THE IMPACT OF ARTIFICIAL INTELLIGENCE ALGORITHMS ON AUTOMATIC EDITING AND PROOFREADING OF TEXTS

Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9604 THE IMPACT OF ARTIFICIAL INTELLIGENCE ALGORITHMS ON AUTOMATIC EDITING AND PROOFREADING OF TEXTS OLEKSII SYTNYK1, NATALIA KNYRIK2, OKSANA TIUTIUNNYK3, YEVGENYA BAZINYAN4, TETIANA TSEPKALO5 1Taras Shevchenko National University of Kyiv, Department of Editorial and Publishing Technologies and Production, Ukraine 2Admiral Makarov National University of Shipbuilding, Department of Information Control Systems and Technologies, Ukraine 3Hryhorii Skovoroda University in Pereiaslav, Department of Ukrainian and Foreign Literature and Teaching Methods, Ukraine 4Armenian State University of Economics, Armenia 5Vinnytsia Mykhailo Kotsiubynskyi Statе Pedagogical University, Department of Journalism, Advertising and Public Relations, Ukraine E-mail: [email protected], [email protected], [email protected], [email protected], [email protected] ABSTRACT The relevance of the study is determined by the rapid development of artificial intelligence (AI) and the need to assess its impact on the quality of automatic editing of English texts in the global informational technologies (IT) environment. The aim was to identify and assess the impact of various AI algorithms on the efficiency and quality of automatic editing and proofreading of English texts in an international context. The objective of this study is to identify and analyze how various AI algorithms impact the efficiency and quality of automatic editing and proofreading of English texts in an international context. The study employed the following methods: experimental methods, content analysis, comparative analysis, and focus groups. The reliability was confirmed by using qualitative statistical analysis, Student’s t-test, chi-square test, and analysis of variance (ANOVA). The assessment was carried out in comparison with human editing. The texts included news, scientific articles, advertising texts, essays, and technical documentation. Readability analysis showed statistically significantly lower indicators of AI compared to editors (FleschKincaid: 69.4 versus 75.2). The users noted the convenience and speed of intelligent technologies, but pointed to insufficient flexibility and stylistic problems. Analysis of variance confirmed the significant advantage of editors in quality (F=12.56, p<0.001, η²=0.45). The results obtained during the study showed a significant advantage of professional editors over modern AI algorithms in editing quality. At the same time, its use demonstrates the potential to accelerate the process, which is important for practical application. Prospects for further research include improving AI algorithms for stylistic editing and studying hybrid methods that combine automation with human expertise. Keywords: Natural Language Analysis; Neural Networks; Proofreading Quality; Text Processing; Text Proofreading 1. INTRODUCTION The evolution of artificial intelligence (AI) algorithms has transformed the paradigm of editing and proofreading texts. This has made the process more accessible, faster and more cost-effective [1]. Algorithms integrated into platforms such as Grammarly, DeepL, and ProWritingAid are capable of processing significant amounts of text data. They detect a variety of errors and offer corrections in real time [2]. Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9605 However, despite rapid progress, the issue of the effectiveness of these systems compared to professional editors remains open. This especially applies to stylistic naturalness, readability, and contextual adaptation of texts [3]. The problem is that algorithms, while demonstrating high accuracy in detecting spelling and grammatical errors, often do not take into account stylistic, genre, and semantic nuances. This can lead to the creation of technically correct but stylistically inappropriate texts [4],[5]. In contrast, professional editors provide higher stylistic harmony and naturalness due to their experience [6]. The focus of this study is on a comparative analysis of the quality of text editing by AI algorithms and professional editors. The study includes an analysis of the quality of error correction, identification of algorithm weaknesses, and study of user experience. The obtained results are intended to deepen the understanding of the capabilities and limitations of AI. This will also contribute to the improvement of technologies and provide valuable insights for the optimal combination of human experience and automation. The aim of the study is to identify and analyse the impact of different AI algorithms on the efficiency and quality of automatic editing and proofreading of English texts in an international context. The objectives of the study: - Compare the quality of automatic editing and proofreading of English texts performed by different AI algorithms with the quality of editing performed by professional human editors; - Identify the types of errors (grammatical, spelling, stylistic, punctuation) that are most effectively detected and corrected by different AI algorithms; - Study the perception and evaluation of the results of automatic editing and proofreading by users (editors, authors) in terms of usability, accuracy of corrections, and time spent on the process. This study differentiates itself from prior work by its applied, comparative, and interdisciplinary nature. While previous research has often focused on isolated aspects of AI editing, our motivation is to provide a holistic empirical evaluation that directly benchmarks multiple stateof-the-art AI algorithms against professional human editors across a diverse set of criteria – including not only error correction rates but also readability scores and qualitative user experience. This approach allows us to move beyond the question of ‘if’ AI can correct errors to ‘how well’ it performs across the full spectrum of editing tasks compared to the human gold standard, and to understand the practical implications for users. The conducted research has academic novelty in theoretical and practical terms. In the theoretical aspect, a comprehensive comparative analysis of the effectiveness of AI algorithms and professional editors was carried out for the first time. The analysis took into account different types of text errors, including grammatical, spelling, stylistic, and semantic. Research hypothesis. AI algorithms demonstrate effectiveness in correcting text errors, but are inferior to professional editors in ensuring stylistic naturalness and readability of texts. Research Questions: 1. Does AI edit texts in English qualitatively compared to human editors? 2. What errors in English texts are best detected and corrected by different AIs? 3. How do users evaluate the convenience and accuracy of AI editing tools? 4. Does the perception of different AI tools differ between editors and authors? 5. What is the general experience of using AI for automatic text editing? Scientific Problem. The study addresses the efficacy of AI-driven text editing and proofreading tools compared to human editors in ensuring high-quality, stylistically natural, and contextually appropriate English texts. While AI algorithms (e.g., Grammarly, DeepL, ProWritingAid) excel in detecting grammatical and spelling errors, their ability to handle stylistic, semantic, and contextual nuances remains limited. This creates a gap between technical correctness (achievable by AI) and human-like readability and coherence (typically ensured by professional editors). 2. LITERATURE REVIEW Basic concepts of artificial intelligence in the context of natural language processing (NLP). Modern natural language processing relies on transformer models that can analyse wide contexts. Lin et al. [7] emphasize the significant advantage of transformers over previous models in the error correction accuracy. However, their limitation is their focus on English texts only, which complicates their application in multilingual environments. Lin [8] emphasizes the unexpected results of transformers, deepening the understanding of their functionality, but its analysis requires better structuring. These studies touch on Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9606 fundamental concepts for understanding AI algorithms in text editing. AI algorithms for automatic editing and proofreading of texts. Deep neural networks, in particular the T5 model, demonstrate significant accuracy in text editing tasks. However, as Mallinson et al. [9] point out in their study, the hybrid approach to editing may not sufficiently cover complex stylistic errors. At the same time, the study by Jiang [10] deeply analyses the processing of large text arrays, but focuses mainly on formal aspects, bypassing stylistic nuances. The work of Kadim and Lazrek [11] makes a significant contribution to improving the processing of natural Arabic language. The authors present an extended version of the Nemlar corpus, which significantly improves the quality of linguistic analysis because of more accurate morphological parsing and tagging of parts of speech. The resource has the potential to improve the efficiency of text editing algorithms, in particular in the tasks of error correction, syntax normalization, and data preparation for deep learning. Nevertheless, the aforementioned studies are an important basis for understanding the potential of the algorithms to be evaluated in our study. Automatic error detection and correction principles. Error correction methods are based on the analysis of syntax, semantics, and style of texts. Wang et al. [12] provides an overview of existing approaches to grammatical correction, including statistical and neural models. Its advantage is the granularity of the analysis, but it does not take into account modern transformer architectures. Raju et al. [13] indicate that Large Language Models (LLMs) demonstrate high efficiency in correcting stylistic errors. At the same time, the authors note that these models can create new errors when editing complex texts. The mentioned studies provide the background to explore in detail the issue of the operation of algorithms with different types of errors, which is important for this study. Modern tools and platforms for automatic editing. Tools like Grammarly and ChatGPT have significantly changed the editing experience. Feuerriegel et al. [14] emphasize the productivity of human-AI collaboration. However, empirical evidence on the effectiveness of these platforms is lacking. Hardack [15] focuses on the ethics of using AI in editing, but her analysis is too broad. Both studies are important for outlining the strengths and weaknesses of the tools considered in our paper. Advantages and limitations of using AI in the editing process. AI offers fast and accessible editing, but sensitivity to style and context remains its weakness. Chakrabarty et al. [16] revealed the tendency of AI to standardize language, which is critical in creative texts. Gasaymeh et al. [17] emphasize the risk of losing cultural diversity, but suggest ways to minimize it. At the same time, the researchers note the increase in adaptability of texts because of semi-autoregressive models. The mentioned studies define important criteria for the comparative evaluation of AI algorithms and human editors in our study. General analysis of the literature and unresolved issues of existing research. Despite the value of the presented studies, the issue of using AI algorithms for editing natural language requires further study and improvement. A critical analysis of the academic literature indicates significant progress in the development of algorithms for automated text editing. At the same time, certain limitations in taking into account stylistic, cultural and ethical nuances are revealed. The shortcomings of a significant part of the research include the predominant concentration on technical aspects, while is strengths are the proposed innovative solutions and the potential for practical application of the obtained results. The considered theoretical principles are the basis for further empirical research on the effectiveness of AI and its synergy with professional editing. A critical synthesis of existing literature reveals that while previous studies have effectively documented the raw capabilities of AI in grammatical error correction (GEC), they often present a fragmented view. Research has predominantly followed two parallel tracks: one focusing on the technical evolution of algorithms (e.g., transformer models, F1-scores) [7, 9, 12], and another exploring broader user perceptions or ethical considerations [14, 15]. A significant gap exists in empirically bridging these perspectives— specifically, in using rigorous, comparative methodology to quantify how these technical shortcomings manifest in tangible quality metrics (like readability indices) and are perceived by professional end-users in a direct A/B test against human expertise. Most studies evaluate AI tools in isolation rather than in a controlled, comparative setting against the benchmark of professional human editing. This article analyses issues that require deeper study in the modern academic paradigm. In particular, the quality of correction of various types Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9607 of errors by AI algorithms in comparison with professional human editors will be empirically assessed. Particular attention will be paid to the ability of algorithms to ensure stylistic perfection and optimal readability of texts. In addition, user experiences regarding the productivity and ergonomics of automated editing tools will be investigated, as well as their impact on the established process of professional editorial activity. 3. METHODOLOGY 3.1. Design The research is experimental, applied, and interdisciplinary. Type – comparative analysis. Sort – empirical, using quantitative and qualitative methods. The research was conducted in three stages. 1. The first — preparatory — stage. Selection of texts of different genres with typical errors. The texts were edited by three AI algorithms (Grammarly, DeepL, ProWritingAid) and professional editors. 2. The second — empirical — stage. A comparative analysis of the quality of error correction was conducted using quantitative methods, including readability assessments (FleschKincaid, Gunning Fog) and ANOVA. Error classification was also performed. 3. The third — analytical — stage. The data from focus groups, including editors and authors, were collected for a qualitative analysis of the perception of AI algorithms. Thematic analysis identified key aspects of convenience, efficiency, and impact on the workflow. 3.2. Participants Five groups of English texts were selected for the study, reflecting genre and stylistic diversity. In particular, the texts contained news, scientific articles, advertising texts, essays and technical documentation. The volume of each text was 300–500 words, and the number of errors reached 15 per 100 words, which ensured their adequate complexity for analysis. The errors included grammatical, spelling, stylistic and punctuation aspects to assess the effectiveness of corrections in different conditions. The group of professional editors included 10 specialists with work experience from 5 to 10 years. The specialists had an education in the field of philology or linguistics and spoke English at a level not lower than C1. All editors had experience working with automated editing systems and texts of varying complexity, which ensured the representativeness of their results. Twenty people were invited to participate in the focus groups: 10 professional editors and 10 copywriters, aged 25 to 45, most of whom had higher education. The participants had between 1 and 5 years of experience working with AI algorithms. This allowed for a wide range of experiences and assessments of usability, accuracy of corrections, and impact on the workflow. The study employed strict inclusion/exclusion criteria to ensure methodological rigor. Texts were selected based on genre representation (news, scientific articles, advertising, essays, technical documentation), length (300–500 words), and error density (15 errors/100 words). Non-English texts, highly specialized content (e.g., legal/medical jargon), and multimodal materials were excluded. Professional editors were required to have 5–10 years of experience and C1+ English proficiency; AI tools were limited to widely adopted, NLP-driven platforms (Grammarly, DeepL, ProWritingAid). These criteria controlled variables while reflecting real-world editing scenarios. 3.3. Study variables This study precisely operationalizes variables within a comparative framework, assessing the efficacy of AI-driven versus human editing. The independent variable is the editing method, categorized into four groups: Grammarly, DeepL, and ProWritingAid (representing AI algorithms), and professional human editors (as the control). Dependent variables quantify editing quality through readability metrics (Flesch-Kincaid Grade Level, Gunning Fog Index), error-correction accuracy (F1-score, BLEU score), and error classification rates across grammatical, spelling, stylistic, punctuation, and semantic categories. User experience is qualitatively evaluated via focus groups, capturing perceived usability, speed, and flexibility. To ensure validity, control variables standardize experimental conditions, including text characteristics (genre, length, error density), editor expertise (experience, English proficiency), and technical settings (uniform AI tool configurations, standardized NLP protocols for metric computation). This design is crucial for isolating the impact of editing methods on outcome variables. 3.4. Data collection 1. The experiment involved comparing the quality of editing English texts performed by Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9608 different AI algorithms and professional editors. For this purpose, texts of varying complexity with typical grammatical, stylistic and other errors were selected. Each text was sequentially processed by several common AI algorithms and an independent group of professional editors.[18] 2. Content analysis aimed to identify patterns in the work of editing algorithms. Errors were classified into the following categories: grammatical, spelling, stylistic, punctuation, and semantic. It was explored which errors were corrected, omitted or created.[19] 3. A comparative analysis of the stylistic characteristics (readability and naturalness) of texts edited by AI algorithms and professional editors was conducted. For this purpose, metrics from the field of natural language processing (NLP), in particular the Flesch-Kincaid and Gunning Fog indices, were applied. Each edited text was evaluated according to readability and language complexity indicators.[20] 4. Focus groups were used to collect qualitative data on the experiences of users of automatic text editing tools. Editors and authors who use similar solutions were involved. During moderated discussions, the participants discussed the ease of use of algorithms, their effectiveness and impact on the workflow. The obtained data were analysed to identify key themes and trends. The study used several groups of tools. Popular AI algorithms such as Grammarly, Deeple, and ProWritingAid were used for automatic text editing. These models were chosen because of their popularity among users and their ease of use. The study used the Flesch-Kincaid and Gunning Fog metrics to assess the quality of editing, which determined the level of readability of texts. BLEU was also used to compare automatically corrected texts with reference ones, and F1 score was used to assess the accuracy of error detection and correction. The Flesch-Kincaid and Gunning Fog readability indices were calculated based on the average sentence length, the number of syllables in words, and the proportion of complex words in the text. The BLEU metric was determined by comparing n-grams (up to the 4th order) of automatically edited texts with versions prepared by professional editors. F1 score was calculated as the harmonic mean of accuracy and completeness based on comparing corrected errors with the total number of existing errors. All metrics were applied using automated tools and provided an objective analysis of the quality of editing. The data obtained during the focus groups were recorded using audio recordings and transcribed as transcripts. The data analysis was performed using thematic analysis software such as NVivo and coding schemes for error classification. SPSS and Python packages with appropriate modules were used for statistical analysis. Tableau tools and Python modules such as Matplotlib and Seaborn were used for visualization of the results. 3.5. Statistical techniques and procedures In the statistical analysis, the sample size was 200 units, of which each group included: Grammarly algorithm (n=50), DeepL algorithm (n=50), ProWritingAid algorithm (n=50), editors (n=50). A set of methods was used for statistical analysis to ensure the reliability of the conclusions. ANOVA compared the average editing qualities between AI and editors. Student’s t-test detailed pairwise differences. Chi-square test revealed patterns in the types of errors processed. Descriptive statistics assessed the basic characteristics of the data. Qualitative thematic analysis of focus groups revealed key user views. The selected methods combine quantitative and qualitative approaches, which provide a comprehensive picture of the effectiveness and peculiarities of the work of AI algorithms and human editing. Such a methodology ensures the reliability of the results and their practical significance. All statistical analyses in this study were conducted with a 95% confidence level (α=0.05), aligning with conventional standards in linguistic and computational research. This threshold established a critical p-value of 0.05 as the boundary for statistical significance. Accordingly, outcomes with p<0.05 were deemed statistically significant, indicating a ≤5% probability that observed differences arose by chance. For pairwise comparisons following ANOVA, the Bonferroni correction was applied to control family-wise error rates, tightening the critical p-value to 0.0083 (0.05/6 comparisons). Significant results (p<0.05) were further contextualized via effect sizes (η2≥0.01 = small, ≥0.06 = medium, ≥0.14 = large), ensuring findings reflected both statistical and practical relevance. This rigorous framework safeguarded against Type I errors while maintaining analytical precision in evaluating AI-human editing disparities. 3.6. Instruments The study used several groups of tools. Popular AI algorithms such as Grammarly, Deeple, and ProWritingAid were used for automatic text Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9609 editing. These models were chosen because of their popularity among users and their ease of use. The study used the Flesch-Kincaid and Gunning Fog metrics to assess the quality of editing, which determined the level of readability of texts. BLEU was also used to compare automatically corrected texts with reference ones, and F1 score was used to assess the accuracy of error detection and correction. The Flesch-Kincaid and Gunning Fog readability indices were calculated based on the average sentence length, the number of syllables in words, and the proportion of complex words in the text. The BLEU metric was determined by comparing n-grams (up to the 4th order) of automatically edited texts with versions prepared by professional editors. F1 score was calculated as the harmonic mean of accuracy and completeness based on comparing corrected errors with the total number of existing errors. All metrics were applied using automated tools and provided an objective analysis of the quality of editing. The data obtained during the focus groups were recorded using audio recordings and transcribed as transcripts. The data analysis was performed using thematic analysis software such as NVivo and coding schemes for error classification. SPSS and Python packages with appropriate modules were used for statistical analysis. Tableau tools and Python modules such as Matplotlib and Seaborn were used for visualization of the results. 4. RESULTS It is necessary to compare the technical characteristics of different AI models to fully understand the predicted differences between the results. Table 1 provides a comparison of the algorithms of the models used in the study. The table demonstrates the strengths of each tool depending on their main purpose, which can help in choosing the optimal solution for specific tasks. Table 1: Technical Comparison Of Algorithms And Models Of Grammarly, Deepl, And ProWritingAid Characteristics Grammarly DeepL ProWritingAid Model type GPT (with NLP optimization) Transformer (advanced translation) GPT-3 with stylistic modifications Main specialization Automatic editing: grammar, style, tone Machine Translation (MT) and Text Editing Stylistics editing, text analysis Grammar algorithms Rule-based + context-aware neural networks Syntactic analysis with translation focus Linguistic analysis with NLP metrics Stylistics module Tone hints, rewriting recommendations Limited (mainly semantics) Deep style analysis, text tone selection Types of errors it corrects Grammar, spelling, style, punctuation Grammatical, semantic (for translation) Stylistic, spelling, semantic Quality metrics (score) F1-score: 0.92, BLEU: 0.87 BLEU: 0.94 (in translation), F1: 0.84 F1-score: 0.88, BLEU: 0.83 Context support Advanced (target audience consideration) Localization for different Languages Medium (focus on authorial styles) User tools Multi-platform, tone tips Translation localization, simple interface Multi-component style analysis Integration complexity High (deep API integration) Average (localized APIs) Medium (settings for editors) Processing speed High Very high High User-centricity Multi-aspect editing, recommendations Focus on translation accuracy Suggestive editing, multistyle approach Typical audience Content writers, editors, students Translators, business users Authors who focus on style Source: created by the authors of the research So, Grammarly stands out with its comprehensive approach to text editing, offering grammatical, stylistic, and tonal cues. DeepL specializes in translation with high localization accuracy, but has limited stylistic analysis capabilities. ProWritingAid focuses on style and readability, offering tools for deep text analysis. Each tool is optimized for different tasks and audiences, which determines their effectiveness in specific scenarios. The next step was to evaluate the readability of texts edited by three AI algorithms and human editors (Figure 1). Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9610 0 10 20 30 40 50 60 70 80 Group 1 Group 2 Group 3 Group 4 Group 5 Grammarly: Flesch-Kincaid DeepL: Flesch-Kincaid ProWritingAid: Flesch-Kincaid Figure 1:. Readability Assessment Of Texts Edited By Three AI Algorithms And Editors Source: created by the authors of the research The data analysis revealed the highest Flesch-Kincaid readability scores (70.4–78.4) for human-edited texts. The editors also showed the best Gunning Fog scores (8.7–10.5), indicating lower complexity of the texts they edited. AI algorithms had similar but higher Gunning Fog scores (9.5–12.5), indicating the need to optimize the elimination of complex syntactic structures. Overall, human editing provides a more natural and easy-to-understand language that is not yet achieved by AI algorithms. Table 2 presents the results obtained from the focus group discussions. Table 2: Results Of Thematic Analysis Of Readability Of AI-Edited Texts By Focus Groups Topic Frequency of mentions Examples of quotes Convenience 15 “The algorithm is easy to use, does not require special knowledge.” “The interface is clear, but some functions need to be configured.” Efficiency 20 “Grammar editing is at a high level, but the style is poor.” “The algorithms correct spelling well, but ignore context.” Flexibility 10 “I would like to have more options for adjusting the tone of the text.” “The system does not always adapt to the specifics of the text genre.” Impact on the Workflow 12 “The editing speed has increased significantly, but quality control is needed.” “The time to complete the project has decreased, but additional refinement is needed.” Source: created by the authors of the research Analysis of tabular data showed that focus group participants most often noted the effectiveness of algorithms (20 mentions). They emphasized their advantages in correcting grammatical and spelling errors, but noted problems with style and contextual understanding. Ease of use also received a positive assessment (15 mentions), but some functionality requires additional configuration. Flexibility and impact on the workflow were discussed less often (10 and 12 mentions, respectively), but were identified as important aspects for optimizing algorithms. The results of the ANOVA are presented in Table 3. Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9611 Table 3: Results Of The (ANOVA) Source of variation Sum of squares (SS) Degrees of freedom (df) Mean square (MS) F-value p-value Etasquared (η²) Between groups (Grammarly, DeepL and ProWritingAid) 325.4 3 108.47 12.56 < 0.001 0.45 Within groups 392.6 36 10.91 - - - Total 718.0 39 - - - - Source: created by the authors of the research The results of the analysis of variance (ANOVA) showed statistically significant differences between groups (Algorithms and Editors) in the quality of text editing, F(3, 36) = 12.56, p < 0.001. The Eta-squared value (η² = 0.45) indicates a large effect, indicating that group affiliation (algorithm or editors) explains 45% of the total variance in editing quality. The high sum of squares between groups (SS = 325.4) compared to the within-group (SS = 392.6) confirms significant differences between editing methods. The results indicate that the quality of work of editors and algorithms has significant differences, which may be important for choosing an appropriate editing tool. A classification of errors that most often remain uncorrected or are created by modern text editing algorithms has been developed on the basis of the results and tables. The classification includes five main categories of errors (Figure 2). Figure 2: Classification Of Typical Errors In The Operation Of Modern Automated Editing Algorithms Source: created by the authors of the research AI algorithms demonstrate high efficiency in detecting and correcting elementary grammatical and spelling errors. However, their capabilities are significantly limited when processing contextually dependent stylistic and semantic aberrations. The developed taxonomy of linguistic errors provides detailed information about the weaknesses of modern algorithms, which determines the directions of future research and development. The results of the study confirmed the hypothesis of higher stylistic naturalness and readability of texts edited by humans. AI algorithms effectively corrected grammatical and spelling errors. However, their work with stylistics turned out to be less precise, as evidenced by qualitative focus group analysis. ANOVA revealed statistically significant differences between editing methods, emphasizing the advantages of editors in creating natural texts. 5. DISCUSSION The findings of this study enter a growing academic conversation concerning the efficacy of AI-driven text editing. Our results are largely consistent with the body of prior work that acknowledges the proficiency of AI in grammatical and spelling correction while highlighting its shortcomings in stylistic and semantic tasks [2, 3, 16]. For instance, our data robustly supports the conclusions of Li et al. [21] regarding the superiority of human expertise in handling nuanced aberrations, and aligns with Heintz et al. [2] on the high accuracy of tools like Grammarly for grammatical errors. Journal of Theoretical and Applied Information Technology 30th November 2025. Vol.103. No.22 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 9612 The asserted superiority of human editors in stylistic and readability aspects [3, 6, 21] is empirically validated, moving beyond qualitative claims with quantitative evidence (e.g., FleschKincaid: 75.2 vs. 69.4; η²=0.45). While tools like Grammarly demonstrate grammatical proficiency [2, 22], their relative performance nuances against other AI tools and human editors are revealed by the comparative design, which also precisely categorizes their persistent weaknesses through a novel error taxonomy (Figure 2). This work thus consolidates fragmented prior knowledge into a unified, evidence-based framework, offering a benchmark for future development. The obtained empirical data correlate with the conclusions of Li et al. [21], which emphasize the superiority of human expertise in correcting stylistic and semantic aberrations. At the same time, the use of tools such as Grammarly and DeepL has demonstrated high accuracy in identifying grammatical errors. This is consistent with the results of the research of Gupta et al. [22]. However, the obtained results are somewhat contradictory to the findings of Ara and Sik-Lanyi [23], which express skepticism about the ability of AI to correct complex multi-level errors. A comparative analysis of the quality of text editing revealed a statistically significantly higher average Flesch-Kincaid index in professional editors (75.2) compared to AI algorithms (69.4). The obtained data are consistent with the findings of Erslev [24], which emphasize the readability advantage of texts edited by humans over automated systems. At the same time, the study by Shanthi et al. [25] indicates the ability of algorithms to achieve comparable results when processing standardized texts. The data demonstrated in the study of the team of researchers partially correlates with the results obtained within standardized syntactic structures. AI algorithms are better at editing texts with standard word order and simple grammatical rules. The classification of detected linguistic aberrations showed that AI algorithms demonstrate the highest efficiency in correcting grammatical and spelling errors. However, correcting stylistic and semantic inaccuracies remains a significant problem for them. The obtained results are consistent with the data of Ma et al. [26], which show that modern generative models, in particular GPT-like architecture. AI models are mainly optimized for formal-grammatical correction, but often make errors in context-dependent stylistic decisions. At the same time, the research of Fiorillo and Mehta [27] expresses justified optimism about the potential of future iterations of algorithms to overcome these limitations by implementing more complex models. User feedback showed a high assessment of the speed and ease of use of AI by editors and authors. However, they indicate the limitations of algorithms in complex editorial tasks. Similar conclusions are presented in the study of Algahtani [28], which emphasizes the irreplaceability of human expertise in the final stages of text processing. At the same time, Balluff et al. [29] suggest the potential superiority of AI over humans in certain scenarios. In particular, this concerns the editing of large arrays of text data containing typical and predictable errors. Our study confirmed the conclusions of earlier studies and outlined key vectors for further improvement. The results of the thematic analysis of focus groups demonstrated the need for contextual and stylistic adaptation of algorithms. This need is also emphasized in the study of Sani [30]. It should be noted that the obtained empirical data emphasize the importance of synergistic AI integration with the human editing process. Such a symbiotic approach can be the foundation for further developments in this field. A detailed classification of errors has been developed identifying the weaknesses of modern editing algorithms. Furthermore, the study deepens the understanding of the interaction of AI algorithms with the stylistic context of text materials. The practical significance is the possibility of using the results to optimize existing automatic editing tools, such as Grammarly, DeepL, and ProWritingAid. The results confirmed the research hypothesis, showing that AI algorithms are inferior to editors in ensuring stylistic naturalness and readability of texts. At the same time, the algorithms demonstrated high efficiency in correcting grammatical, spelling, and punctuation errors. The results of the study can be used by translation agencies to automate translation quality control. Editing platforms can integrate algorithms to reduce text processing time and increase efficiency. In the field of education (EdTech), the results will be useful for the development of teaching tools to improve written literacy. However, achieving stylistic naturalness of texts will remain a task that requires further research and improvement of algorithms. 5.1. Limitations The study has some limitations that need to be considered when interpreting the results. The analysis was reduced to three AI algorithms, which