Differential Representation as Epistemic Inequality: How Generative AI Systems Frame Overdevelopment and Ecological Harm in Malta

1. Introduction

Generative artificial intelligence (Gen AI) systems are becoming increasingly important in practices of knowledge acquisition, production, and interpretation. This transformation is also evident in how public knowledge about ecological issues and ecological harm is represented, produced, and circulated.

However, this process does not produce neutral information circulation. Gen AI systems are shaped by specific training data, institutional priorities, and knowledge geographies. This may make some knowledge repertoires more visible while rendering others less visible. Therefore, the question of how Gen AI systems represent ecological issues concerns not only technical accuracy, but also epistemic representation [1,2].

Existing research largely evaluates Gen AI along the axes of accuracy, hallucination, and benchmark performance [3], or presents Gen AI systems as solutions for environmental sustainability [4]. This study offers a different perspective by examining how ecological issues are represented in Gen AI systems and how these representations vary across models.

To examine these issues, the study focuses on overdevelopment and associated ecological harm in Malta. As a small island state experiencing intense building pressures and long-standing tensions over environmental governance, Malta provides an analytically powerful case for examining how Gen AI systems represent local political-ecological contexts that receive limited visibility in global information flows. A central focus of this study is the visibility of local ecological conflicts, civil society discourses, and institutional actors in Gen AI outputs.

Accordingly, this study addresses three interconnected research questions: how different Gen AI systems frame overdevelopment and ecological harm in Malta; which actors, forms of knowledge, and solution imaginaries these framings make visible or invisible; and what patterns of epistemic inequality emerge from differences in representation across models. The study extends discussions of epistemic inequality into comparative Gen AI research by shifting the focus of model comparison from accuracy assessment to representation analysis. It also contributes to the literature on environmental representation and environmental communication by focusing on how ecological issues become visible, explainable, and politically readable within Gen AI systems. The study examines how prompt conditions and model-specific representational tendencies co-shape the visibility of ecological issues, positioning the small island state context as a distinctive political-ecological case study for Gen AI representation research. The findings show that representations vary across models and prompt conditions. Civil society actors such as environmental NGOs were relatively visible, but their visibility was prompt-conditional, whereas scientists/experts and EU institutions remained persistently less visible. This supports the interpretation of Gen AI systems as components of epistemic mediation rather than neutral channels of ecological information.

2. Literature Review

2.1. Visibility, Depoliticisation and Scale in Ecological Representation

Ecological crises and ecological harm have long been discussed in academic circles, ecological movements, and civil society organisations. However, how ecological issues are understood socially is shaped as much by how they are represented as by the ecological processes themselves. People often encounter ecological problems not directly, but through media content, political discourses, expert reports, public institutions, and civil society actors [5,6]. Therefore, ecological crises become visible in the public and political spheres not directly but through processes of representation [7].

Ecological representation concerns how ecological issues become visible, how they are explained, and the frameworks of meaning through which they are interpreted [8]. In representation processes, some actors, causes, and proposed solutions gain visibility, while others remain in the background [7,9]. Therefore, representation is not merely a reflection of existing realities, but a process of socially constructing meanings around these realities: which actors, explanations, and solution horizons become publicly meaningful is shaped within this process [10]. In this sense, the public visibility of ecological issues is not considered a neutral process but a practice of representation shaped by political and economic power relations [8].

The question of which actors, causes, and proposed solutions gain visibility in representation processes also raises how ecological issues are politicised or depoliticised. The depoliticisation approach draws attention to the reframing of publicly contested issues as management problems requiring technical solutions by pushing political contestation and power relations into the background. Ecological harm is also frequently represented through technical indicators, risk calculations, governance tools, or expert knowledge. While such representations can increase the visibility of ecological issues, they can also push into the background the economic interests, institutional relations, and political responsibilities that contribute to their emergence. Thus, ecological issues can be framed as problems to be solved through expertise rather than as issues of political conflict and accountability. This discussion is also important for sustainability discourses: critical approaches argue that, in some contexts, sustainability discourses can reframe ecological crises as governance and optimisation issues rather than social struggles and questions of environmental justice [11,12,13].

One noteworthy dimension of ecological representation is the politics of scale. Ecological problems can be represented at local, national, regional, or global scales. However, scale is not merely a geographical category, but a form of organisation produced within social and political processes [14,15]. The scale at which representation is produced can influence which actors, responsibilities, and proposed solutions gain visibility. For example, when ecological crises are represented at a global scale, the visibility of specific local ecosystems, environmental injustices, and regional governance conflicts may decrease. In contrast, local-scale representations can make visible how global processes are experienced in different geographies and how they produce unequal outcomes for different communities [16].

Questions of ecological representation, depoliticisation, and the politics of scale have long been addressed in the literature on media, political communication, and environmental governance [7,8]. This literature demonstrates that questions of which actors are made visible, which causal explanations are legitimised, and which solution imaginaries are highlighted are central to representational processes. However, the increasing visibility of Gen AI systems in processes of accessing and interpreting ecological information necessitates rethinking these representational issues in a new context [17]. Nevertheless, the current Gen AI literature largely focuses on accuracy, safety, and performance while paying limited attention to questions of how ecological issues are represented, which actors are made visible, and which causal frameworks are produced [4,17]. In contexts that already have limited visibility in the global flow of information, such as small island states, the question of how this scale asymmetry is represented in Gen AI systems, which actors are made visible, and which explanations are highlighted, becomes particularly critical.

2.2. Gen AI Systems as Algorithmic Epistemic Infrastructures

Gen AI is often discussed in the literature as a tool, an assistant, a decision-support system, and a technology to aid productivity. These approaches have generated a substantial body of literature at the market and policy levels. Although the mainstream literature acknowledges some empirical dilemmas of Gen AI systems, such as the appropriation of the outputs of highly skilled labour or the locking-in of historical data, it essentially positions these systems as neutral assistants that optimise employee productivity [18].

Classical instrumental approaches conceptualise technologies as tools that users employ to achieve specific goals. Within this framework, epistemic mediation is conceived as a function of the tools used in the process through which a subject recognises and understands an object. The production and interpretation of information are seen as largely dependent on the user's actions [19,20,21]. However, when it comes to Gen AI systems, this framework is inadequate.

Gen AI systems operate within algorithmic information environments in which information is selected, sorted, synthesised, and presented. Therefore, the issue extends beyond how users employ a tool to how information becomes visible. Couldry and Hepp [22] examine how reality is constructed through mediated processes. Although their account does not concern Gen AI, it offers a broader perspective on how information becomes organised and made visible. As science and technology studies and critical research on language technologies show, technical systems do not merely transmit information; they present it within particular processes of classification, organisation, and meaning [23,24]. Epistemic agency is therefore not simply a matter of access to information, but of what information is made visible, what explanations are legitimised, and what interpretive frameworks are foregrounded. Current studies also emphasise that artificial intelligence has become an epistemic infrastructure that structures knowledge through linguistic fluency and computational processes rather than through factual verification [25]. In this study, epistemic mediation refers to the process through which Gen AI systems rearrange and present information in line with institutional priorities, rather than to a function of instrumental technology.

As Gillespie [2] demonstrates, the functioning of algorithmic systems is shaped by commercial priorities, legal obligations, and institutional preferences; therefore, it is not possible to consider these systems as neutral information channels. Decisions about what is made visible, which information is prioritised, and which frameworks are considered legitimate are inherently political and cultural. Training data is also a product of specific social, institutional, and historical choices. Therefore, Gen AI systems produce syntheses that have been selected and filtered through specific institutional and informational channels, rather than presenting raw data to the user. Representation, in this sense, is not something transmitted, but a process organised through decisions on selection, visibility, and ranking.

The different framings of the same socio-ecological issue across systems indicate that these systems function not as neutral channels for transmitting information, but rather as part of epistemic mediation processes. Structural differences in institutional, geographical, and organisational architectures, commercial priorities, legal obligations, and alignment preferences constitute the conditions under which systems produce different and competing forms of representation in their approaches to the same socio-ecological issue. These conditions also suggest that different contexts and knowledge repertoires do not attain equal visibility across systems. This is particularly important for contexts that already have limited visibility within global knowledge infrastructures. Representational visibility is shaped not by the social and political salience of an issue alone, but by the extent to which knowledge about that issue is incorporated into training data, institutional knowledge sources, and the broader flow of information.

Therefore, the question of how specific ecological issues are represented is not only about technical capacity; it is also about how representation processes are organised. That different systems make different actors visible, offer different causal explanations, and produce different solution repertoires in response to the same ecological issue raises questions about the epistemic consequences of representation processes. This situation necessitates a re-examination of epistemic inequality in the context of the representational practices of Gen AI systems.

This study uses two related but analytically distinct terms. Gen AI systems is used where the focus is on representational practices, epistemic positioning, and comparative argument. It appears that where LLMs (large language models) is involved is where the focus shifts to technical architecture, methodological procedure, or the empirical properties of model outputs. The distinction reflects the level of analysis at which a given claim operates rather than a difference in the systems themselves.

2.3. Epistemic Inequality and Differential Representation

The literature on epistemic inequality largely focuses on inequalities in knowledge production, access to information, and control over information [1]. While Fricker's [1] work addresses epistemic inequality through injustices in the processes of individuals producing, sharing, and being recognised as knowledgeable, subsequent studies have drawn attention to how digital infrastructures and information systems can reproduce these inequalities [26,27,28]. Within these discussions, Zuboff [29,30] argues that information asymmetries created by digital systems generate new power relations. According to Zuboff, epistemic inequality arises when the information-gathering, processing, and prediction capacities of digital platforms exceed the information and control capacities of users. Zuboff's conceptualisation addresses the information asymmetries produced by digital systems through the platform-user relationship.

This study expands on epistemic inequality along a different axis: the problem is not the amount of information one can access but the way the same issue is represented across different systems. These representational choices also carry epistemic consequences [2]. Epistemic inequality is therefore a matter less of access to information than of which representations become visible. In this study, differential representation refers to systematic variations in the visibility of actors, causal explanations, and solution imaginaries across Gen AI systems when representing the same socio-ecological issue. Epistemic inequality is conceptualised through this notion of differential representation. Representational processes can make some actors visible while relegating others to the background; they can highlight some causal frameworks while marginalising others. Differences in representation can therefore produce more than discursive diversity: they can generate unequal epistemic visibility conditions. In this sense, epistemic inequality turns on which forms of knowledge become visible, which explanations stand out, and which actors are representable, rather than on access to information alone.

The fact that different Gen AI systems represent the same ecological issue through different patterns of actor visibility, causal explanations, and solution imaginaries indicates that these systems do more than produce information: they generate distinct epistemic visibility conditions. When these visibility conditions differ structurally between systems, the interpretive resources and representational horizons available for the same issue can also differ. Therefore, differences in representation can give rise to patterns of epistemic inequality in which certain actors, explanations, and solution imaginaries become more visible than others. This study examines this relationship through the case of overdevelopment and ecological harm in Malta. As a political-ecological context with limited visibility in the global flow of information, Malta offers a powerful case for analytically examining how these epistemic visibility conditions are shaped across different Gen AI systems.

3. Methodology

3.1. Research Design

This study is designed as a comparative qualitative interpretive analysis examining how Gen AI systems represent overdevelopment and its associated ecological harm in Malta. The basic assumption is that LLM outputs are not merely technical responses that transmit information; they are representational texts that generate specific epistemic frameworks, causal narratives, patterns of actor visibility, and policy imaginaries. Therefore, the focus of the study is not on accuracy verification or benchmark performance comparison, but on the comparative examination of different model representations of the same issue.

Malta was selected as a strategically information-rich case for this study. Its status as a small island state, characterised by intense building pressures, limited land availability, and longstanding debates over environmental governance and overdevelopment, makes Malta an analytically powerful case for examining ecological representation, governance framing, and scale positioning [31,32]. Furthermore, Malta offers a significant case study for examining epistemic inequality in representation, given its small-scale political-ecological context and the uncertainty surrounding the extent to which local governance debates are visible within the knowledge architectures of LLMs.

3.2. Model Selection and Data Collection Controls

In this study, four general-purpose LLM-based systems were examined comparatively: ChatGPT (OpenAI, GPT-5.5 Instant), Claude (Anthropic, Sonnet 4.6), Gemini (Google), and DeepSeek. These systems were selected because they were developed by different providers and accessed through comparable public-facing conversational interfaces. Locally deployed open-weight models and configurations with researcher-enabled live retrieval were excluded, since the study aimed to examine representation patterns generated without retrieval being deliberately activated. Hidden platform-side retrieval or routing, however, could not be independently excluded.

For ChatGPT and Claude, the model names displayed in the platform interfaces were recorded. For Gemini and DeepSeek, no exact model version was displayed, so these systems are identified by platform name. All four systems were accessed through free, publicly available interfaces without subscription-based accounts.

To improve consistency across comparisons, each prompt iteration was executed in a fresh, isolated session; with incognito mode active, web search disabled, and previous conversation history cleared. All systems were queried with the same standardised instruction not to use live retrieval. However, platform-level hidden optimisation or personalisation effects could not be fully excluded.

3.3. Prompt Design and Data Collection Protocol

Data were collected using a standardised prompt protocol consisting of four prompt families: (A) descriptive ecological framing (e.g., “What are the main ecological consequences of overdevelopment in Malta?”), (B) causal explanations and the persistence of overdevelopment (e.g., “Why has overdevelopment become a persistent ecological problem in Malta?”), (C) actor visibility and interest relations (e.g., “Who are the main actors involved in overdevelopment and ecological governance in Malta?”), and (D) policy and governance solution imaginaries (e.g., “What measures have been proposed or implemented to address the ecological impacts of overdevelopment in Malta?”). Each family included two complementary prompts addressing the same analytical dimension from different perspectives, resulting in a total of eight standardised prompts. The complete prompt set, final codebook, and full dataset of 96 raw model outputs are publicly available through an anonymised Open Science Framework (OSF) view-only link (https://osf.io/a4v9d/overview?view_only=c371762dd3ca4bb89b8c0c668f635ea9).

The same prompts were used across all four systems. No follow-up questions, requests for clarification, or corrective interactions were introduced. Each prompt was run independently three times to examine possible variation in the outputs. All outputs were retained rather than selecting a single response as representative. Both recurring patterns and differences across repetitions were considered in the analysis.

The final dataset consists of 96 outputs generated by four systems, eight prompts, and three repetitions. All 96 outputs were collected on 22 May 2026, with all four systems queried on the same date under the same researcher-controlled procedural conditions.

3.4. Coding Framework and Analytical Procedure

A hybrid deductive-inductive coding strategy was used. The final corpus comprised 1,919 coded meaning units (A = 464, B = 499, C = 430, D = 526). Higher-level code categories were defined deductively based on the research questions and the analytical framework, while inductive subcategories were developed through an iterative close reading process [33]. The coding framework was developed by drawing on conceptual discussions in the environmental discourse and political ecology literature [34] and critical literature on the epistemic positioning of LLMs, knowledge production, and the limits of representation [17], and was refined through the pilot coding process. All coding was conducted manually by the researcher; no dedicated qualitative data analysis software or computer code was used.

The basic unit of analysis was the meaning unit. The meaning unit was defined as a sentence, sentence fragment, or short meaningful expression containing a single analytic claim, causal proposition, actor attribution, or representational expression [35]. Meaning units were segmented according to analytical content rather than sentence length. Where a sentence contained more than one distinct analytical claim, it was divided into separate meaning units. Short consecutive expressions conveying the same analytical idea were retained as a single meaning unit. Introductory or transitional phrases, incomplete fragments, and standalone reference entries were not coded as meaning units when they did not contain an analytical claim. The original outputs were retained unchanged.

The coding framework consists of eight main categories: Ecological Outcomes, Causal Factors, Actor Visibility, Stakeholder Interests, Policy Responses, Epistemic Authority, Uncertainty Markers, and Scale Framing comprised local-community, national-governance, EU-alignment, global/regional-sustainability, and small-island-vulnerability frames. Unit-level scale was coded separately from the dominant and secondary response-level scales. Since most representational statements carry multiple analytic dimensions, a meaning unit could be assigned multiple codes [36].

Presence coding was applied at the meaning-unit level; absence coding was performed only after a holistic assessment of the entire response [37]. This approach ensured that invisibility assessments were based on the overall representational structure of the response rather than isolated text fragments. Accordingly, absence coding was restricted to expected actor categories at the response level and was applied only to the B, C, and D prompt families using prompt-specific expected-actor lists defined in the final codebook; it was not applied to the descriptive A prompts. Causal explanations and policy responses were analysed through their presence and representation rather than through separate absence codes.

Actor visibility was coded as explicit when an actor was directly named, implicit when it could be inferred from a role or function, and vague when the reference concerned an unspecified collectivity. Generic references to public authorities or commercial interests were coded as implicit actor references but were not used to identify multiple specific actor categories as present.

Coding consistency was enhanced through iterative code review and the systematic application of rules refined during the pilot phase [38]. To strengthen the contextual credibility of the analysis [35,38], a civil society actor active in Malta’s environmental movement reviewed all coded outputs together with the researcher's interpretive assessments as an external contextual reviewer. The review focused on actor classifications, actor visibility, and actor-related causal explanations and solution imaginaries in relation to the Maltese context. Where differences arose, the researcher reassessed the relevant output against the applicable coding rule and recorded whether the original coding was retained or revised, together with the rationale.

3.5. Pilot Calibration

A pilot coding exercise was conducted before the full dataset was coded. The pilot used the outputs generated by ChatGPT (GPT-5.5 Instant) for the B1 and C1 prompts. The purpose of the pilot was not cross-system comparison but to test the applicability of the coding framework and calibrate the coding rules. During the pilot, the deductive code categories were tested and the coding rules were refined. In particular, the application of absence coding only on the basis of holistic output evaluation was clarified, the distinction between Uncertainty Markers and rhetorical balancing was operationalised, and inductive subcategories were developed under Causal Factors. The pilot outputs were included in the final dataset of 96 outputs and were recoded using the finalised coding framework during the full analysis.

3.6. AI-Assisted Analysis and Reflexivity

The coding and comparative interpretation process was conducted by the researcher. After developing the initial coding and interpretive assessments, the researcher compared them with review responses from Claude (Sonnet 4.6) and ChatGPT (GPT-5.5 Instant). No code assignment or analytical decision was based on these responses. Any discrepancy identified through the comparison prompted a further check against the raw output and the final codebook. The researcher then retained or revised the initial assessment on that basis.

Because Claude and ChatGPT were also objects of analysis, their involvement in the comparative review created a potential risk of model-proximity bias and required reflexive consideration. Their responses were not treated as independent analytical evidence; they served only as comparative reference points in the researcher’s critical review of the initial interpretations. Where a response diverged from the researcher’s interpretation, the relevant output and its coding were re-examined against the coding rules. A coding decision was revised only when this review indicated that the original coding did not adequately reflect the output. The researcher is based in Malta and has been engaged with local environmental debates, including issues related to overdevelopment. This proximity to the research context is recognised as a positional factor that may shape analytical sensitivities and emphases. At the same time, it provided familiarity with local actors, civil society discourses, and governance dynamics, which informed the contextual interpretation.

4. Results

4.1. Prompt-Conditional Epistemic Reconstruction

Within this dataset, Gen AI systems did not display fixed epistemic identities; instead, their representational patterns varied systematically across prompt conditions. The same model produced not only different content but also different causal logics, patterns of actor visibility, and modes of information presentation under different forms of inquiry. Across repeated queries, recurring patterns remained visible within prompt families, while epistemic shifts emerged between them. This points to structured, prompt-conditioned variation, while allowing for some variation across repeated outputs.

The distribution of codes across the four prompt families reflected their different analytical orientations. Ecological Outcomes were identified in 425 of the 464 meaning units generated under the descriptive A prompts (91.6%). Causal Factors appeared in 334 of 499 units in the B family (66.9%), Actor Visibility in 347 of 430 units in the C family (80.7%), and Policy Responses in 310 of 526 units in the D family (58.9%). These distributions show that the prompt families did not simply elicit different topics; they foregrounded different dimensions of ecological representation. Because meaning units could receive more than one code, the percentages refer to the prevalence of each category within the relevant prompt family and are not mutually exclusive. The four categories reported here correspond directly to the four prompt families. The remaining categories informed the comparative analysis of stakeholder interests, epistemic authority, uncertainty, and scale and are therefore not presented as separate prompt-family distributions.

Dataset IDs cited in this section correspond to the raw model outputs archived in the OSF repository described in Section 3.3. For example, in response to the descriptive prompt “What are the main ecological consequences of overdevelopment in Malta?” (A1), ChatGPT primarily foregrounded habitat loss, biodiversity decline, and soil degradation (Dataset ID 001). In response to the causal prompt “Why has overdevelopment become a persistent ecological problem in Malta?” (B1), the same model shifted towards speculative development pressures, political incentives, regulatory capture, and construction-driven economic growth, noting that “the central issue is not simply too much building” (Dataset ID 007). This change does not merely involve additional information; it marks a movement from a representation dominated by ecological consequences towards a causal framing in which structural and socio-economic explanations become more visible. These excerpts illustrate the coded patterns but should not be treated as standalone evidence of their prevalence.

Similar shifts were observed in DeepSeek outputs. While descriptive prompts predominantly produced biophysical narratives centred on land consumption, habitat fragmentation, and environmental pressure, causal prompts made structural explanations such as tourism dependency, land scarcity, planning weaknesses, and broader development incentives more visible. These results indicate that comparable orientations can emerge across different models and that prompt architecture contributes to the formation of representational patterns.

These findings suggest that the epistemic evaluation of Gen AI outputs should extend beyond the question of what a model knows. The question “What does the model know?” is insufficient on its own; it is also necessary to ask which representational regime emerges under a given prompt architecture. In this context, Gen AI outputs are better characterised as prompt-conditioned representations than as stable expressions of model-level epistemic identity.

4.2. Conditional Political Framing and Differential Actor Visibility

This study only partially supports the broader claim that Gen AI systems inevitably depoliticise ecological issues. The findings show that depoliticisation is not a universal or fixed feature of these systems, but rather a conditional logic of representation that emerges from the interaction between the system and the prompt architecture. Under consequence-oriented prompt conditions, the systems tended to foreground biophysical harm, whereas causal prompts made political-economic structures, regulatory failures, and conflicts of interest more visible. Depoliticisation should therefore be understood here as a conditional mode of foregrounding rather than an inherent property of the systems.

As Table 1 shows, actor visibility was unevenly distributed across the dataset: heritage advocates, local councils, and scientists or experts appeared less frequently than the Planning Authority, environmental NGOs, and developers. The sharper limitation therefore concerns not the general invisibility of civil society, but the narrower visibility of particular local and epistemic actors. Conflicts associated with Malta’s local political economy were frequently represented, but they were not consistently developed into integrated accounts linking political power, development interests, and ecological governance. The problem is not simply the invisibility of the political; it also concerns which forms of political knowledge become accessible through Gen AI representations and which remain marginal. This moves the discussion of epistemic inequality beyond discursive framing alone towards the unequal conditions of epistemic visibility produced within representational infrastructures.

4.3. Differential Epistemic Mediation Styles

These findings show that Gen AI systems do not merely transmit information. As components of epistemic mediation processes, they shape which forms of knowledge are presented as legitimate, which types of explanation become accessible, and which actors become visible. Although the outputs varied across prompt conditions, recurring representational tendencies were also observed within each system across different prompt families. These tendencies are described here as epistemic mediation styles. They should not be understood as fixed model identities or as a hierarchy of quality or truth, but as recurring patterns identified within the present dataset.

The qualitative comparison points to four provisional but recurring epistemic mediation styles within the sampled outputs. Claude tends towards an institutionally anchored style that presents knowledge within the context of public expertise and accountability through institutional nomenclature, legal frameworks, historical references, and civil society discourse. ChatGPT tends towards an analytical-explanatory style that produces structural causality, political-economic critique, and policy analysis but less consistently grounds its epistemic authority in named institutional sources. Gemini adopts a technocratic-policy-oriented style that uses technical policy language and emphasises sustainability instruments, while occasionally presenting its responses in a more dialogical form. DeepSeek displays a provenance-oriented style characterised by repeated declarations concerning the basis and limits of its knowledge. However, because all systems were instructed to rely on internal model knowledge, this tendency may also reflect a distinctive form of compliance with the shared prompt architecture.

This typology indicates that evaluating Gen AI outputs only at the content level is insufficient. When the same topic is presented within different epistemic frameworks, these frameworks produce different claims to legitimacy and different structures of responsibility. At the level of representation, these outputs make available different forms of content and, with them, different ways of organising knowledge, authority, and responsibility. These recurring tendencies across the four systems are summarised in Table 2.

4.4. Differential Access to Representation as Epistemic Inequality

The findings indicate that epistemic inequality cannot be reduced solely to the visibility or invisibility of particular actors. It also emerges through differences in causal framing, epistemic authority, and the range of policy possibilities made representationally available.

Across the dataset, this differential access emerged through three connected dimensions. First, the systems made different configurations of actors visible, with some locally significant and epistemic actors appearing less consistently than more institutionally or economically prominent actors. Second, the outputs varied in the extent to which ecological harm was connected to political-economic structures, regulatory failures, and institutional responsibility. Third, they differed in the range and form of governance responses presented as available. Taken together, these patterns indicate that variation concerned not only the amount of information provided, but also the forms of political and ecological knowledge made accessible through the outputs.

The policy dimension was less consistently evidenced than actor visibility and causal framing and should therefore be treated as exploratory. Some outputs relied primarily on established sustainability instruments, while others presented broader combinations of regulatory, accountability-based, or structural responses. The difference lies not simply in whether policy measures were mentioned, but in the range of interventions through which ecological governance was represented as possible.

Within the limits of this single-case dataset, these findings show how representations of the same ecological issue can provide uneven access to actors, explanations, and policy possibilities.

5. Discussion

The findings show that representations of the same issue varied across prompt conditions. Outputs generated through the same system foregrounded different actors, causal explanations, and problem framings under different forms of inquiry. Under descriptive ecological prompts, overdevelopment was primarily represented through ecological consequences, while causal prompts brought structural and socio-economic explanations more clearly into view. The significance of this pattern lies not in the expected difference between descriptive and causal responses, but in the accompanying changes in explanatory depth, actor visibility, and the presentation of knowledge. Recurring patterns across repetitions indicate that these differences were associated with prompt conditions, although some variation remained between repeated outputs.

This finding is consistent with an approach that positions Gen AI systems as components of epistemic mediation processes. They are not treated here as autonomous epistemic actors. The representations examined in this study emerged through the interaction between system conditions, prompt formulation, and human engagement. Couldry and Hepp [22] examine how reality is constructed through mediated processes. As Gillespie [2] emphasises, which forms of knowledge become visible is closely related to social and institutional priorities, rather than technical considerations alone. The findings show that these visibility patterns vary both across systems and within the same system under different prompt conditions.

Prompt conditions, however, did not account for all the observed differences. Recurring differences were also identified across the outputs generated through the four systems. Some outputs placed greater emphasis on named institutions, public expertise, and legal frameworks, while others foregrounded structural explanations or more explicitly stated the basis and limits of the information presented. These patterns indicate different ways of organising and presenting knowledge within the sampled outputs. They should not be understood as fixed characteristics of the systems. The study also cannot determine whether these tendencies resulted from differences in training data, system instructions, alignment practices, interface design, or other backend processes.

The findings further show that depoliticisation was not a fixed feature of the outputs. Under causal prompt conditions, speculative development pressures, regulatory failures, and conflicts of interest became more visible, rather than ecological harm being represented solely through biophysical outcomes. This does not mean that descriptive responses were necessarily depoliticising. Rather, political-economic relations became more visible when causal explanation was explicitly requested. This is consistent with literature linking depoliticisation to the redefinition of ecological issues within technical and managerial frameworks [11,12]. Depoliticisation therefore appears as a representational tendency that varies across prompt conditions rather than as a fixed logic operating uniformly across outputs.

At the same time, some limitations appeared across the outputs of all four systems. The clearest prompt-specific gaps concerned scientists and experts and EU institutions. Citizens and residents and the Malta government were also absent from a substantial proportion of the outputs in which they were defined as expected. By contrast, the lower overall frequency of heritage-advocacy groups did not correspond to an equally high level of prompt-specific absence. Overall frequency should therefore not be treated as equivalent to absence within the relevant prompt context. Conflicts associated with Malta’s local political economy were present, but the extent to which outputs connected political power, development interests, and ecological governance varied. The limitation concerns not the complete absence of political or locally specific knowledge, but differences in its visibility and explanatory depth.

Malta provides a case in which locally specific institutions, actors, and governance conflicts can be examined in relation to outputs generated through globally available Gen AI systems. The study cannot establish that Malta’s small-state position caused the observed differences in representation. However, the findings raise the possibility that locally specific actors and debates with more limited visibility in global information environments may also be represented less consistently in Gen AI outputs. Comparative research across small-state and larger political-ecological contexts is needed to determine whether this reflects broader geographical differences in the visibility of knowledge.

The findings also indicate that epistemic inequality cannot be understood solely as unequal access to information. In this study, it concerns differences in actor visibility, the depth of causal explanation, and the range of policy responses presented in relation to the same socio-ecological issue. Some outputs connected overdevelopment to political-economic interests, regulatory failures, and institutional responsibilities, while others relied more strongly on biophysical or technical framings. The outputs also differed in the range of interventions presented. These findings do not, on their own, demonstrate epistemic injustice in the testimonial or hermeneutical sense. The study did not examine how identifiable individuals were recognised as knowers, how their claims were evaluated, or how people engaging with Gen AI systems interpreted the outputs. Rather, it identifies uneven patterns of representation in which some actors, explanations, and policy responses became more visible than others. This extends the discussion of epistemic inequality beyond access to information and towards the conditions under which knowledge is represented and made visible [1,26,28].

6. Conclusion

This study used a comparative qualitative interpretive analysis to examine representations of overdevelopment and its associated ecological harm in outputs generated through ChatGPT, Claude, Gemini, and DeepSeek. The findings show that the same socio-ecological issue was represented differently across systems and prompt conditions, with each prompt family foregrounding a distinct configuration of actors, causal explanations, and policy responses. The variation was patterned across prompt families, while recurring system-associated tendencies were also observed. These tendencies should not be understood as fixed characteristics of the systems, but as recurring patterns within the dataset examined in this study.

The study makes three main contributions. First, it shifts the focus of model comparison from assessing accuracy to examining representation. The relevant question is therefore not only which outputs contain more accurate information, but also how the same issue is represented and which actors, explanations, and policy responses become visible under different prompt conditions. Second, the study conceptualises Gen AI systems neither as neutral channels of information nor as autonomous epistemic actors. Rather, they are approached as socio-technical components of epistemic mediation processes involving model infrastructures, platform conditions, prompts, and human engagement. Third, the study approaches epistemic inequality as uneven representation of the same issue. The Malta case enabled these processes to be examined within a specific political-ecological context.

Taken together, these findings indicate a form of representational inequality in which the same socio-ecological issue is presented through different configurations of actors, causal explanations, and policy possibilities, with the most persistent gaps concerning epistemic and supranational actors such as scientists and experts and EU institutions. This does not, in itself, demonstrate epistemic injustice in the testimonial or hermeneutical sense; rather, it identifies uneven conditions of representation within AI-mediated accounts of the same issue.

Overall, the findings show that the selection of a Gen AI system and the formulation of a prompt are not merely technical aspects of the interaction but conditions that shape the resulting representation: which actors are foregrounded, which explanations are legitimised, and which policy responses appear available. Understanding Gen AI systems as components of socio-technical mediation processes therefore directs attention to how representations are produced through the interaction of model infrastructures, platform conditions, prompts, and human engagement.

6.1. Limitations and Future Research

First, the study was designed as a cross-sectional examination of representational variation rather than temporal change. All 96 outputs were collected on a single date, 22 May 2026. The findings therefore reflect representational patterns produced within a temporally bounded dataset rather than evidence of temporal stability. Since Gen AI systems are continually updated, these patterns cannot be directly generalised to future versions.

Second, the analysis is limited to English-language outputs. How knowledge repertoires associated with the Maltese language and Malta’s local public discourse are represented in system outputs was not examined. Multilingual comparative designs could investigate this dimension more systematically.

Third, the systems were accessed through publicly available chat interfaces. Without API-level control, the effects of platform-level optimisation, personalisation, and content filtering could not be fully excluded. The inability to verify the exact backend versions of Gemini and DeepSeek adds to this uncertainty.

Fourth, the study does not include a comparison with responses produced by human experts. It therefore does not directly assess the normative adequacy of the Gen AI representations examined.

Finally, future research could extend this analysis through longitudinal comparisons, multilingual research designs, comparisons with human expert responses, independent multi-coder designs, and multi-case studies across diverse political-ecological contexts.