Evaluating Generative AI for Identifying Ethical, Legal, and Social Dimensions in Migration Narratives: A Case Study of Ukrainian Discourse

Abstract

Collective endorsement of shared values across diverse social groups is essential for the development and sustainability of democratic societies, yet capturing the perspectives of marginalised populations remains a persistent challenge, particularly when examined through ethical, legal, and social (ELS) lenses. This study develops a structured Migration ELS taxonomy to guide a GenAI-assisted semantic classification model designed to identify ELS dimensions in textual data. The model is fine-tuned and evaluated within a human-in-the-loop framework using expert annotations to ensure reliability and interpretive accuracy. As an empirical case, the approach is applied to migration-related official policy documents and narratives of Ukrainian migrants published on the Telegram platform. The resulting framework enables the analysis of alignment between governmental and migrant perspectives, revealing thematic and temporal divergences in ELS dimensions across institutional and user-generated discourse. The findings demonstrate the potential of this scalable framework, which combines taxonomy-driven modelling with generative AI and expert-in-the-loop validation, to reveal patterns of alignment and temporal dynamics in the representation of values across different social groups.

1. Introduction

One of the fundamental priorities in the development of democratic societies is the enhancement of social cohesion, a core dimension of which encompasses institutional trust and the collective endorsement of shared values across all social groups (United Nations High Commissioner for Refugees (UNHCR) 2024b). Obviously, the stability of communication between the government and citizens, facilitated by long-established communication and feedback mechanisms such as the media, social surveys, and monitoring assessments, as well as more recent open government platforms (Schmidthuber et al. 2021), plays a crucial role in strengthening social cohesion.

At the same time, the reflections and narratives of marginalised segments of society, for instance, refugees or immigrants, are not always immediately apparent to governmental institutions, often due to limited representation within conventional feedback mechanisms, constraints in outreach efforts, or language barriers. Traditional data collection methods, such as social surveys, often fail to engage enough individuals from these closed communities (Ojo et al. 2024). Additionally, marginalised individuals may lack awareness of existing communication and feedback mechanisms, further widening the gap in understanding their reflections and communicative expressions (Reddick et al. 2017).

A possible solution for integrating the opinions and views of all social groups is the analysis of large volumes of user-generated data from social networks, which enables a more comprehensive representation of the full range of opinions, ways of thinking, and perspectives within society. At the same time, the analysis of large-scale data offers unique insights into the lived experiences, concerns, and opinions of marginalised communities, while allowing this to be done without accessing personal or private information. Big data technologies and artificial intelligence (AI) methods, including generative AI (GenAI), can reveal critical trends in internal discourse across different population groups and detect misalignments between government policy and public sentiment that are often overlooked in traditional policy processes.

Moreover, in analysing societal discourse, it is essential to consider the convergence of social groups around shared public values, the negotiation of which plays a key role in fostering social cohesion. These values encompass the ethical, legal, and social aspects embedded in public communication. Particular attention should be paid to examining how these ethical, legal, and social dimensions are manifested in narratives related to marginalised groups, as well as to assessing the interaction between governmental regulation and public responses across these three normative domains. The successful alignment of policies and public opinion with the ELS dimensions ensures that institutional actions resonate with societal values and are recognised and shared across all social groups, thereby fostering long-term trust and cooperation (OECD 2022).

However, despite the growing use of generative AI for analysing large-scale social data, the extent to which these models can reliably capture key aspects of shared value narratives across different social groups remains uncertain. In particular, it is unclear whether generative AI can effectively identify and interpret ethical, legal, and social dimensions embedded in complex textual content. To address this issue, the study examines the following research question:

[RQ1] Can generative AI models identify the ethical, legal, and social (ELS) components embedded in textual content?

To examine this question, the study employs a human-in-the-loop semantic classification approach, in which human experts guide and fine-tune the model while also validating its outputs.

As an empirical case to demonstrate the capability of generative AI to identify ELS dimensions in textual data, the study examines the alignment between governmental narratives and those of Ukrainian migrants, who were displaced as a result of the Russian invasion. Considering Ukrainian migrants as a case of a marginalised group provides a unique opportunity for analysis due to the scale and urgency of the crisis: over six million Ukrainians were forcibly displaced by the end of 2024, according to the United Nations High Commissioner for Refugees (UNHCR) (2024a). This humanitarian emergency not only exposed the limitations of existing regulatory frameworks but also will be able to test the adaptability of public institutions in addressing key needs across ethical, legal, and social dimensions.

Accordingly, as a practical demonstration of the applicability of this approach, the study examines the extent to which European governmental regulations align with the lived experiences and normative expectations of Ukrainian migrants, viewed through the lenses of ELS dimensions. To this end, the study seeks to answer the following second research question:

[RQ2] How can the alignment between governmental and Ukrainian migrant narratives be evaluated across ethical, legal, and social aspects?

The research focused on the twelve main host countries for Ukrainian refugees: Poland, Canada, the United States, Slovakia, Moldova, Germany, the Czech Republic, Spain, the United Kingdom, Italy, Sweden, and Ireland (Fransen and De Haas 2022). It analyses Telegram messages posted by Ukrainian migrants between February 2022 and February 2024 in these countries and compares them with official EU migration-related regulations.

The study therefore combines two complementary objectives. First, it contributes methodologically by developing and evaluating a GenAI-assisted framework for identifying and analysing ethical, legal, and social dimensions in complex textual data. Second, it demonstrates the applicability of this framework through an empirical analysis of governmental and Ukrainian migrant narratives, using migration discourse as a real-world case study. The methodological and empirical components are intentionally integrated, as the migration case both motivates and validates the proposed analytical approach.

To address these two research questions, the study adopts an interdisciplinary approach that integrates generative AI with computational social science methods. Specifically, the analysis proceeds in four main steps: (i) the development of the Migration ELS Taxonomy, comprising hierarchically organised ethical, legal, and social dimensions; (ii) the application of a GenAI-assisted semantic classification model fine-tuned using this taxonomy; (iii) the validation of the model’s performance through comparison with expert annotations to assess reliability and interpretive accuracy; and (iv) the analysis of temporal and thematic variations in ELS dimensions across official policy documents and Ukrainian migrant narratives.

In this study, the term “migration narratives” refers to discursive and thematic representations of forced migration experiences, refugee integration, institutional responses, and public communication related to the displacement of Ukrainians following the Russian invasion. Although the analysed population primarily consists of refugees, the broader term “migrants” is occasionally used in relation to migration-related discourse and communication contexts discussed in the prior literature and policy documents.

The remainder of the paper is structured as follows. The Methods section presents the development of the ELS taxonomy and the implementation of the generative AI-based classification approach to address RQ1. The Results section applies this framework to analyse governmental and migrant narratives, addressing RQ2.

2. Literature Review

2.1. Generative AI for Analysing Social Discourse

The rapid expansion of social media platforms has produced vast amounts of user-generated content, including text posts, comments, images, and videos. Applying GenAI to the analysis of this large-scale data presents both challenges and opportunities for researchers aiming to understand public sentiments, opinions, and attitudes toward diverse societal issues. One of the key applications of GenAI in this field is sentiment analysis, reflecting a growing recognition of its potential for understanding and forecasting public opinion. For instance, Cerina and Duch (2026) presented a methodology for public opinion research by employing GenAI to derive structured, survey-like insights from unstructured social media content. Similarly, Wei et al. (2025) proposed a computational framework that integrates an enhanced LLM with retrieval-augmented generation (RAG) to predict both the opinions of influential figures and the emotional responses of general users. These approaches effectively anticipated public sentiment trends by synthesising influencer perspectives and aligning them with broader social reactions. Recent studies (Li et al. 2023; Krugmann and Hartmann 2024) demonstrated the successful use of LLM-based social media sentiment analysis to capture and forecast nuanced consumer emotions, opinions, and perceptions that shape social dynamics.

Other recent research also showed that GenAI-based approaches extend traditional sentiment analysis by enabling the identification, clustering, and interpretation of thematic patterns in large-scale social conversations (Wang and Zhai 2017). Unlike traditional unsupervised models such as Latent Dirichlet Allocation (LDA) and Non-Negative Matrix Factorization (NMF), which primarily identify statistical co-occurrences of words, GenAI models were shown to be capable of capturing higher-order semantic relationships and contextual dependencies between themes (Chiarello et al. 2024; Matwin et al. 2023). This shift from probabilistic to generative modelling expanded analytical precision and allowed for a better representation of the complexity of social discourse. As explored by (Khairova et al. 2024), identifying the thematic distribution of social attention and opinion dynamics over time can contribute to strengthening social cohesion by revealing the issues that matter most to specific communities. In turn, Gokcimen and Das (2024) applied topic modelling to climate change discourse, uncovering dominant themes, emerging trends, and key stakeholders.

In recent years, LLMs have demonstrated the capacity to mirror how humans structure knowledge and meaning within specific domains, encapsulating this through the conceptual organisation of language. According to (Du et al. 2025), this capability was evident in their ability to answer questions, generate coherent text, and infer information about entities, objects, and properties that they had not directly encountered. Further evidence was provided by (N. Xu et al. 2025) and (Q. Xu et al. 2025), who argued that although LLMs do not possess grounded representations beyond text, the relational structures they capture among linguistic concepts can approximate the cognitive organisation of conceptual knowledge. In their study, Patel and Pavlick (2022) examined whether LLMs were able to map widely recognised conceptual categories, such as spatial directions, cardinal directions, and colour names, onto unseen textual contexts. The results suggested that these models were able to successfully align such abstractions, including emotional and temporal concepts, with coherent internal representations.

There is a growing research focus on fine-tuning LLMs to convert unstructured text into structured conceptual representations capable of capturing complex and overlapping issues. One example is the work by (Perin et al. 2023), who employed fine-tuned models to extract structured representations from unstructured text, identifying 51 distinct concepts with a maximum precision of 0.71. Similarly, Wan et al. (2024) developed taxonomies of user intent and conversational domains based on transcripts from Microsoft’s Bing Consumer Copilot system. The study employed GPT-4 and GPT-3.5-Turbo models for taxonomy generation and evaluation. These taxonomies, comprising 10 and 25 categories, respectively, were subsequently used to fine-tune the LLM for text classification, with performance evaluated using Cohen’s Kappa coefficient. In recent years, increasing attention has been directed toward applying LLMs to the classification of abstract conceptual categories, including behavioural emotions (Merrill et al. 2021) and societal and cultural norms (Perin et al. 2023).

Taken together, these studies demonstrate a progressive shift from surface-level language modelling toward the effective interpretation of complex, abstract, overlapping, and often morally significant categories. However, while existing research highlights the utility of GenAI in extracting structured sentiment, identifying influential opinions, and mapping broad public discourse, it rarely addresses how such tools can be used to compare divergent social narratives through the lens of ethical, legal, and social norms. This gap motivates the present study, which investigates whether generative AI models can systematically identify ELS dimensions in textual data.

2.2. Taxonomic Challenges in Analysing ELS Dimensions from Migration Narratives

Despite the growing application of GenAI in analysing social media content, several challenges remain unresolved. As noted by (Chimbga 2023), key issues include ethical concerns related to data privacy, algorithmic and data bias, and the potential misuse of digital content. Further complications arise when attempting to extract insights from social media generated by marginalised communities, such as migrant groups. These challenges stem not only from the limitations and inconsistencies of official migration records but also from the limited accessibility and visibility of migrants’ own perspectives. Datasets related to migration often suffer from incompleteness, inaccuracies, and inconsistent reporting standards across countries and institutions (Ahmad Yar and Bircan 2025). For example, differences in classification practices, as well as underreporting, can distort the interpretation of migration patterns. Moreover, these datasets frequently lack information about migrants’ motivations or socioeconomic backgrounds, which restricts the ability of researchers and policymakers to fully understand the complexity of migration flows (Sahin-Mencutek 2020).

Importantly, quantitative data alone cannot capture the complex and context-dependent aspects of migrant narratives. A more comprehensive understanding therefore requires the integration of qualitative insights, particularly those derived from social network content. However, analysing such textual data introduces additional challenges, including information oversaturation (Abkenar et al. 2023) and linguistic limitations associated with low-resource languages (Dash and Ramamoorthy 2019). In such cases, automated translation may fail to preserve linguistic or cultural nuances essential for accurate analysis (Vanmassenhove et al. 2019).

These challenges, particularly those related to linguistic variability and the interpretation of complex textual data, further complicate the integration of social media data with official datasets. Therefore, despite its potential, social media data remains difficult to integrate with official sources, especially when working with migrant narratives. Existing methods still lack robust frameworks to bridge this gap. Building on these methodological challenges, recent research has increasingly turned toward the development of taxonomies as a way to organise complex migration-related textual information.

One illustration of this approach is provided by (Mendola and Pera 2022), who developed a multilayered taxonomy of refugee vulnerability that considers multiple dimensions of the concept, including legal status, physical safety, socio-economic exclusion, and psychological harm. (Pisarevskaya et al. 2020) developed a meta-taxonomy of migration studies that traced the field’s evolution across disciplines, thematic domains, and methodological approaches. Their analysis identified shifting thematic clusters over time, including integration, governance, and identity. The taxonomy developed by IMISCOE (International Migration, Integration and Social Cohesion Network) and described by (Scholten 2022) provides a systematic framework for organising the field of migration studies. It categorises research across thematic areas and offers a structured overview of how the field has evolved and diversified over time.

However, these studies, while useful for structuring information related to migration processes, primarily conceptualise and classify relatively concrete dimensions of the migration experience, such as research domains, psychosocial support needs, or economic conditions (Iacus et al. 2025). Although existing taxonomies provide important insights into refugee vulnerability and methodological challenges in migration research, they rarely address the structuring of ethical, legal, and social dimensions that emerge from migration discourse.

Moreover, although interest in applying LLMs to abstract and normative domains is increasing, as demonstrated in recent work analysing ethical, legal, and social aspects in responsible AI for the agri-food sector (van Hilten et al. 2025), there remains limited research that systematically examines the capacity of LLMs to classify and interpret ethical, legal, and social norms simultaneously. This gap is particularly evident in multilingual settings and in complex domains such as migration and public policy.

To address this gap between the capabilities of generative AI and the lack of conceptual taxonomies for abstract, intersectional, and complex interpretive concepts, we propose a novel taxonomy specifically designed for analysing ethical, legal, and social dimensions in migration-related narratives. Using this taxonomy, we evaluate the ability of a generative AI model to identify and map ELS concepts across both citizen-driven migration discourse and corresponding governmental policy texts.

3. Materials and Methods

This section addresses RQ1 by presenting the development of the Migration ELS Taxonomy and the generative AI-based framework used to identify and classify ethical, legal, and social dimensions in textual data. It also describes the data collection and preprocessing procedures underpinning the analysis.

3.1. Data Collection

To achieve the objectives of this study, two datasets were utilised: (1) Dataset 1, consisting of official governmental publications from multiple countries addressing regulations related to Ukrainian refugees (Redozub and Khairova 2025), and (2) Dataset 2, comprising textual messages posted by Ukrainian refugees on public Telegram channels (Khairova et al. 2025). Data collection for both datasets was guided by the query: “migration” OR “migrants” OR “refugee” AND “Ukrain*”, ensuring thematic relevance to Ukrainian refugee issues across both institutional sources and user-generated social media content.

The data selection process for Dataset 1 involved collecting official governmental documents from a variety of national and international sources. Web scraping was performed using Python’s BeautifulSoup library (beautifulsoup4 version 4.11.2), and data from PDF files were extracted using PyPDF2 version 1.26.0. The LangDetect library (version 1.0.9) was used to identify the language of each document, while the WordNinja package (version 2.0.0) was applied to segment and clean raw text. For audio and video content, multilingual transcriptions were generated using the Sonix web service. In cases where automated scraping was restricted or unfeasible due to the limited scale or structure of certain sources, manual extraction methods were applied.

In total, Dataset 1 comprises documents collected from 34 information sources across multiple countries, resulting in a dataset of 708 official texts in English, Ukrainian, and Russian. These documents address issues related to Ukrainian refugee migration and living conditions abroad between February 2022 and February 2024. A substantial proportion of the dataset (33.05%) was sourced from the Global Data Institute of the IOM, as well as from the official websites of the IOM and its Ukrainian branch. Government portals of host countries contributed 32.06% of the documents, while an additional 15.25% originated from national ministries and other governmental institutions. A further 12.15% of the documents were collected from the State Migration Service of Ukraine and migration services in host countries. Social media and multimedia sources were also included, with 1.98% of the documents derived from Twitter posts by government officials and 5.51% transcribed from official YouTube channels.

Table 1. Distribution of official documents by selected information sources (Dataset 1).

Source of Documents	Source Details *	Number of Documents
International Organization for Migration	IOM Ukraine: Displacement Tracking Matrix	234
Government portals	UK, Polish, Spanish, Irish, Swedish, Italian, Czech Republic	226
Migration Agencies	German, US, Swedish, Moldovan, Czech, Canada	53
Ukraine Migration Services	State Migration Service of Ukraine	51
Foreign Ministries	Spanish, Slovak, Italian, Czech Republic	49
Ukrainian Embassies	Germany, Slovakia	18
Video Platform	YouTube	39
Social Media Platform	Twitter/X	14
Other	Czech, Slovak Education and Health Portals	25

* Depending on the source category, the column includes countries, institutions, governmental agencies, official portals, or digital platforms associated with the collected documents.

summarises the distribution of documents across these sources.

Telegram was selected as the primary platform for this research due to its central role in public discourse among Ukrainian refugees (Ghasiya and Sasahara 2023) and its accessibility for collecting textual data through open channels (Navumau and Matveieva 2025; Willaert 2023). While the analysed Telegram messages cannot be assumed to fully represent all Ukrainian migrants, previous studies have shown that Telegram constitutes one of the central communication platforms among displaced Ukrainian communities. Consequently, the discourse observed in these Telegram channels may, with appropriate methodological caution, provide important insights into broader communication and narrative patterns among Ukrainian migrants displaced by the war.

The data selection process for Dataset 2 involved collecting messages from Ukrainian refugee group chats on Telegram. Data collection focused on messages published between February 2022 and February 2024 and was conducted using the Telethon Python library, which enables automated access to publicly available Telegram content. To minimise potential geographical bias and ensure the reliability and integrity of the dataset, the selection of Telegram channels was guided by United Nations migration flow statistics identifying the main host countries that receive the largest numbers of Ukrainian refugees (United Nations High Commissioner for Refugees (UNHCR) 2024b). In total, 632,372 messages were collected.

To ensure linguistic consistency and prepare the textual data for further analysis, the texts from both Dataset 1 and Dataset 2 were processed using several cleaning steps, including removal of emojis, symbols, and icons; basic text cleaning; and word lemmatisation. To harmonise the textual content across datasets and align it with the predominantly English-language training models used in GenAI systems, Dataset 2 (messages posted by Ukrainian refugees) was translated from Ukrainian and Russian into English using the Google Translate API, while Dataset 1 (official governmental publications) was originally in English.

Table 2. Country-level distribution of official governmental documents (Dataset 1) and Telegram messages (Dataset 2).

Country	Dataset 1		Dataset 2
	Official Texts (n)	Tokens (n)	Telegram Messages (n)	Tokens (n)
Canada	2	487	1473	233,233
Czech	61	23,270	173,607	6,754,443
Germany	28	91,246	235,737	1,862,742
Ireland	27	26,055	157,858	3,712,852
Italy	17	9403	169,560	421,089
Moldova	4	1098	63,122	1,682,927
Poland	67	24,670	235,108	5,739,128
Slovakia	11	2307	94,691	2,513,625
Spain	51	36,444	43,569	1,245,337
Sweden	20	15,924	139,560	3,225,243
United Kingdom	88	70,194	60,631	1,707,510
USA	10	6699	359,514	8,731,600
Ukraine	70	23,623	-	-
IOM	252	697,128
Total	708	1,028,548	1,734,430	37,829,729

presents a comparative overview of the data distribution across the selected countries, including both official governmental documents (Dataset 1) and Telegram messages (Dataset 2).

3.2. Four-Step Methodological Framework for Measuring ELS Dimensions

To assess the interaction between government regulations and public discourse across the three normative dimensions (ELS), a GenAI-assisted mapping process was applied, followed by comparative and temporal analyses. The methodological framework is presented in Figure 1 and comprises a preliminary stage followed by four core processing steps.

The preliminary stage of the methodology involved selecting a generative AI model aligned with the analytical objectives of the study. Since the aim was to investigate the applicability of generative AI for identifying fundamental ELS components embedded in textual content, rather than conducting a benchmark comparison between multiple models, the selection process was guided primarily by practical and operational considerations. These considerations included ease of use, latency, estimated cost per million processed tokens, and the approximate cost of processing a typical unit of analysis (250 input tokens and 50 output tokens). A comparative assessment was conducted using three widely used generative AI models: LLaMA-2, Claude 3, and GPT-3.5. Although all three models demonstrated the capability to support ELS-oriented text classification, GPT-3.5 was selected due to its comprehensive documentation, stable response latency, accessible API infrastructure, and relatively lower operational costs.

The first core component of the methodology is the development of a dedicated Migration ELS taxonomy. This taxonomy is required due to the abstract, overlapping, and context-dependent nature of ethical, legal, and social categories, which lack the concrete and distinct features typically relied upon by machine learning models for reliable classification. These categories are expressed through normative language and are shaped by cultural, institutional, and situational factors (Kumari 2024). This variability is particularly evident in the context of migration, where policy frameworks, public discourse, and lived experiences intersect.

Guided by these considerations, the taxonomy was developed through an iterative, expert-informed process involving cycles of refinement, evaluation, and feedback. During the evaluation and comparative analysis stages, it serves both as a classification framework and as an interpretive bridge between human reasoning and computational processing (Akata et al. 2020).

In the second step, texts from both datasets were mapped to one of the predefined domains (ethical, legal, or social) using a generative AI model. By embedding the Migration ELS Taxonomy into the prompt design, the model was guided towards more accurate, consistent, and transparent mapping of ELS concepts across both governmental documents and migrant-authored texts.

In the third step, the capability of the generative AI model to map governmental and migrant narratives to the ELS categories was evaluated through an expert-based validation procedure. This involved human experts manually annotating a subset of the dataset according to the three predefined dimensions. Model performance was then assessed using standard metrics, including accuracy, F1-score, and Cohen’s kappa, to evaluate agreement between annotations.

In the final step, comparative and temporal analyses were conducted to examine how governmental and Ukrainian migrant narratives align across the ELS dimensions, and how this alignment evolves over time. The four-step methodological framework for using generative AI to measure alignment between official and public narratives across ELS dimensions in Ukrainian migration discourse is presented in more detail below.

3.2.1. Human-in-the-Loop Development for the Migration ELS Taxonomy

The development of the Migration ELS Taxonomy follows a four-step pipeline, beginning with the initial conceptualisation of ELS categories generated by the GenAI model and progressing toward a refined and robust definition of ethical, legal, and social concepts within the migration domain through a human-in-the-loop refinement process.

In the first step, we examined how the GenAI model represents the ELS categories by prompting it to define and describe the key features of each domain (initial conceptual mapping). Based on the model’s responses, a preliminary set of representative concepts corresponding to each category was identified, as presented in

Table 3. Preliminary representative concepts identified through GenAI-assisted conceptual mapping of ELS categories.

Ethics	Legal	Social
morality principles, decision-making responsibility, social norms agreements, professional ethics, ethical dilemmas	rules and regulations, enforcement authority protection of rights, legal system adaptability, interpretation and application	social interactions, social norms, social institutions, socialisation, social change, social stratification

.

In the second step, the taxonomy was expanded and contextualised for the migration domain through a review of the academic literature. To identify relevant publications, six keyword combinations were used in Google Scholar searches: “migration legal,” “migration legal Ukraine,” “migration social,” “migration social Ukraine,” “migration ethics,” and “migration ethics Ukraine.” For each query, the first five pages of search results were screened, and studies published from 2022 onwards were included to ensure relevance to the context of Ukrainian refugees. This process resulted in 14 ethics-related, 24 legal, and 26 social studies. Based on these sources, and supported by iterative prompting with GPT-3.5, three extended sets of migration-related concepts were generated, comprising 250 ethical, 400 legal, and 450 social terms.

In the third step, following a human-in-the-loop framework, the taxonomy was manually refined through an iterative, co-creative process incorporating expert feedback and continuous revision. This step involved removing duplicate entries, excluding overly abstract or ambiguous terms, and assigning concepts that appeared across multiple domains to the most appropriate category.

In the fourth step, the Migration ELS Taxonomy was finalised by verifying its internal consistency and organising the entries hierarchically. Each category was structured using an “is-a” relationship framework, in which high-level concepts were divided into sub-concepts to ensure semantic coherence.

3.2.2. GenAI-Based Mapping of Texts to ELS Categories

To address RQ1, a GenAI-based classification approach was applied to assign the content of both datasets to one of the three ELS dimensions. For Dataset 1, the full text of each document was analysed without reduction. For Dataset 2, a stratified sampling strategy was employed to ensure geographical and thematic diversity across countries. From national Telegram channels in each country, a balanced random sample of 1000 messages was selected to capture variation in refugee discourse. Due to token limitations and API constraints, longer messages were truncated to fit within the model’s prompt capacity.

The model was then used to label each document and message as belonging to one of the three categories: ethical, legal, or social. To improve the accuracy and consistency of the model’s alignment with the ELS categories, the Migration ELS Taxonomy was incorporated into a fine-tuned prompt framework. This framework was refined over three phases through iterative adjustments to prompt both structure and content.

In the first prompting phase, a basic instruction was applied that asked the generative AI model to assign each text from both datasets to one of the predefined ELS categories or, where classification was uncertain, to assign the label Other. As generative models tend to default to concrete and easily recognisable topical classifications, the second phase introduced stricter constraints. In this phase, the prompt was modified to require the model to assign each input to one of the three predefined ELS categories, thereby eliminating the Other option.

In the final phase, the prompt was further refined using the Migration ELS Taxonomy as a conceptual foundation. Based on this taxonomy, detailed definitions and descriptors of the ethical, legal, and social domains were incorporated into the prompt, providing clearer semantic guidance for the model.

3.2.3. Validation of GenAI-Driven Classification

To assess the effectiveness of the generative AI model in mapping governmental and migrant narratives to the ELS categories, an expert-based validation procedure was conducted. This evaluation was performed after the final classification stage and focused on the outputs generated by the model, which had been refined through taxonomy-guided prompt engineering.

A balanced subset of 300 texts was randomly selected for validation, comprising 150 texts from the official policy dataset and 150 from the migrant-generated Telegram dataset. This sampling strategy ensured equal representation of source types, as well as diversity in topical domains and linguistic variation.

The selected subset was independently annotated by two human experts with domain expertise in migration-related legal, social, and ethical issues. Annotators were instructed to assign each text to one of four categories: Ethical, Legal, Social, or Other, relying primarily on the taxonomy definitions to enhance labelling consistency and minimise interpretative variation. The use of the Other category was discouraged. In cases of disagreement between the two annotators, a third expert adjudicated to determine the final label. This expert-annotated subset served as the gold standard for evaluating the model’s classification performance.

Model effectiveness was assessed using standard classification performance metrics, including accuracy, F1-score, and the confusion matrix. In addition, to evaluate the reliability of expert judgements and the consistency of the labelling process, inter-rater agreement between the two primary annotators was measured using Cohen’s kappa coefficient.

3.3. Research Ethics

Throughout the study, ethical considerations were addressed at each stage of the research process, particularly in relation to data handling, human involvement, and the interpretation of normative ELS components. These considerations were guided by the ELS framework and ethical design principles (Dignum et al. 2018; Teixeira et al. 2025; Qian et al. 2024).

During data collection, the primary ethical priority was to ensure that no personal, private, or sensitive information, particularly relating to a vulnerable population such as Ukrainian refugees, was included in the analysis in order to avoid exacerbating existing vulnerabilities (Lutz 2022). To address this, only publicly available data were used, including official documents and posts from open-access Telegram channels. All data were collected and processed in full compliance with data protection regulations, including the General Data Protection Regulation (GDPR). No identifiable user information was collected, stored, or analysed. The analysis focused strictly on public discourse rather than on individual users, thereby minimising the risk of harm, re-identification, or misrepresentation. This approach is consistent with best practices in ethical social media research (Williams et al. 2017; Townsend and Wallace 2016).

During the research design stage and the development of the Migration ELS taxonomy, ethical considerations focused on ensuring that the ethical, legal, and social dimensions were clearly defined to prevent bias and misinterpretation. To mitigate conceptual bias, the research team applied a transparent, iterative, and co-creative refinement process (Bowker and Star 2000). Ambiguous expressions, including politically charged language or vague references such as “imperial nations” or “social media information,” were excluded. Furthermore, overlapping concepts (e.g., “asylum seekers,” “human rights”) were carefully reassigned to the most contextually appropriate category. These refinements were guided by contextual analysis grounded in the migration literature rather than relying solely on automated tools (Arrieta et al. 2020).

During the narrative mapping stage using GenAI models, one of the primary ethical challenges was the risk of misclassification. Large language models often struggle with abstract categories and may generate biassed or reductive outputs when prompts are not carefully designed (Mittelstadt et al. 2016). To mitigate this risk, a human-in-the-loop approach was implemented, involving iterative prompt refinement and the integration of the taxonomy into the model’s classification process (Holzinger 2016). This approach reduced the likelihood of misrepresenting refugee voices or official documents. Care was also taken to prevent the model from reproducing harmful narratives (Leese et al. 2022). Only sanitised and relevant content was processed, and the use of broad categories such as Other was minimised through clearer conceptual guidance in the prompts.

To ensure ethical consistency during the expert evaluation stage, each expert was briefed on the purpose of the task, the sensitivity of the content, and the expectations for fair and consistent classification (Fort et al. 2011). Experts were instructed to use the Other category sparingly and only when the content clearly did not correspond to ethical, legal, or social concepts. To enhance transparency and reduce subjective bias, each set of texts was independently evaluated by two experts.

In the comparative and temporal analysis stage, a key ethical priority was to avoid attributing blame to any specific actor or policy. The analysis was descriptive and framed to highlight divergence or convergence in value expression, rather than to make normative judgements.

4. Results

This section presents the results of the study. Section 4.1, Section 4.2 and Section 4.3 report on the development and validation of the ELS taxonomy and the generative AI-based classification framework, while Section 4.4 applies this framework to analyse the alignment between governmental and Ukrainian migrant narratives.

4.1. Migration ELS Taxonomy Development

By implementing a four-step human-in-the-loop pipeline, a comprehensive Migration ELS Taxonomy comprising 450 elements was developed. The refinement process described in Section 3.2.1 progressively enhanced the precision, internal consistency, and normative clarity of the ethical, legal, and social domains.

Table 4. Phased refinement of the Migration ELS taxonomy using a human-in-the-loop pipeline.

Taxonomy Dimensions	First Step	Duplicates Deleting	Manual Adjustment	Final Number of Concepts/Top Level
Ethics	250	165 (85 deleted)	77 (88 deleted)	77/7
Legal	400	324 (76 deleted)	145 (179 deleted)	145/8
Social	450	350 (100 deleted)	228 (122 deleted)	228/4

presents the phased improvements achieved at each stage of the taxonomy’s development.

The hierarchically organised taxonomy comprises three normative dimensions: ethical (77 concepts, 17.1%), legal (145 concepts, 32.2%), and social (228 concepts, 50.7%).

Table 5. Top-level hierarchy of the Migration ELS taxonomy.

Ethics	Legal	Social
Migration status and intentions	International Law and Frameworks	Migration and Social Structures
Ethical Considerations	European Union Migration Framework	Migration and Economic Factors
Migration and Emotional Impacts	National Migration Policies and Laws	Migration and Social Change
Research Ethics	Migration Control and Enforcement	Migration and Individual Factors
Migration Policies	Migration Types and Categories
Human Rights and Social Justice	Migration Governance and Policy
Migration Research Methodologies	Regional and National Bodies
	Research, Advocacy, and Public Discourses

presents the top-level hierarchy of the finalised Migration ELS taxonomy, while Figure A1 in Appendix A provides the full taxonomy structure.

4.2. ELS Taxonomy-Guided GenAI Classification

In accordance with the described approach, during the initial prompting stage, a basic instruction was applied in which the GenAI model was asked to categorise texts from both datasets into one of the predefined ELS categories or, where classification was uncertain, to assign the label Other. This approach resulted in a considerable proportion of texts being classified as Other: 12% for official policy documents and 20% for Telegram messages. Figure 2 presents examples of text themes assigned by the model to the Other category during this stage. Examination of these outputs indicates the model’s tendency to default to concrete, easily identifiable thematic classifications when uncertainty is present.

In the second phase of the fine-tuning process, the prompt was modified to impose stricter constraints. The model was explicitly instructed to assign each input to one of the three predefined ELS categories, thereby strongly discouraging the use of the Other option. This adjustment led to a notable reduction in Other classifications, decreasing to 4.5% in official texts and 10.5% in social media texts.

In the final, taxonomy-informed phase, the prompt was further refined using the Migration ELS taxonomy as a conceptual foundation. By embedding detailed descriptions of the ethical, legal, and social categories into the prompt, the model achieved greater contextual awareness and improved its ability to classify content. As a result, the vast majority of texts were successfully assigned to one of the three ELS categories, with only a small fraction remaining in the Other category. In this phase, the proportion of Other responses decreased to 0.7% for official documents and 1.9% for Telegram messages.

Table 6. Three phases of GenAI-based text mapping (initial, constraint-based, taxonomy-informed).

Dataset	Three Phases (Initial/Constraint-Based/Taxonomy-Informed)
	Ethics (%)	Legal (%)	Social (%)
Dataset 1: Official documents	0/<1/0.3	16/25.5/32.7	72/70/66.3
Dataset 2: Telegram texts	1/<1/0.3	15/21.5/22.8	64/68/75

presents the three phases of GenAI-based ELS text mapping (initial, constraint-based, and taxonomy-informed). Across these phases, the model progressively reduces the proportion of texts assigned to the undefined Other category. At the same time, the model increases the proportion of texts mapped to the legal category (from 16% to 32.7%) and slightly decreases social classifications (from 72% to 66.3%) in Dataset 1. In Dataset 2, a similar pattern is observed, with increases in texts mapped to both the social category (from 64% to 75%) and the legal category (from 15% to 22.8%).

Appendix B presents selected excerpts from official governmental documents and Telegram-based migrant discourse corresponding to each ELS category as examples of the model’s classification.

4.3. Results of Expert-Based Validation of GenAI-Driven Mapping

To evaluate the generative AI model’s ability to identify ethical, legal, and social aspects embedded in textual content, an assessment was conducted using the results obtained after the final prompt refinement stage, in which the Migration ELS taxonomy was incorporated into the prompt.

An expert-annotated subset of 300 text instances served as the gold standard for evaluation. Model performance was assessed using precision, recall, and F1-score. In cases where the two expert annotations did not agree, disagreements were resolved by a third expert, whose judgement was used to assign the final label. Figure 3 compares classification performance across the two datasets, comprising official documents and texts from the Telegram social network.

The results indicate that discrepancies between expert annotations and model classifications most frequently occur when the model assigns a text to the legal category, while experts classify the same text as social.

Table 7. Examples of classification mismatches between expert annotations and model predictions across Dataset 1 (official documents) and Dataset 2 (Telegram messages).

Official Documents (Dataset 1)	Expert Label	Model Label
…The aim of the project is to provide support for the functioning of national asylum and migration management systems * by increasing the efficiency of activities in the area of voluntary returns. On the one hand, raising knowledge on the mechanism of voluntary return… *	social	legal
You have the right to work when you receive a temporary protection permit in accordance with the Temporary Protection Directive. Your children have the right to go to school, and you are entitled to health care as soon as you submit a defense application in accordance with the Temporary Protection Directive. The right to work. A person who is 16 years old is entitled to work after receiving a residence permit with protection in accordance with the Temporary Protection Directive…You or your employer must submit an application for a pre-employee income tax (A-Skatt) to the Skatteverket…	legal	legal
…Because we have monitored and understand how many our citizens need documentation according to passports, this is first and foremost what is very pleased. The fact that they simultaneously design our citizens, at the same time draw these cards. This means that they do not lose this sacred connection with Ukraine. We also expand the range of services we provide with the Ministry of Internal Affairs of Ukraine. It is an exchange of driver’s licenses…	legal	legal
Telegram Messages (Dataset 2) **	Expert Label	Model Label
Spain provides temporary protection to Ukrainians until 4 March 2024. What you will receive: The right to legal residence in Spain (until 4 March 2024). The right to work in Spain. The right to free state medical care in Spain Free housing is NOT PROVIDED!!… Language courses—are free for Ukrainians, you look for it yourself. !! That is, you need to rely only on yourself and your financial capabilities. !! What they write about programs is from organizations such as the Red Cross and others. They provide shelter (without animals) and food…	social	legal
How to get an electronic copy of your diploma If you forgot or were unable to take all the necessary documents and were forced to leave your homes, you can read the instructions on how to independently get an electronic version of your higher education diploma…	social	legal
Vacancy Free Responsibilities: sewing car seats and completing …Accommodation: • provided by the employer FREE OF CHARGE • … We guarantee: • Official employment in accordance with Czech legislation • Timely payment of wages (to a card or cash) • Weekly advances of 1000–1500 CZK • Payment of all state taxes for the employee • Free processing of all employee documents necessary for employment • Assistance and consultation in processing tolerance visas…	social	social

* Italicised text denotes fragments that influenced class definitions, as identified by experts. ** Telegram messages were translated from Ukrainian or Russian.

presents several examples of such mismatches across both the official document dataset and the social media dataset.

4.4. Comparative and Temporal Analysis of ELS Mapping Results

To address RQ2, a comparative analysis was conducted using the expert-validated mappings as a basis for comparing governmental publications and migrant-generated social media posts across the ELS dimensions. The analysis comprises three components.

First, an examination of the conceptual alignment between the two datasets was conducted by analysing the distribution of content across the three ELS categories in both official policy documents and migrant-generated Telegram messages. This analysis enabled the identification of dimensions that were emphasised or underrepresented in each dataset and allowed us to assess the extent to which governmental communication aligns with refugee discourse across the ethical, legal, and social dimensions. As shown in Table 6, social values are predominant in both official and migrant communications, though they are more pronounced in refugee discourse at 75 percent than in government texts at 66.3 percent. Legal content is more strongly represented in official documents at 32.7 percent, compared to migrant messages at 22.8 percent. Ethical concerns are almost absent in both datasets, at 0.3 percent. The proportion of content classified as Other, referring to texts that could not be assigned to any ELS dimension, was minimised through model fine-tuning.

To capture temporal dynamics, we analysed how discourse priorities evolved in both governmental documents and migrant-generated Telegram messages, assessing patterns of convergence or divergence over the study period. By segmenting each dataset into yearly intervals and reapplying the ELS classification to these temporal subsets, we traced shifts in governmental and migrant priorities across different phases of the Ukrainian migration crisis.

Table 8. Temporal distribution of ELS categories in official and migrant-generated texts.

	Dataset 1 (%)		Dataset 2 (%)
	2022	2023–February 2024	2022	2023–February 2024
Social	69.0	62.3	75.4	74.5
Legal	29.4	37.7	23.6	21.9
Ethics	0.5	<1	<1	<1
Other	1.2	<1	0.8	3.0

presents the distribution of classified content across two time periods, 2022 and 2023–2024, for each dataset.

The results indicate a shift in governmental documents, with the legal component increasing from 29.3% to 37.7%, while the social component decreased from 69.0% to 62.3%. This pattern suggests a growing formalisation of governmental response strategies as the refugee crisis evolved. In contrast, refugee discourse remained consistently centred on social themes, accounting for about 75% of the content, with a slight decline in legal references and a modest increase in Other content. The ethical category remained marginal in both datasets, indicating a continued underrepresentation of ethical considerations in governmental discourse and migrant-generated Telegram discourse.

Finally, cross-country variation in the distribution of the three ELS dimensions was examined, with attention to potential explanatory factors underlying differences observed across countries.

Table 9. Distribution of ELS categories in migrant Telegram messages by country (percentages).

	IT	MD	PL	SK	SP	SE	UK	IR	DE	CZ	CA	USA	Average
Social	77.4	77.9	76	71.6	75.7	69.5	77.7	83.2	72.3	86.5	70	62.8	75
Legal	20.7	20.8	23	27.4	23.6	30	20.9	15.8	25.8	12.5	28.5	24.5	22.8
Ethics	0.2	0.3	0.2	0.2	0.2	0	0.2	0.5	0.2	0.1	0.1	1.5	0.3
Other	1.7	1	1.1	0.8	0.5	0.5	1.2	0.5	1.7	0.9	1.4	11,.2	1.9

presents the distribution of Telegram messages by country and ELS category.

In migrant discourse observed on Telegram, social values dominate across all national contexts, indicating a widespread focus on daily life, inclusion, and integration. The proportion of the legal component varies by country, reflecting differences in host-country regulatory environments. Ethical issues remain marginal in all cases, while the share of content classified as Other is low.

Taken together, these analyses enabled an assessment of both the overall alignment between the two datasets within the ELS framework and the evolution of ethical, legal, and social priorities over time among authorities and marginalised social groups.

5. Discussion

This study evaluates the capacity of generative AI models to identify and classify ELS dimensions in textual data, using a comparison between governmental and Ukrainian migrant narratives as an empirical context to demonstrate the applicability of this approach. The findings align with previous research showing that LLMs often struggle to process abstract or complex topics without tailored prompts or domain-specific guidance (Mittelstadt et al. 2016; Weidinger et al. 2022). This limitation is reflected in the model’s strong tendency to assign a substantial proportion of inputs to the “Other” category during the initial mapping phase. However, through three iterative prompt fine-tuning phases, including the integration of the Migration ELS Taxonomy, the proportion of unclassified content was reduced by over 26% in official documents and 32% in migrant messages. This result is consistent with prior work emphasising the importance of human-in-the-loop design in AI systems tasked with interpreting ethically and socially grounded discourse (Fjeld et al. 2020).

Expert validation of the AI-assigned ELS labels indicates that the model achieved high accuracy for governmental texts (0.89) and moderate to high accuracy for social media data (0.77), demonstrating substantial alignment with expert judgements. Notably, similar levels of inconsistency were observed among human experts, as reflected in Cohen’s kappa values ranging from 0.65 to 0.72 across official documents and Telegram messages. Thus, while not perfect, the model, which was fine-tuned using expert-informed taxonomies, achieved a level of performance comparable to that of human annotators when handling ambiguous or abstract categories. This finding is consistent with well-documented challenges in interpreting normative language, particularly in cross-cultural and policy contexts (Ghallab 2019).

Building on the demonstrated capacity of generative AI to reliably map textual data onto ELS dimensions, this study extends the analysis to examine broader patterns in the data. The findings suggest that, although further validation is required, there is an emerging trend of increasing misalignment in narrative priorities between governmental documents and migrant-generated discourse. This misalignment appears in two main forms: first, thematic misalignment, referring to differences in ethical, legal, and social emphasis between institutional discourse and migrant-generated discourse observed on Telegram; and second, temporal misalignment, referring to divergent trends in the emphasis of these dimensions over time across these two forms of discourse.

Thematic misalignment is evident in the proportional emphasis of ELS components. Although both governmental discourse and Telegram-based migrant discourse are dominated by social themes, official communications consistently underrepresent social concerns compared to migrant-authored texts (66.3% vs. 75%). In governmental discourse, social dimensions primarily focused on institutional support, integration programmes, accommodation policies, and financial assistance. Although similar themes were also present in Telegram-based migrant discourse, migrant-generated messages more frequently emphasised everyday lived experiences, including housing and employment difficulties, language adaptation, access to healthcare and education, emotional stress, and community support networks. At the same time, the legal dimension is more prominent in policy documents (32.7%) than in Telegram-based migrant discourse (22.8%), reflecting the stronger regulatory focus of official institutions. This legal emphasis in official documents was primarily associated with temporary protection regulations, registration procedures, and residence requirements. Although similar legal themes were also present in Telegram-based migrant discourse, migrant-generated messages more frequently focused on practical legal procedures and everyday administrative challenges related to residence status, documentation requirements, temporary protection regulations, and access to rights and services within host countries.

Temporal misalignment is illustrated by two contrasting trends. Compared with 2022, governmental discourse during the 2023–February 2024 period shifted toward legal concerns, increasing by 8.3%, while references to social issues decreased by 6.7%. In contrast, migrant-generated messages remained relatively stable, with the social dimension declining only marginally from 75.4% to 74.5% and no comparable increase in the legal dimension. This pattern suggests that, while migrant concerns have remained relatively stable, focusing on housing, employment, integration, and everyday resilience, governmental discourse became increasingly oriented toward legal and administrative aspects of migration management. Such temporal divergence may indicate a gradual decline in the responsiveness of state communication and may provide indirect evidence of policy approaches that increasingly prioritise regulatory control over social and humanitarian considerations in migration governance.

Cross-national analysis reveals additional nuances. In some countries, such as Ireland and Sweden, the alignment between institutional discourse and migrant discourse observed on Telegram appears relatively strong, with both emphasising social and legal dimensions. In contrast, in countries such as the United States and the Czech Republic, the divergence is more pronounced, with migrant discourse focusing on community-building and everyday survival, whereas official communication emphasises security and immigration control. Notably, the ethical dimension is either absent or nearly absent across all countries, indicating a substantial gap in ethical reflection.

One of the most unexpected findings is the minimal presence of the ethical category in both datasets (approximately 0.3%). Within the Migration ELS Taxonomy, the ethical dimension refers to normative considerations related to moral responsibility, social justice, human rights, inclusion, vulnerability, and the ethical treatment of migrants and refugees. It includes issues such as fairness in refugee treatment, protection of vulnerable groups, discrimination, humanitarian responsibility, moral obligations toward displaced populations, socially responsible behaviour within host communities, prevention of exploitation and misinformation, and the protection of social cohesion during crisis conditions. However, despite the relevance of these concerns in migration-related contexts, the study demonstrates that their explicit representation in both governmental and migrant-generated discourse remains very limited. This limited representation of ethically relevant content may stem either from limitations of the model and the taxonomy or from a genuine lack of explicit ethical discourse in the source material. Furthermore, it may be assumed that, unlike legal or social dimensions, ethical considerations are often embedded indirectly within broader narratives and expressed implicitly through discussions of legal procedures and social difficulties, which may make them more difficult to identify consistently through the model and taxonomy. However, given the similarly low proportion of ethics-related labels assigned by human experts (approximately 1.5% to 3%), the results allow us to conclude with a reasonable degree of confidence that both governmental and migrant discourses devote limited attention to explicit ethical considerations. This marginal presence of ethical reflection may indicate a broader institutional and societal tendency to prioritise legal and social-administrative concerns over moral and humanitarian considerations, an imbalance widely discussed in the field of public administration ethics (Mittelstadt et al. 2016).

Limitations

Despite the strengths of the proposed approach, several limitations should be acknowledged.

First, the definition and operationalisation of abstract concepts such as ethical, legal, and social dimensions remain inherently challenging. These categories are overlapping and difficult to delineate semantically, which affects both model performance and expert agreement. This complexity is reflected in the moderate classification performance and Cohen’s kappa scores observed in the study. Although ethical, legal, and social dimensions are inherently overlapping normative concepts, the study attempted to reduce category vagueness through the application of a four-step human-in-the-loop development pipeline for the Migration ELS Taxonomy, incorporating GenAI-assisted conceptual mapping, literature-based contextualisation, iterative expert refinement, and hierarchical semantic structuring. However, despite these methodological procedures, some degree of ambiguity between normative categories remains unavoidable in complex migration-related discourse.

Second, the representativeness and accessibility of social network data pose important limitations. Although prior research supports the relevance of Telegram data for Ukrainian migrants, social media platforms tend to be used more frequently by younger and more affluent populations (Blank 2017). In addition, access to aggregated data is often restricted by platform providers (Lutz 2022), which may introduce biases when generalising findings to broader populations or other marginalised groups.

Third, linguistic differences between datasets (Ukrainian/Russian versus English) may introduce semantic inconsistencies. The translation of normatively sensitive expressions may alter meaning and affect the accuracy of ELS classification. Moreover, reliance on generative AI may introduce biases in the interpretation of nuanced expressions, particularly ethical ones. In this study, differences between Ukrainian, Russian, and English texts may have limited the model’s ability to capture implicit ethical meanings, which could result in the underrepresentation of ethical dimensions compared to more clearly expressed legal and social categories.

Fourth, while the Migration ELS Taxonomy provides a useful structure for classification, certain concepts inherently span multiple normative domains. Although the human-in-the-loop approach helps resolve such ambiguities, further refinement of the taxonomy, including more granular categories, could enhance conceptual clarity and classification performance.

Finally, another limitation of the study is the reliance on a single generative AI model (GPT-3.5) for the classification and mapping of ELS dimensions. The objective of the study was not to compare the performance of different generative AI systems, but rather to investigate the capability of a GenAI-assisted framework for mapping and analysing ethical, legal, and social dimensions in migration-related discourse. Future studies may expand this approach through comparative evaluations involving multiple generative AI models and architectures.

6. Conclusions

This study makes the following contributions. First, it develops a novel Migration ELS Taxonomy for structuring ethical, legal, and social dimensions in migration-related discourse. Second, it evaluates the capacity of generative AI models to identify and classify these dimensions in textual data using a human-in-the-loop approach. Third, it applies this framework to analyse the alignment between governmental and Ukrainian migrant narratives, providing empirical insights into thematic and temporal divergences in ELS dimensions.

Ultimately, this study introduces an initial and scalable framework that combines a domain-specific ELS taxonomy with generative AI and human-in-the-loop validation to systematically assess the alignment between governmental narratives and the perspectives of marginalised populations under crisis conditions.