ARTIFICIAL INTELLIGENCE IN THE UKRAINIAN MEDIA: Integration, Challenges, and Sector Needs

Collection and Analysis of Publicly Available Sources

The collection and analysis of publicly available sources (scraping desk research) covers the period from 2023 to the end of April 2026. An approach combining elements of desk research and a scoping literature review—referred to as scraping desk research—was developed for data collection and analysis. 

The choice of this hybrid method is dictated by the nature of the project: it is not an academic study and does not involve a full-fledged systematic literature review, but it does require a sufficient level of methodological transparency and reproducibility. Scraping desk research—or the automated collection of secondary information from publicly available sources—ensures structured and consistent data collection within the limits of available resources. From the scoping literature review—that is, a review of the literature on a given topic—we adopted the definition of time frames, languages, and keywords prior to the start of the search, a sequential search organized by annual cycles, and a two-stage classification of texts into thematic groups. From desk research, we adopted the inclusion of non-academic sources, the use of general search engines, and a focus on industry materials alongside academic ones. Thus, this hybrid approach ensured a structured and consistent collection of data within the limits of the project’s available resources.

At this stage, the sample includes materials written in Ukrainian and English. The pool of materials comprises both academic and non-academic texts—since academic literature cannot keep pace with the rapid changes in AI development and, in particular, in the editorial practices of media organizations. Instead, industry reports and analytical materials from international organizations fill this gap, allowing for an analysis of the phenomenon at both the theoretical and practical levels. 

Google Scholar was used to collect academic texts. The search keywords were formulated based on the research questions and specified in two languages—Ukrainian and English. 

Keywords in English: AI, journalism, media, integration, barriers, adaptation, newsrooms. 

Keywords in Ukrainian: ШІ, журналістика, медіа, інтеграція, бар’єри, адаптація, редакції.

The same set of keywords was used to search for both academic and non-academic texts.

The search for academic materials was conducted sequentially, in annual cycles. In the first step, all texts published in 2023 were filtered using the specified keywords; subsequently, a separate search query was performed for each subsequent calendar year through April 2026 inclusive. If no relevant material was identified on the search results page (SERP—Search Engine Results Page), the researcher moved on to analyzing the next annual cycle without further iterations.

The next step was to review and select relevant texts, followed by their analysis and classification. This process took place in two consecutive stages. The first stage involved assessing the relevance of articles using large language models (LLMs): the claude-haiku-4-5-20251001 model from Anthropic and the gpt-4.1-mini model from OpenAI, with the researcher verifying the results. Irrelevant articles were excluded from further analysis. Due to the small number of Ukrainian-language texts, their relevance was assessed by the researcher without the aid of LLMs. 

We manually labeled the relevance of 49 articles on a scale of one to five, where 5 represents the highest relevance, and compared this with the scores from the two LLMs mentioned above. We then converted these scores into a binary classification: scores from one to four were considered irrelevant, while a score of 5 was considered relevant. The results were verified using the metrics of precision and recall, which compare the LLM’s rating with the human rating, where the latter is considered the correct answer (the “gold standard”). Precision is calculated as the proportion of true positives—that is, articles that a human identified as relevant—among all articles marked as relevant by the LLM. Recall is defined as the proportion of articles identified as relevant by the LLM among all truly relevant articles. 

Texts that both models identified as highly relevant were classified into thematic groups, with each text potentially belonging to one or more groups. In the same prompt, each LLM also identified barriers to AI implementation, if any were mentioned in the article. Classification into thematic groups was performed using the aforementioned LLMs (the claude-haiku-4-5-20251001 model from Anthropic and the gpt-4.1-mini model from OpenAI) based on predefined categories from a preliminary qualitative review of the texts collected by the researcher. The results of the thematic grouping were also verified: the thematic groups for 34 articles, which had been manually classified by the researcher, were compared with the groups identified by the LLMs for those articles. The prompt we used for automated classification is available in the appendix. 

The following metrics were used for verification: Jaccard coefficient, primary-group accuracy, precision, and recall. The Jaccard coefficient is a measure of similarity between sets, calculated as the number of categories assigned by both the human annotator and the model (the intersection), divided by the total number of unique categories in their combined sets (the union). That is, if a person assigns the categories {A, B, C} to an article, and the model assigns them {B, C, D}, then the common part is {B, C}, the union is {A, B, C, D}, and the Jaccard coefficient is 2/4, or 50%.

The accuracy of the primary group is the proportion of articles where the model’s primary thematic group (the first in the list) matches the true primary thematic group. Precision and recall were calculated as the average of the corresponding metrics for each article. That is, for each article individually, we calculated what proportion of the thematic groups assigned by the model matched those assigned by humans (precision), and what proportion of the human labels were identified by the model (recall)—then these values were averaged across all 34 articles.

To analyze the overlap between thematic groups, we represented them as a graph and applied the “eigenvector centrality” metric: it assesses the importance of a node not only by the number of its connections but also by the importance of the nodes to which it is connected. Higher values of this metric indicate topics that occupy a central position in the co-occurrence structure.

To classify barriers—if they were mentioned in the articles and identified by the LLM during clustering into thematic groups—we used a two-stage approach employing the LLM. Classification was necessary because the identified barriers were often synonymous or overly detailed. In the first stage, using a separate prompt (see “Appendices”), the model matched each barrier—if mentioned in an article—to a more general category, or created a new one if none of the existing ones were suitable. In the second stage, we classified the identified general categories into abstract categories using another prompt (also in “Appendices”) through the same process. After processing the entire corpus, certain general categories were moved to other abstract categories following a manual review. A complete list of abstract categories of barriers, along with their general categories, is also provided in the appendix.

This approach to classification allowed us to organize the collected literature into clear thematic areas, identify major research trends, and note groups of barriers to AI implementation in newsrooms that were mentioned in the literature. Additionally, it provided a better understanding of existing patterns of AI use in the media environment based on secondary research.

The search for and selection of non-academic texts (reports) followed a selective approach: we chose well-known media and nonprofit/nongovernmental organizations and their reports from the specified time period (2023–2026). This selective approach was adopted for several reasons. First, to avoid duplication of materials: a large share of the texts already included in the academic corpus were also present on Google Search’s search engine results pages (SERPs) during the search for and selection of non-academic texts. Second, the non-academic information space is significantly less structured compared to academic databases. Search results on Google Search depend on ranking algorithms, personalization, SEO optimization, and the popularity of individual resources, which complicates the formation of a stable and reproducible sample based on the principles of systematic search. Under these conditions, a comprehensive selection of all relevant materials does not guarantee a higher-quality sample; rather, it significantly increases the volume of low-relevance or duplicate results. 

A selective approach allowed us to focus on materials containing aggregated empirical data, survey results, case studies, and analytical conclusions, rather than on individual news reports or opinion pieces. This ensured greater comparability between non-academic sources and academic literature and enhanced the analytical value of the collected corpus. Because such reports focus on the practical aspects of journalistic work in the age of AI, they formed the basis for analyzing practical approaches to the use of AI in newsrooms. Specifically, the analysis examined which tasks AI addresses and how AI tools are used. In summary, the use of a selective approach to the selection of non-academic texts was dictated by the desire to ensure the relevance, analytical quality, and manageability of the sample given the high heterogeneity and dynamism of the non-academic information environment.

The sample of non-academic texts included works by Nieman Lab, the Reuters Institute for the Study of Journalism, the European Broadcasting Union, FT Strategies, the UNC Hussman School of Journalism and Media, the Lviv Media Forum, and the Deutsche Welle Akademie. In particular, the last two organizations—the Lviv Media Forum and the Deutsche Welle Akademie—covered the context of Ukrainian newsrooms. These organizations were included in the sample based on three criteria. First, the organization must have recognized expertise in the field of journalism, media, or the digital transformation of newsrooms. Second, its materials had to contain systematic analysis, research findings, survey results, or practical case studies regarding the use of artificial intelligence in journalism. Third, the organization had to regularly publish analytical materials throughout the study period (2023–2026), which made it possible to track changes in editorial practices over time. In total, the sample comprised 16 sources: 9 reports from the aforementioned institutions, 6 publications from the Lviv Media Forum—each focusing on a single newsroom and its approach to using AI—and one news article about a newsroom’s policies on AI use (The New York Times). We found the last article using the snowball method while reviewing materials from the aforementioned organizations.