In recent years, generative artificial intelligence systems have become central mediators of how people access and understand information about contentious social issues, including migration and forced displacement. Integrated into dominant search infrastructures, these systems do not merely retrieve existing content; they actively synthesise and generate new textual and visual representations (Bommasani et al., 2021) that can shape collective imaginaries about vulnerable groups.

This article addresses a specific research problem: how Google’s generative AI systems (AI Overviews and Gemini) construct and circulate imaginaries about refugees in the Spanish-speaking context. The study is guided by the following research question: How do Google AI Overviews and Gemini generate textual and visual representations of refugees, and what imaginaries of refuge and displacement emerge from these outputs? Building on literature on algorithmic bias, socio-technical imaginaries, and refugee representation (Noble, 2018; Makhortykh et al., 2021; Wijnhoven and van Haren, 2021; Seo and Kavakli, 2022), we propose two main hypotheses: (H1) that Google’s generative AI systems tend to privilege institutional and highly structured sources, leading to a narrow representation of refugees; and (H2) that the visual and textual outputs reproduce pre-existing stereotypical iconographies that emphasise vulnerability and institutional dependence while obscuring refugee agency and heterogeneity.

The term “refugee” (refugee in English, réfugié in French), although previously in use, was made official in 1951 with the adoption of the Convention Relating to the Status of Refugees, approved by the United Nations on 28 July 1951 and entered into force on 22 April 1954. The 1951 Convention defines a refugee as: “Any person who, owing to a well-founded fear of being persecuted for reasons of race, religion, nationality, membership of a particular social group or political opinion, is outside the country of his nationality and is unable or, owing to such fear, is unwilling to avail himself of the protection of that country.” (United Nations, 1951). Since then, although previous studies exist, the figure of the refugee has taken centre stage in studies on mobility, human rights and migration, as it is an element of intersection between humanitarian protection and national security (Loescher, 2021). Thus, the figure of the refugee is both a legal and narrative element that produces effects of inclusion and exclusion within states and their migration governance policies (Owen, 2024). We must therefore understand the figure of the refugee in a holistic, way not just a legal one, as it places the person in a symbolic and political position on the margins of citizenship and recognition (Chimni, 2009).

The study of and interest in refugees has grown as migration has become one of the central elements of the political narrative in the Western world. In this context, the choice of Spain and Spanish as the geographical and linguistic framework for this study responds to its specific relevance within contemporary migratory flows and the European public debate on refuge and asylum, making it a case study of great international relevance. Spain is one of the main countries of entry into the European Union via the western Mediterranean and Atlantic routes (International Organization for Migration, 2025), which has placed the issue of refugees at the centre of political, media and social discourse. The scientific approach has evolved from a humanitarian perspective to multidisciplinary research known as “forced migration studies,” which analyses governance structures, underlying problems, and regulatory or discursive frameworks that shape displacement policies (Fiddian-Qasmiyeh et al., 2014). In this process, refugees have gone from being passive subjects of study viewed from a paternalistic perspective to key actors in processes of identity, community and political reconstruction (Brankamp and Weima, 2021). Today, the figure of the refugee has become almost an element of identity or a symbol of globalisation and contemporary inequality, where their media visibility shapes imaginaries of compassion, threat and otherness (Novais and Calderón, 2024). These representations are not neutral and have a two-way impact on social perceptions and political responses to migration (Sutkutė, 2023).

Thus, between the legal and political dimensions of refuge and its media representation, we observe that the imagined social construction, narratives and discourses comprise and value the figure of refugees. This indicates that the construction of the figure of the refugee is not only produced by institutions through regulations and legislation, but also within the symbolic realm of the mainstream media, social networks and, more recently, digital platforms and AI. Within this evolving landscape of algorithmic systems, this research specifically focuses on Google’s generative AI tools, namely Google AI Overviews and Google Gemini, as a significant case study. This selection is justified for several reasons: First, Google AI Overviews represent a transformative shift in search interfaces (Garanko, 2025), moving from indexing existing content to synthesizing and generating novel summaries that present themselves as authoritative information sources. Second, the integration of generative capabilities into Google’s search ecosystem—the most widely used search platform globally, commanding 89.94% of the global search market share (StatCounter Global Stats, 2025) and 93.53% within Spanish-speaking contexts—makes this analysis particularly relevant for understanding how algorithmic authority shapes public knowledge construction. Third, analysing both AI Overviews and Gemini allows us to see how the same Google system generates textual and visual representations of refugees, providing a more complete picture of its algorithmic construction of meaning. Finally, the scale and ubiquity of Google’s reach are particularly pronounced in Spanish-language search markets, where Google accounts for more than 94% of all search queries in Spain and similarly dominant shares across Spanish-speaking countries (Clayton, 2025).

This set of elements shapes the social imaginaries that legitimise certain perceptions and institutional responses (Crawley and Skleparis, 2018; Chouliaraki and Georgiou, 2016). As a whole, the worldview of refugees can be encompassed within a duality that oscillates between the humanitarian framework, which appeals to empathy and protection (albeit with clear doses of paternalism), and the security framework, which emphasises control and threat (Esses et al., 2013; Bleiker et al., 2013). At the current frontier of AI as a tool for creating discourse for both civil society and decision makers (Medina Grajales et al., 2024), it is necessary to understand how AI is contributing to the creation of these imaginaries, whether it is neutral, reproducing previous narratives or establishing its own pattern in the worldview of refugees.

The image is not a visual element, a simple representation of reality that is more or less figurative; it is a symbolic element that mediates in the social perception. This perception is rarely neutral and is related to the projection of power games derived from the delimitation of collective identities (Chouliaraki and Georgiou, 2016). Collective imaginaries are communal elements of self-definition that must be shared by the members of a group in order to give meaning to their world (Bottici, 2014). Bottici (2014) introduces three phases of imaginative structures: radical imagination, social imagination and, in between, with the capacity to generate political power, we find marginal imagination. This imaginary is inherent to society itself, its articulation and institutionalisation (Castoriadis cited in Bottici, 2014, p. 52). Thus, imaginaries construct or cement collective identities, which are the elements that constitute the core of the formation of meaning for a group of people (Castells, 2003). Portasany (1995) refers collective identities as those in which the subject of analysis is not the individual, but a group. In this sense, collective identities tend to be created in opposition to the other in a common universe of rules and norms which can be transformed and modified (Portasany, 1995). In this sense, refugees and the associated discourse are recurring elements in the creation of identities in opposition to the “other” through the cementing of collective imaginaries.

These imaginaries associated with migration and refugees have historically been sustained by institutional discourses or narratives at the state or supranational level or by the media, based on definitions in international law (Cox et al., 2019). They are largely dual simplifications between paternalism and violence, between victim and threat, producing a high degree of symbolic polarisation (Smets and Bozdağ, 2018, p. 294). This duality necessarily occurs through stereotypes or prejudices. In this sense, Lippmann (2003) defines stereotypes as mental images or reductionist patterns of reality that humanity uses to have a comprehensible idea of the world, even if these are fictitious, irrational and manipulated both by power and by our pre-established cultural and moral values (Lippmann, 2003). Similarly, Noya (2007), from the perspective of geopolitics and governance of countries, defines them, citing Furnham and Schofield (1986), as “oversimplified, rigid and generalised beliefs about groups of people according to which all individuals in the group are considered to possess the same set of dominant characteristics” (Noya, 2007, p. 31). Thus, although stereotypes are a simplification of reality, in the case of emigration, this trivialisation is often laden with negative connotations, which can be defined as prejudice.

In this context, research on migrants and refugees is based on a filter of imaginaries from which science is not exempt. In this sense, although research on the subject has increased in recent years, it focuses on the European or North-American context (Seo and Kavakli, 2022), which hinders a general theoretical approach. With regard to the media, during the 2015 migration crisis, the media used frameworks of “uncertainty” to represent migration, which directly affected social perceptions of insecurity as Boomgaarden and Gottlob (2020) show in the French and Australian media. Since this serious crisis, far-right populist movements have found in refugees an element of confrontation and polarisation that is electorally profitable (Brubaker, 2017; Caiani and Graziano, 2019), with a high degree of manipulation and normalisation of criminalisation (Loner et al., 2025). In the case of Spain, this dynamic has followed the same pattern as in other European countries (López Sala et al., 2021; Olmos-Alcaraz, 2022). As a means of eliminating this type of bias, Nayar (2022) argues that the “Refugee Imaginary” should take into account both social representations of refugees (by the host society) and the self-representations of those who migrate. In any case, the iconographic analysis of images of migrants allows us to interpret an overview of society and the context of refugees (Clark, 2020).

Until a few years ago, this iconographic analysis was based on the analysis of the final result of the image, taking into account the positions of the producer of the image (Bleiker et al., 2013). However, the emergence of AI, due to the volume of creation and the “democratisation” of image production, makes it necessary to understand not the final image but the underlying logic and internal mechanisms. Unlike documentary photography or traditional art, AI image generation is based on patterns extracted from huge data sets, which can perpetuate visual stereotypes. It has been shown that image generation systems tend to reproduce biases present in their training data, creating representations that can be both inaccurate and potentially harmful (Sun et al., 2023; Huang and Huang, 2025).

In this context, AI has an impact on the creation of imaginaries that “guide” specific visions of specific elements. According to Jasanoff and Kim (2009), we can say that these imaginaries are part of what can be called “socio-technical imaginaries,” which have the capacity to articulate shared visions of the social order or the technological future. These can be related to the “algorithmic imaginary” described by Bucher (2016) based on Facebook’s algorithms insofar as they explain how people imagine, perceive, and negotiate the power of algorithms in their daily lives. Algorithms can be understood as cultural practices insofar as they shape, or influence the shaping of, regimes of authority and imaginaries, and can establish meanings for a specific perspective (Seaver, 2017), such as migratory dynamics. According to Cave and Dihal (2019), AI algorithms shape collective imaginations ranging from dystopia to utopia, structuring hopes and fears.

The perspective provided by Cave and Dihal (2020) is particularly significant in that it analyses raciality in these technological components, where there is a homogeneity in robots, assistants and images of markedly white people, which implies the recognition of this group as legitimate technological users. Generative AI potentially has this underlying problem insofar as it accumulates web image and language models, giving rise to requests and analyses for the establishment of limits and documentation on processes (Bender et al., 2021). With regard to images, this type of racial bias has been detected in multimodal datasets (Birhane et al., 2021). Therefore, the literature on algorithmic equity is particularly important for synthesising sources of bias (Mehrabi et al., 2021).

Thus, various studies have shown that generative artificial intelligence systems are not neutral technologies. Especially in large language models, it has been observed that the increase in scale and the opacity of datasets can generate systematic distortions, reinforce dominant narratives, and marginalise alternative perspectives (Bender et al., 2021). On the other hand, research focusing on multimodal datasets has identified the persistence of racialised, gendered and stereotypical representations in generative systems (Birhane et al., 2021; Mehrabi et al., 2021). Finally, recent analyses of image generation models show that these systems systematically reproduce representational biases, especially in relation to gender, race, and vulnerability. This is a consequence of imbalances in the training data and the evaluation criteria used (Sun et al., 2023).

Finally, this context becomes more complex with the introduction of Google AI Overviews into generative processes. AI Overviews represent a significant evolution in how information is presented online, transforming search engines that only indexed content into synthesisers and generators of new content, and assuming the role of “authority” on the subject. They provide users with a version that has already been interpreted and validated by an algorithm, rather than leaving them the task of comparing and verifying information. These systems use language models to generate coherent responses based on multiple sources, but their selection and synthesis process is not transparent. Research on AI Overviews has revealed that these systems tend to prioritise consensus on a topic over accuracy, especially when faced with contradictory or irrelevant information. This is particularly problematic for topics related to refugees, where information may be scarce, controversial, or politicised. Studies have shown that AI Overviews may ignore accurate and up-to-date content if it contradicts a broader but outdated consensus (Yuan et al., 2025; Duan and Wang, 2024; Ye et al., 2023).

This research employs a primarily qualitative design combining content analysis of AI Overviews’ textual summaries with iconographic analysis of Gemini-generated images, supported by descriptive counts of source selection patterns. This methodological approach falls within the tradition of algorithmic auditing (Raji et al., 2020; Metaxa et al., 2021), a technique that allows for the systematic examination of how algorithms process and represent information about specific groups. Algorithmic audits are a research method that allows the performance of algorithms to be evaluated in relation to specific criteria, identifying possible biases, errors or discrimination that may be present and that could have consequences in the construction of collective imaginaries (Lovelace, 2020; Raji et al., 2020).

Imaginaries about refugees are constructed through multiple mediations, including visual representations, textual narratives, and interpretive frameworks. These imaginaries constitute an emerging interdisciplinary field that examines how refugees imagine the world and how the world imagines them. Research has shown that dominant representations tend to oscillate between two main discourses: the threat discourse, which presents immigration as a challenge to security, and the humanitarian discourse, which emphasises the adverse conditions and suffering of migrants. These reductionist binary representations have far-reaching effects on the formation of public policy and social attitudes.

This study examines how Google’s generative AI systems—specifically AI Overviews and Gemini—construct and generate representations of refugees through synthesized content generation in AI Overviews and image generation in Gemini. The results reveal that these generative systems tend to select and synthesize sources such as Wikipedia for specific queries, while UNHCR dominates in general queries. Significant gaps in AI content selection are identified for some high-volume searches on refugees, for example, on environmental refugees and specific historical events, highlighting how the scarcity of quality content directly influences AI’s ability to generate coherent and non-stereotypical textual and visual representations.

This research emphasises the idea that digital inclusion and participation are key vectors for developing equitable communication environments. This is especially significant when the image-creating element is a collective in contexts of disability or vulnerability. If communication technologies must be designed with accessibility criteria in mind in order to encourage the participation of certain groups in the digital sphere (Keeley and Bernasconi, 2023; Wahl and Weiland, 2023), AI must ensure the development of meaningful and unbiased texts and images to promote the inclusion of migrant or refugee groups in the use of generative technologies. This element opens the door to a series of new research projects that evaluate the digital divide in the use of generative AIs among different social groups and how this affects first the collective imagination, then the communication ecosystems, finally and political decisions.

In this sense, we can find synergies in this geopolitical and communicative dimension of media governance, which is closely linked to the ability of governments to play a dual role (Dragomir, 2025). In this regard, it is necessary to develop new research that evaluates how the centralisation of power in generative AIs is affecting, or may affect in the future, the creation of discourses on refugees, what role states are playing in the regulation or deregulation of AIs, and how this affects the dynamics of power, informational control and symbolic exclusion. This fact may have enormous repercussions for the very foundation of Western democracy since, as Cano-Orón et al. (2024) points out, the public sphere is currently fragmented, polarised and subject to emotional manipulation. The promotion of biased images in systems with the potential to replace global communication processes may lead to a radical change in the immediate future of global governance. This process must be investigated in greater depth and from a more diverse perspective on the basis of this work, which analyses the creation of the image of refugees, one of the elements around which current political discourse in the West revolves. In this regard, we believe that this work provides a solid foundation for further comparative research on the imagery surrounding refugees in artificial intelligence platforms and traditional media.