More human than human? Differences in lexis and collocation within academic essays produced by ChatGPT-3.5 and human L2 writers

2.1 Generative AI in education

Generative AI (GenAI) in this study refers to software tools utilising Large Language Models (LLMs) trained on extensive text data that are able to generate human-like text, answer questions, and perform various language-related tasks with high accuracy (Kasneci et al. 2023). These tools demonstrate capabilities in producing intelligent-sounding responses to users’ natural language prompts and are proficient in generating coherent and contextually relevant text reflecting advanced natural language processing abilities (Stokel-Walker 2022). ChatGPT as a prominent GenAI tool, offers numerous advantages by providing real-time access to vast amounts of information, thus facilitating the acquisition of (nearly) up-to-date and accurate data for literature and research (Wang et al. 2023).

The application of GenAI has gained significant attention in educational research as a tool to assist teachers and educators in creating more flexible, efficient, inclusive, engaging, and customised learning environments. For instance, GenAI can collaborate with teachers to accomplish tasks that are challenging for teachers to complete alone, such as grading a large volume of student assignments and providing the necessary support and scaffoldings for learners, including real-time feedback (Chan and Hu 2023). Human-machine collaboration not only reduces the workload of teachers but also enhances the effectiveness and efficiency of education, enabling learners to receive a more personalized educational experience. Cooper (2023) highlighted that educators can gain a deeper understanding of the learning process by incorporating GenAI across multiple domains. e.g., providing a mentor for each learner, helping learners and teachers acquire new 21st-century skills, enhancing data interaction, enabling the use of global classrooms, and supporting lifelong learning. Yang and Kyun (2022) provided similar findings and conclusions through an analysis of 25 empirical research papers on AI-supported language learning, namely that: (a) AI provides and supports more effective language learning and collaboration among learners; and (b) combining AI tools with teacher guidance can produce better learning outcomes and achieve higher aims.

2.2 (mis)Applications of generative AI for L2 writing

AI has also demonstrated itself as a potentially powerful tool in supporting L2 writing and pedagogy, encompassing a range of tools such as automatic translation, error correction, and automated scoring programs, all of which are now widely used in L2 writing classrooms (Stokel-Walker 2022). The impressive capabilities of GenAI writing tools in generating intelligent text in response to user prompts have potential for creating high-quality texts, increasing learners’ interest in L2 writing and helping them to achieve higher learning outcomes (Zhao 2023). Previous research has emphasized that incorporating GenAI technology into the classroom significantly promotes students’ L2 writing via providing instant, personalised feedback (Atlas 2023). Wang et al. (2023) highlight that even many academic practitioners already use GenAI to summarize literature, discuss and revise manuscripts, draft research proposals, and identify research gaps. For instance, Kung et al. (2023) and Cotton et al. (2024) found that ChatGPT assists researchers in rapidly collecting and analyzing relevant literature. Chan and Hu (2023) noted that integrating GenAI into L2 writing offers benefits such as reducing learners’ writing time, costs, frustration, and anxiety, providing quick assessments through timely feedback, and predicting learners’ future performance. ChatGPT also excels in L2 writing assessment, providing more objective and faster scoring than humans, with Xia et al. (2024) finding ChatGPT achieved an accuracy rate of 84.38 % when scoring various types of essays.

However – and rightly so – educators remain concerned about potential academic integrity issues, and the debate on how to integrate GenAI tools to enhance English writing skills while adhering to academic integrity principles is ongoing. For example, Guo et al. (2023) discovered that the text generated by ChatGPT-3.5 is remarkably similar to human-written text. In an experiment, participants unfamiliar with ChatGPT were asked to distinguish between answers provided by ChatGPT and those generated by humans. More than half of the participants could not differentiate between the two and found ChatGPT’s answers to be more detailed and useful. This challenge to distinguish between GenAI generated outputs and learners’ original writing is crucial in maintaining academic integrity (Cotton et al. 2024). Studies suggest that traditional university assignments are in danger as ChatGPT can now generate high-scoring essays and provide correct answers to problem questions (Al-Zahrani 2023; Crawford et al. 2023). This challenge is compounded by the fact that the ethical and acceptable boundaries of using ChatGPT in academic writing remain unclear to both students and staff (Tlili et al. 2023). Given the rapid evolution of generative AI tools and their increasing integration into educational contexts, it is imperative to adopt a critical perspective towards these technologies and educate leaners on their appropriate and ethical use, as underscored by recent research (e.g., Guo et al. 2023; Tlili et al. 2023).

Specifically, regarding lexis, studies have found that GPT exhibits notable performance in lexical choices and their application in generated academic texts. For instance, Li et al. (2024) analysed the vocabulary usage of GPT in academic English witing and discovered that it excels in terms of lexical diversity and precision. Additionally, AlAfnan and Mohdzuki (2023) examined GPT-generated academic texts highlighting the model’s high consistency and accuracy in employing technical terms and academic vocabulary. Advances in LLM technology are also closing the gap, given ChatGPT-4 shows significantly higher lexical diversity when compared with ChatGPT-3 (Herbold et al. 2023). Enhancements in GPT’s ability to express politeness and handle deixis within texts have also improved the naturalness and contextual awareness of its generated content (Rafique et al. 2024). Nonetheless, despite these advancements, ChatGPT’s performance remains inferior to that of human writers in terms of lexical diversity, nuanced collocational usage and sensitivity to contextual appropriateness (e.g. Jiang and Hyland 2024).

However, almost all research to date has focused on the differences between GenAI-produced writing and that of monolingual L1 writers (e.g. Bašić et al. 2023; Li et al. 2024). Studies comparing differences between GenAI-produced writing and those of developing L2 academic writers are far less prominent. Examples include the impact of ChatGPT’s automatic text generation on undergraduate students’ perceptions and behaviours in learning L2 writing skills (Yan 2023), or studies focusing on the use of ChatGPT as a tool for influencing and guiding responses in higher-education students’ L2 writing (Zou and Huang 2023). Understanding differences in texts produced by L2 writers compared with GenAI is important because educators need to identify the specific advantages and disadvantages of AI-generated content compared to L2 human writing if they are to use it correctly as a more effective writing aid, supplementing human teaching and thereby enhancing the overall quality of L2 writing education. Furthermore, this understanding helps set appropriate expectations for the use of AI in academic contexts, ensuring that L2 students and educators can make informed decisions when integrating these technologies into their learning and teaching processes.

3.1 Data

The data is divided into two sources: one containing L2 human-produced academic writing (L2 human corpus) and the other GenAI-generated academic writing (GPT corpus). In the L2 group, essay topics are sourced from a corpus of 30 argumentative essays each on different topics taken from an undergraduate level academic writing course at an Australian University. The course is designed to improve undergraduate students’ L2 academic writing skills, while the essay task is designed to encourage students to incorporate critical thinking into their writing, using evidence and citations concisely and clearly to articulate their positions. Feedback for each essay was provided at the planning stage by English-speaking academic writing instructors, so as to help students develop their drafting, revising and critical thinking abilities. Students are required to use formal vocabulary and academic language to successfully complete the task. Although the students’ names and ages are not disclosed, they are all undergraduates, aged between 18 and 20 years. They are non-native English speakers with diverse first languages and an English proficiency level ranging from upper intermediate to advanced (CEFR B2-C1), with IELTS scores between 6.5 and 7.0.

Essay topics cover various fields according to the students’ own selection. In selecting the texts for analysis, care was taken to diversify the range of topics to prevent overlap and to cover various fields, as listed in Appendix 1. By including a wide range of topics, any bias that could emerge from lexical variation tied to a single subject (e.g., specialised terminology in a particular field) is reduced. Each essay in the corpus follows APA 7th edition citation and referencing format with an average length of 1,170 words excluding references.

ChatGPT-3.5 was employed as an additional source of data, comprising 30 essays generated by ChatGPT-3.5 using the following prompt: “Create a 1,000-word academic text about [topic], using APA 7th edition citations excluding references and without subheadings”. A new chat window was opened for each essay. The topic element was taken verbatim from the equivalent topic from the L2 corpus via prompting to ensure comparability. The generated texts were copied into Microsoft Word documents to facilitate analysis. The average length of the ChatGPT-3.5-generated corpus was 896 words with the longest essay being 1,100 words, not including reference lists.

The analysis of lexical and collocational patterns in both the L2 and GPT corpora was conducted using R statistical software. As an open-source, high-level programming language and environment, R is highly flexible and supports faster processing on all major operating systems, including Linux, Windows, and Mac, without requiring licensing compliances (Kumar and Tyagi 2024). While corpus linguistics software such as AntConc and Sketch Engine are commonly employed in linguistic studies (Zih et al. 2020), R provides improved flexibility for automating processes, conducting custom statistical tests, and visualising the results (Schweinberger 2023), facilitating detailed comparisons between the lexical and collocational choices in each corpus.

3.2 Analysis

R Studio version 4.33 (2024.04+735, Posit Software, PBC) (within the R programming language) was employed as the instrument for all data analysis. R is a comprehensive programming environment encompassing complex natural language processing, statistical data analysis, data visualisation, and the creation of applications and websites (Schweinberger 2023; Verzani 2011).

Before importing the data into R Studio, the reference list from each text was manually removed and the file format was converted from “doc.” to “txt.” format. In R Studio, several essential data analysis packages were installed, including ‘here’, ‘stringr’, ‘dplyr’, ‘ggplot2’, ‘quanteda’, ‘quanteda.textstats’, ‘tm (Text mining)’ to process and promote the textual data with loading, searching, cleaning, annotating, and visualizing within a single tool (Arnold et al. 2019). Subsequently, the command ‘library’ was used to activate all analysis packages. Initially, the data paths for both groups were imported to identify the locations of the 60 academic essays. Then, the data were imported, and new filenames were assigned to each data group, with ‘gpt’ respesenting the GPT corpus and ‘hmn’ representing the L2 human corpus. The data for both corpora were reviewed to ensure successful importation. The next step involved cleaning all the data, conducted using the “qyanteda” package in R (Banks et al. 2018). This process assisted in converting all words in each text to lowercase, and numbers, functional words, prepositions (such as ‘the’, ‘an’, ‘of’, ‘from’), pronouns (such as ‘I’, ‘she’, ‘he’), and non-words (such as ‘et al.’ in citation) were removed. The most frequently used words in any corpus are often found together with function words that lack substantive meaning, making it essential to remove these function words at an initial stage. All remaining tokens were then compiled into a data frame named “texts_df”, with the corresponding frequency of each token in both “gpt” and “hmn” recorded before the calculation of cross-corpora frequency, collocation and keyness according to the procedures below (Figure 1).

Figure 1:

Overview of analysis procedures within R programming language.

Figure 1 describes the analysis procedures within R programming language.

3.2.3 Keywords

Regarding cross-corpus keyword calculation, the odds ratio (OR) is a crucial metric for determining differences across groups. An OR greater than 1 implies a strong likelihood of difference, an OR less than 1 implies a lower likelihood of difference, while an OR equal to 1 means there is no association between the exposure and the outcome. The formula for calculating the odds ratio is following:

Odds ratio  ( OR ) = odds   of   outcome   in   the   exposed   group odds   of   outcome   in   the   non − exposed   group  

( OR ) = O 11 × O 22 O 12 × O 21  

The log odds ratio (Log OR) is the natural logarithm (log) of the OR. Log OR is widely used in statistical models, particularly in logistic regression analysis, due to its preferable mathematical properties compared to the raw OR. Log transformation aids in stabilising variance and linearising relationships between variables, which is beneficial for various statistical analyses. The formula for the Log OR is below:

OR = a × d b × c

log OR = ln ( OR ) = ln ( a × d b × c )

In the current study, the formula used is based on the Log OR, with an addition of 0.5 to each cell in a 2 × 2 contingency table. This adjustment, known as the Haldane Anscombe correction, addresses the issue of zero. Without this correction, an OR is involving a zero count would be undefined due to division by zero. Given that some terms in the generated data have zero counts in one corpus or another, applying an appropriate correction method is essential. The formula is used in this study:

OR = log ( ( O 11 + 0.5 ) × ( O 22 + 0.5 ) ( O 12 + 0.5 ) × ( O 21 + 0.5 )   )

Due to the large volume of data, this study applied a filtering criterion with a p-value less than 0.001 to focus on statistically highly significant data. This criterion ensures that the likelihood of the observed association occurring by chance is extremely low, less than 0.1 %.

To visualise the Log OR as linguistic feature keywords, the top 20 Log OR keywords with p < 0.001 from each corpus were extracted and displayed in figures. The Y-axis represents tokens, while the X-axis indicated the association strength between tokens and Log OR.

Figure 2:

Keywords in human-produced texts.

Figure 3:

Keywords in GPT-produced texts.

4.1 Word frequency

As previously mentioned, there are 30 different topics covering various fields, including the nuclear wastewater issue, education, AI, violence, the fashion industry, health, economy, and society (Appendix 1). As the students were free to choose their own topics, we could observe individual preferences and lexical tendencies in L2 academic writing based on students’ personal experiences and perspectives.

Table 1 describes the highest frequency words in the L2 human corpus.

The token “health” emerges as the most frequently mentioned word in the L2 corpus, appearing 141 times across 30 essays, equating to an average of at least 4.7 mentions per essay. Given the freedom students had to select their own topics, this frequency likely reflects a strong personal interest in health-related issues, possibly tied to the global relevance of health discussions in academic and societal contexts. Closely following are “sleep”, “students”, and “children”, each cited over 100 times, averaging more than three mentions per essay. This pattern indicates a strong focus on well-being within the essays, with related vocabulary including “intelligence” and “therapy” (mentioned 90 times), and “vaccines” and “human” (71 times). These words likely reflect broader societal concerns that are commonly discussed in academic settings, which supports the idea that L2 writers engage with socially relevant issues in their writing. The high frequency of words such as “students” and “children” also highlights a significant focus on educational themes and issues related to younger demographics. These results suggest that L2 writers frequently choose topics that are closely connected to their personally related factors, cognitive, affective (Kormos 2012), a trend observed in other studies on L2 composition.

Table 2 describes the highest frequency words in the GPT corpus.

In the GPT corpus, the most frequently mentioned word is the same as in the L2 human corpus (“health”), highlighting a shared focus on health-related topics, given the same topics as those from the human data were used to prompt GPT. Given this information, it is unsurprising that certain words like “health” are frequent in both corpora. However, the frequent appearance of abstract terms such as “challenges”, “impact”, and “energy” all appearing over 100 times in the GPT texts suggests a more generalised focus on societal and technological themes, contrasting with the more personal, human-centred themes in the L2 corpus.

The frequent mention of “AI (artificial intelligence)”, which appears 99 times, also underscores the importance of technology (namely it’s own) in GPT’s output as compared with L2 writers. Other words not appearing in the L2 corpus include “economic” and “potential”, “concerns”, “contribute”, “individuals”, “essential”, further indicating the GPT-produced texts’ different focus on addressing global issues rather than personal ones, despite sharing the same essay topics. Comparing the words of the two groups reveals that the human L2 writers are more focused on human-related aspects, such as education, therapy, and health, whereas the GPT corpus is more focused on relatively less-human content, including AI, economics, and technology. These differences suggest that even when prompted with similar topics, GPT’s training data and model prioritises more abstract, generalised content, while L2 writers tend to draw on their personal experiences and educational backgrounds.

Additionally, the top 20 words in the ChatGPT corpus generally have higher frequencies than those in the L2 human writing corpus, with each ranking differing by about 10 occurrences. This difference in frequency distribution reflects the varying focus and content generation strategies between the two groups. For example, despite “students” being a common term in both corpora, in the human group, “students” appeared 122 times, ranking third, whereas, in the GPT corpus, it appeared only 64 times, ranking second to last. This suggests that although both groups possibly only briefly touched upon the subject, human L2 writers might place a greater emphasis on discussions involving students - reflecting their personal experiences - while GPT-generated content has a broader thematic focus.

4.2 Collocations

All collocations listed in Tables 3 and 4 focus on bigrams or word pairs. These were chosen over longer collocation spans as shorter collocation occur more frequently in writing due to their flexible use across different contexts (Shin and Nation 2008), providing a broader basis for our analysis. Table 3 describes the highest frequency collocations in the L2 human corpus.

As seen in the word frequency data above, the collocational content of these essays primarily focuses on leadership and work-related themes, as evidenced by the most frequent collocation is “transformational leaders”, appearing 15 times, followed by “transformational leadership”, which appears 11 times, revealing the importance placed on leadership within this L2 writing sample. This may reflect both academic and real-life interests of the students, as leadership is often a key topic in educational and professional development writing tasks, such as autobiographies or biographies where personal growth and leadership experiences are frequently highlighted (English 2006). “Job satisfaction” and “career development” discuss work and career planning issues, reflecting the writers’ emphasis on professional aspirations.

Despite the variety of essay topics, health-related collocations such as “mental health” and “childhood obesity” also feature prominently aligning with previous research showing that L2 writers often draw on familiar, socially significant topics that reflect their language learning goals and self-efficacy beliefs, which influence their choice of personally of academically relevant subjects (Kormos 2012).

Regarding collocation strength, a high Z-score indicates a high significance with a low likelihood of each word being found together by chance. In the L2 human corpus, the collocation “transformational leaders” exhibits the highest Z-score, indicating a statistically significant and strong association between these terms. Conversely, “team members” demonstrates the highest lambda value 10.65, suggesting a robust relationship despite its lower frequency. This pattern suggests the collocations with the strongest statistical associations are not necessarily those with the highest frequency. The data therefore indicate that L2 writers tend to rely on familiar word combinations while also making efforts to diversify their lexical choices. Overall, the lambda values of these 20 collocations are mostly concentrated around 6 to 8, indicating that their associations strengths are relatively similar.

Table 4 describes the highest frequency collocations in the GPT corpus. The data reveal some common collocations, such as “mental health”, “childhood obesity”, and “nuclear wastewater”, indicating that the vocabulary and collocations used in these areas are highly similar between the two corpora. However, there are significant differences between the two datasets. The content and collocations produced by GPT are more extensive and diverse, covering topics for health, renewable energy, vaccines, public health, academic writing, performance, fashion, and climate. Also, each collocation appears more frequently in the GPT data than in the L2 human writing data, with the lowest frequency GPT-produced collocation “employee motivation” appearing 15 times, the same as the highest frequency collocation in the L2 human data, “transformational leaders”. This discrepancy may suggest that GPT-3.5, due to its training on extensive datasets, has the capacity to produce texts that consistently employ high-frequency collocations across a wide range of topics, whereas L2 writers may be more constrained by their familiarity with specific subject matter and depth of acquired vocabulary.

Regarding collocation strength, in the Z-score column, the highest score in Table is 19.10 for “emotional intelligence”, far higher than the highest score of 13.12 in Table 3. The lambda values in Table 4 range from 5.37 to 10.65, with higher values indicating stronger associations between collocated words than those found in Table 3. These differences suggest how ChatGPT-3.5 and L2 human writers approach topics differently, reflecting distinct priorities and influences in their writing. GPT-3.5’s extensive training data enables it to produce a greater variety of collocations with stronger internal associations, suggesting that it can deliver more lexically diverse content that L2 writers. The L2 writers’ reliance on fewer collocations and topics more closely aligned with their experiences might indicate a tendency to engage deeply with familiar themes rather than exploring a broad range of subjects, as well as a limited range of productive collocations from which to draw on when writing.

It is important to note that these human authors are L2 writers, not L1 writers. This factor may explain many of the observed results, such as lower collocation frequencies and reliance on a smaller range of topics and collocations that the students are familiar with, given their limited L2 vocabulary. Also, in contrast to the data in the L2 corpus, none of the collocations in Table 4 are repeated, which suggests that GPT tends to diversify its collocational use across different contexts. This could potentially suggest GPT-generated texts tend to discuss all topics superficially without deep association or understanding of any particular field, relying solely on extracting content from its database when asked questions or given instruction rather than drawing on personal experience. The L2 writers’ limited vocabulary and familiarity with certain topics result in lower collocation frequencies and a narrower range of discussed topics, contrasting with GPT’s broader and more varied content generation.

4.3 Keywords

Table 5 describes the tokens and Log OR in the L2 human corpus.

The Log OR values describe the relative frequency of these words generated by ChatGPT-3.5 compared to L2 human-produced L2 academic essays compared with those generated by ChatGPT 3.5. 35 tokens meet the high significance criterion for inclusion in Table 5. The Log OR values predominantly fall within the range of −1 to −3, with only one token below −1 and none exceeding −3. The more negative the value, the higher the association with the human group, so item 35 in Table 5 is the most likely to appear in the L2 data and not the GPT data. The word most closely associated with the human group and appearing most frequently is “food”, with a Log OR of −2.972, it means that this token appears more significantly in human-produced content and has a higher association, whereas its frequency and association are lower in ChatGPT-generated content. Similarly, other words such as “youtube (−1.122)”, “growth (−1.419)”, “autism (−1.982)”, “employees (−1.993)”, “explore (−2.633)”, and “children (−1.469)” are also more common in the L2 human corpus due to their connection with personal experiences, authentic context, and emotions. As previously mentioned, “leaders” also is associated with human L2 writing as compared with that of GPT.

Table 6 describes the tokens and Log OR in the GPT corpus. The data illustrate several trends in lexical selection by GPT that do not feature in L2 writing, indicated by positive Log OR. Notable examples e.g. the tokens “technical” and “initiatives” with relatively high Log OR values indicate a strong association with GPT-generated texts and suggest a broader and more abstract lexical focus than that observed in the L2 corpus. Another significant token is “sustainable (2.869)”, indicating a high frequency of discussions on sustainability in ChatGPT-3.5 academic essays. Furthermore, tokens such as “vehicles (2.169)” and “renewable (1.738)” demonstrate that ChatGPT-3.5 frequently engages with subjects related to transportation and renewable energy.

When comparing the GPT-generated text table with the human-produced text table, several key differences are immediately apparent. In addition to the overall higher Log OR for tokens in the GPT corpus compared to the L2 human-writing corpus, the vocabulary in the ChatGPT table is mainly concentrated on themes such as technology and sustainability. In contrast, the L2 writer-produced text shows a prevalence of terms related to leadership, health, and specific real-world issues. Tokens such as “leaders”, “youtube”, “radioactive”, and “children” appear more frequently in human-produced texts as evidenced by their minus Log OR values in the GPT corpus. This suggests that human texts focus more on personal experiences, current events, and in-depth discussions of social issues. These distinctions are particularly relevant when considering that the human texts are produced by L2 writers. L2 writers might have a narrower range of vocabulary and less fluency compared to L1 writers, which could result in a stronger focus on familiar and concrete topics, such as leadership and health, and a greater reliance on real-world experiences and current events. This contrasts with ChatGPT-3.5, which has access to vast database of information and can generate content on a wide range of topics, including more abstract or technical themes.

Figure 2 visually describes the association strength between keywords and Log OR in the L2 human corpus. The horizontal axis represents the association strength (Log OR), while the vertical axis lists the keywords. Each point on the chart corresponds to a specific keyword and its Log OR value, indicating the frequency and importance of that keyword in L2 human-generated content compared to GPT-generated content. Keywords with more minus Log OR values are more strongly associated with L2 human texts, suggesting a higher frequency in human writing.

Figure 2 further illustrates the frequency and importance of certain keywords in the L2 human corpus relative to the GPT corpus. High association strength keywords include “food” and “career”, suggesting that these words are more common and frequently discussed in human-generated texts. These keywords reflect topics that are likely grounded in the real-life experiences or concerns of L2 writers. Keywords with medium association strength keywords such as “censorship” and “children”, while low association strength keywords like “leaders” and “youtube”, although more common in human texts, show less frequency difference compared with that of GPT-generated texts.

Figure 3 describes the association strength between keywords and Log OR in the GPT corpus.

Figure 3 displays keywords such as “educational” and “technical” have Log OR values exceeding 3.0, indicating that these tokens are used significantly more than in the GPT corpus compared to the human corpus. These two keywords represent the highest association strengths, highlighting ChatGPT’s focus on topics related to education, technology, and entertainment. Keywords with moderate association strengths include “play” and “initiatives”, suggesting a considerable amount of discussion on new projects or plans in ChatGPT texts. “individuals” and “essential” are examples of keywords with lower association strengths. While these tokens are still common in ChatGPT content, the frequency difference compared to human texts is less, and they do not stand out as much as the high-strength keywords. The Log OR values for the L2 human corpus are minus, with larger absolute values indicating higher association strength. Therefore, the keywords for the L2 human corpus are arranged from nearly −3.0, starting with “food” and “career”. Conversely, the Log OR values for the GPT corpus are positive, meaning that large values indicate higher association strength. Again, the variation in keyword association strengths highlights the thematic focus of each corpus. GPT-generated texts emphasise structured, formal discussions on education and technology, whereas L2 human writers tend to concentrate on topics grounded in authentic experiences and societal issues. This could be attributed to the limitations in L2 writers’ vocabulary and their preference for familiar and concrete topics, which contrast with ChatGPT’s ability to generate content on a broad range of subjects, including more abstract and technical themes. Understanding these distinctions provides a deeper comprehension of the characteristics of GPT-generated content and the focal points of L2 human-produced texts, offering valuable insights for L2 educators.