing examples, criteria to be considered for the task, interacting with the AI tool on the possible mistakes or gaps in the output, and giving pointed instructions for the required modifications in the output.

The pattern of using mostly generic or complex, but not iterative, prompting among the Croatian participants indicates that while most were able to initiate interaction with ChatGPT, only a minority engaged in strategic prompt engineering which is an essential element of GenAI literacy. At the same time, more than half of the students in Finland engaged in iterative prompting, showing a higher level of engagement.

Based on analysing the students’ submitted reflections on the writing process and the opinions they shared about AI use and its limitations, three primary patterns or strategies emerged in terms of their use of ChatGPT-generated content: (1) minimal revision (37%, or 15 students from Croatia, and 25%, or 14 students from Finland) where the GenAI output was only superficially edited by changing wording, fixing grammar, or reordering sentences, (2) hybrid writing (34%, or 14 students from Croatia, and 46%, or 26 students from Finland) where the students significantly edited the ChatGPT-generated content, adding their own examples, rephrasing ideas, or integrating external facts, and (3) inspiration-only approaches (29%, or 12 students from Croatia, and 28%, or 16 students from Finland) where the students used ChatGPT only for brainstorming or background knowledge, composing their final paragraph independently. The students in the hybrid and inspiration-only groups produced more natural academic tone and stronger arguments, while texts from the minimal revision group tended to mirror ChatGPT phrasing and completely lack originality.

Most students (85% from Croatia and 83% from Finland) followed the Claim-Evidence-Analysis model that they have been taught in their courses and that we explained to them at the start of the session. The strongest paragraphs contained a clear claim, a specific example and an explanation of relevance (34%, or 14 students from Croatia, and 53.5%, or 30 students from Finland). On the other hand, about 39% (16) Croatian and 16% (9) Finnish students submitted weaker paragraphs, often containing text pasted from ChatGPT with little editing. These texts were frequently generic, repetitive, and lacked analytical depth. It was possible for us to see the writing process and identify the source of the information because the students submitted the prompt they used, the AI output received, the other prompts used for clarification and modification of the output, and any further output they received. Since we asked the participants to submit all the stages of revision they did for the text in addition to the final paragraph, and they were also asked to submit their reflections on the writing process and opinions about AI use and limitations, this created a comprehensive understanding of how they interacted with the AI chatbot. For example, patterns of AI use which reveal lifting the terms directly from the AI output, such as the phrase “an economic model aimed at minimising waste,” appeared in multiple responses, indicating limited student revision or engagement with the content. The remaining (27%, or 11 students from Croatia, and 30%, or 17 students from Finland) produced paragraphs that merged ChatGPT-generated content with personal insights or research. These paragraphs demonstrated improved tone, structure, and flow, although they varied in the depth of their discussion of the designated topic while commonly showing reliance on the ChatGPT content as they incorporated to their final text. However, they differ from the simply copy-pasted paragraphs because a degree of active engagement is visible through the incorporation of their personal insights as edits to the text.

The students expressed mixed views on the usefulness of ChatGPT’s usefulness. 59% (24) Croatian and 91% (51) students from the Finnish institution valued ChatGPT’s speed, clarity, structure, and idea generation, and 75% (42) Finnish students were positive about the tool. Typical comments included notions on how GenAI can be “A great starting point,” or “Useful for getting my thoughts organised.”

On the other hand, 41%, or 17 students from Croatia, while only 5%, or 3 students from Finland were negative or sceptical about ChatGPT and they criticised its generic tone, vague references, and robotic phrasing. These students were more likely to revise or discard AI output. At the same time, 59% (33) of the Finnish students expressed their concerns about the limitations of GenAI tools and stated the need for being critical and selective while using content generated from such tools. The typical phrases describing these concerns were “It could be an assistant but not the sole content creator” and “These tools often hallucinate, questioning the rest of the legible content they generate,” or “I use them to do my information gathering, but I want to control the narrative.” This pattern correlates with the high percentage of iterative prompting by the students from Finland, indicating more critical engagement with the AI application. These results on the awareness of the students on the kind of AI use they engage in and the limitations of such chatbots aligns with the results of the study conducted earlier by Boillos and Idoiaga (2025).

Citation accuracy was generally weak. Only 17%, or 7 students from Croatia attempted proper APA referencing, while 59%, or 33 students from Finland used strong citations, probably following the instructions given during the session. 24%, or 10 students from Croatia, and 14%, or 8 students from Finland included URLs or named sources without APA formatting. However, a majority of the students – 59% (24) of participants from Croatia and 27% (15) from Finland – did not cite any sources, including ChatGPT, which is a concerning finding. What is more, although the students were instructed to cite also the GenAI output, most failed to acknowledge ChatGPT as a co-contributor. Overall, any ethical reflection on authorship, originality, or source transparency was rare.

The abovementioned data suggest three user profiles: (1) Passive AI consumers who copied and slightly edited the AI output (21% from Finland and 38% from Croatia), (2) guided collaborators who used AI but revised content meaningfully (44% from Finland and 42% from Croatia), and (3) critical human writers who used AI only for ideation and produced the most academically sound writing (32% from Finland and 20% from Croatia).

Overall, this study supports findings in earlier research which show that while GenAI tools like ChatGPT can support students in academic writing, their effectiveness depends on how thoughtfully and critically they are used. Because of this, it is important for educators to go beyond teaching technical use and also to help students learn how to revise their work, cite sources properly, and reflect on the ethical implications of using AI-structured tasks – such as the paragraph-writing assignment explored here – in order to create meaningful opportunities that can build AI literacy in real academic contexts. It is equally important to guide the students in evaluating sources, developing their academic voice, and in understanding their responsibility as authors.

5. Conclusion

This study offers additional insight into how Croatian and Finnish undergraduate students from two institutes of higher education engage with GenAI tools, such as ChatGPT, in academic writing contexts. The findings show diverse prompting strategies, varying levels of revision, and differing attitudes toward AI-generated content. While some students were found to show critical engagement and ethical awareness, a significant portion relied passively on AI outputs, raising concerns about originality, citation practices, and authorship responsibility.

These results emphasise the importance of including AI literacy into writing pedagogy – not only to enable its technical use but also to encourage reflective, ethical, and critical approaches to AI-assisted writing. Structured tasks, such as the paragraph-writing assignment examined in this study, can provide a practical framework for promoting these skills.

When comparing the two cohorts of this study, it becomes evident that the Finnish students demonstrated higher levels of AI literacy and autonomy. They engaged more actively in iterative prompting, critical evaluation of AI outputs, and ethical citation. The participating Croatian students, on the other hand, showed enthusiasm for AI as a supportive tool but relied more heavily on default outputs and less on reflective revision. Both cohorts displayed awareness of AI’s limitations and the importance of human oversight, confirming that cross-institutional collaboration and comparative pedagogical reflection can effectively improve responsible AI use in writing classrooms.

Authors