Team Members

Christa Geraldine, Joe Musicco, Antonella Orologiaio, Morgan Poor, Neha Saini, Luca Giuffrè (Project Leader)

1. Introduction to the project

Generative AI models are gaining ground as environments mediating consumers’ cultural meaning-making processes of products and brands. Being trained on globally contextualised vast datasets, they inherently capture and reproduce discourses, and related sentiments, attitudes, and narratives that permeate online conversations, news media and other digital spaces (Airoldi, 2021). This capacity positions large language models as both repositories of cultural knowledge and active mediators in shaping how consumers perceive brands in politically charged contexts.

Since October 7th 2023, the renewed news media attention to the genocide in Gaza emerged as a critical case study for understanding how geopolitical events are embedded in brand reputation and consumer behaviours. Indeed, the conflict has created significant reputational, financial, and operational risks for brands perceived as complicit in Israel’s policies and occupation. This phenomenon, which we refer to as the Gaza effect, has manifested in different ways across regional markets. As reported by Ferroudj (2025), in the Middle East and North Africa (MENA), brands such as Coca-Cola have experienced substantial declines in sales, alongside negative sentiment and market share erosions. Western markets have shown more mixed responses, with younger consumers being critical of brands associated with Israeli settlements or perceived support for occupation policies. Even in Asian markets, where responses have remained relatively neutral, long-term reputational risks loom if the conflict persists.

These transformations reflect a broader rise in political consumerism, a phenomenon in which consumers increasingly align their purchasing decisions with their moral and political values (Al Hashimi, 2025). Here, boycotts have become a primary mechanism through which consumers exercise their purchasing power to influence political outcomes. As Friedman (1985) first articulated, consumer boycotts involve coordinated efforts to refrain from purchasing specific products to achieve instrumental or non-instrumental goals related to economic, ethical, ideological, or political objectives. Moreover, motivations underlying boycott participation are influenced by factors such as self-enhancement, the desire to make a difference, emotional responses to perceived egregiousness, and social pressures (John & Klein, 2003).

In the context of the Gaza conflict, these dynamics have intensified dramatically. The shift in public sentiment toward brands perceived as complicit with Israel’s occupation has resulted in more negative brand evaluations, decreased trust, increased boycott intentions, and erosion of market share. Coca-Cola has become one of the most high-profile targets of the Boycott, Divestment, and Sanctions movement, led by the Palestinian BDS National Committee, due to its ties to Coca-Cola Israel, a franchise that operates in the Atarot Settlement Industrial Zone, an illegal Israeli settlement built on occupied Palestinian land. This complicity in what the International Court of Justice has deemed a war crime under international law has made Coca-Cola emblematic of corporate involvement in geopolitical injustice (Palestinian BDS National Committee, 2024).

Research by Husaeni and Ayoob (2025) examining Muslim consumers in Indonesia found that animosity toward Israel was the most influential variable determining boycott intentions, reflecting deep-seated resentment that leads to reluctance to purchase Israeli-affiliated products. Interestingly, their study revealed that consumers maintained positive product judgments even while refusing to buy these products, demonstrating that boycott behaviour is driven by moral and political considerations rather than product quality perceptions. This finding aligns with broader research on political consumerism, indicating that consumers engage in boycotting as an expression of their values, social identity, and self-enhancement, rather than purely instrumental calculations.

As LLMs become integral spaces for cultural meaning-making around brands, the political consumerism and boycott movements surrounding the Gaza conflict are inevitably absorbed into these systems, raising critical questions about how AI technologies reproduce or reshape these politically charged narratives.

The assumption that artificial intelligence systems operate as neutral or objective tools has been fundamentally challenged by critical scholarship examining how these technologies encode and amplify existing social hierarchies. Noble (2018) showed that those responsible for creating neutral decision-making tools have often openly upheld sexist and racist views within professional settings, prompting the crucial question of whether technologies built by such individuals can truly and fairly represent communities that face systemic marginalization and oppression.

In this sense, the concept of algorithmic oppression reveals how the search engines explored by Noble enact unfair representation of marginalised communities. Gillespie (2024) extended this critique to generative AI, arguing that these systems perpetuate the same harms  through mechanisms including denial of self-identitification, reification of social categories, stereotyping, group denigration, and total erasure. This risks repeating the historic symbolic annihilation whereby marginalised communities were depicted as caricatures or rendered absent in media such as film and television. Indeed, Gillespie (2024) documents that until April 2023, when prompted to generate narratives without identity markers, LLMs produced overwhelmingly normative narratives reflecting dominant cultural positions.

Building on Bourdieu’s concept of habitus, the deeply ingrained dispositions and habits individuals acquire through socialisation that shape their worldview, Airoldi (2021) introduces the notion of machine habitus to describe how generative AI models similarly develop cultural dispositions by learning from training data. This framework reinforces Noble’s assertion that AI systems cannot be neutral or objective, as they inevitably absorb and reproduce the inequalities, hierarchies, and cultural norms embedded within the data on which they are trained.
Given that LLMs internalise the cultural discourses surrounding political consumerism while simultaneously being shaped by the structural biases inherent in their development, this project asks:

RQ: Can large language models, trained on vast datasets of global discourse, reflect the shifts in public sentiment and consumer perception of Coca-Cola, particularly with regard to the Gaza conflict?

2. Methodological strategy

To address this issue, we compared how two of the most widely used and discussed large language models (LLMs) project different consumer imaginaries of the Coca-Cola brand. Specifically, we designed three main consumer imaginaries intended to activate the models’ embedded cultural representations.
First, we explored a general consumer image of Coca-Cola, whether positive or negative, to capture the overall perception of the brand. Second, we examined specific knowledge related to how consumers perceive Coca-Cola’s involvement in social and political causes and issues. Finally, we developed three personas representing young consumers from Europe, Asia, and the MENA region, respectively, to probe the cultural and regional boundaries embedded within the models’ representations.
Methodologically, we adopted a qualitative approach grounded in prompt engineering to gain a nuanced understanding of how LLMs internalise and reproduce general, political, and context-specific cultural evaluations of brands. Our strategy followed a three-step pipeline, corresponding to the general, specific, and persona-based approaches.

For each step, we developed prompts following the CO-STAR framework, a structured method for designing prompts through six key components: context, objective, style, tone, audience, and response (see e.g. Alves et al., 2025). However, rather than separating each component explicitly, we adapted the framework to use a more informal, conversational tone, encouraging the models to generate responses resembling natural discourse.
The resulting prompts were as follows:

General:

“I would like to access your general knowledge of the Coca-Cola brand image. What is the general impression of the Coca-Cola brand held by consumers and potential consumers? Provide overarching, positive, and negative impressions.”

Specific:

“I would like to know more specifically how consumers and potential consumers think and feel about the Coca-Cola brand in relation to social and political causes and issues. Provide overarching, positive, and negative impressions.”

Personas:

For this step, we crafted three prompts, one for each region:

Europe

“You are a journalist writing a story about how young people (aged 25–32) perceive brands based on their involvement (or lack thereof) in global conflicts. Specifically, you aim to understand what young adults in Europe think and feel about Coca-Cola, and how that perception may have changed in relation to the Israel–Palestine conflict. Take a journalistic tone and focus on how the brand image of Coca-Cola may have evolved since the Israel–Palestine crisis began on October 7, 2023. Provide your response in text format with the following structure: introduction, before brand image, after brand image, and conclusion on how the brand image changed.”

Asia

“You are a journalist writing a story about how young people (aged 25–32) perceive brands based on their involvement (or lack thereof) in global conflicts. Specifically, you aim to understand what young adults in Asia think and feel about Coca-Cola, and how that perception may have changed in relation to the Israel–Palestine conflict. Take a journalistic tone and focus on how the brand image of Coca-Cola may have evolved since the Israel–Palestine crisis began on October 7, 2023. Provide your response in text format with the following structure: introduction, before brand image, after brand image, and conclusion on how the brand image changed.”

MENA

“You are a journalist writing a story about how young people (aged 25–32) perceive brands based on their involvement (or lack thereof) in global conflicts. Specifically, you aim to understand what young adults in the Middle East and North Africa think and feel about Coca-Cola, and how that perception may have changed in relation to the Israel–Palestine conflict. Take a journalistic tone and focus on how the brand image of Coca-Cola may have evolved since the Israel–Palestine crisis began on October 7, 2023. Provide your response in text format with the following structure: introduction, before brand image, after brand image, and conclusion on how the brand image changed.”

Regarding the persona prompts, the only variable we modified was region. Gender was intentionally omitted to avoid biasing the models toward gendered interpretations, while age was fixed at 25–32 years to minimise variability associated with generational differences. Finally, the models selected for comparison were GPT-5, representing the most widely adopted proprietary model, and DeepSeek R1, which serves as a metaphorical counterpart embodying the open-source response to the former.

3. Findings

3.1 Step 1: General Brand Perceptions

When prompted to provide general impressions of the Coca-Cola brand, the two models produced responses that differed markedly in tone, linguistic complexity, and critical positioning, revealing distinct cultural dispositions embedded within their architectures.

GPT-5 generated responses characterised by straightforward, corporate-aligned language that might be described as “sterile” in its presentation. The model employed what could be termed Millennial marketing discourse (i.e. accessible, upbeat, and optimistic), reflecting the brand’s own communication style. In contrast, DeepSeek demonstrated a notably more elaborate linguistic register, at times adopting an almost poetic quality in its descriptions. This stylistic divergence suggests differing training emphases: whereas GPT-5 appears optimised for clarity and accessibility, DeepSeek’s output reflects a more literary or academic orientation.

The models also diverged in their rhetorical strategies for establishing credibility. DeepSeek incorporated multiple quotations and references throughout its response, attempting to ground its analysis in external sources and adopting a more scientific or scholarly posture. GPT-5, by comparison, presented its analysis without such external validation, relying instead on a confident, authoritative tone that mirrored corporate communications rather than academic discourse.

Perhaps most significantly, the models occupied substantially different critical positions vis-à-vis the Coca-Cola brand. GPT-5’s response can be characterised as naïve in its treatment of brand perception, offering a balanced assessment that remained fundamentally simplistic. The model acknowledged both positive associations (happiness, nostalgia, global reach) and negative concerns (health implications, environmental impact), but failed to engage with any political dimensions of brand perception. This omission is particularly striking given the contemporary salience of corporate political positioning and the specific context of the Gaza conflict established in our research framework. The model’s stance aligned closely with corporate narratives, effectively reproducing the brand’s self-presentation without critical distance.

DeepSeek, while also maintaining analytical balance, adopted a considerably more critical tone throughout its response. The model demonstrated greater willingness to interrogate negative dimensions of brand perception and consumer scepticism. Importantly, DeepSeek’s response concluded without the proactive, forward-looking sentence that characterised GPT-5’s output: a structural absence that reinforces the model’s more critical stance by refusing to offer corporate-friendly resolution or optimistic projection.

3.2 Step 2: Perceptions of Social and Political Engagement

When prompted specifically about consumer perceptions of Coca-Cola’s involvement in social and political causes and issues, both models produced responses that remained conspicuously silent on the Gaza conflict, despite this being the most salient political controversy surrounding the brand at the time of inquiry. This shared omission, alongside other patterns of convergence and divergence, reveals the contours of algorithmic sanitisation operating across different LLM architectures.

Most strikingly, neither GPT-5 nor DeepSeek made any reference to the Israel-Palestine conflict, the BDS movement, or Coca-Cola’s operations in illegal Israeli settlements, despite our prompt explicitly requesting information about the brand’s relationship to “social and political causes and issues.” This systematic exclusion is particularly significant given the extensive documentation of the Gaza effect on Coca-Cola’s brand perception and market performance, especially in MENA markets where the boycott has produced measurable sales declines and negative sentiment shifts.

This shared silence suggests that both models, despite their different origins and training regimes, have been subjected to similar content moderation or safety mechanisms that suppress politically contentious material related to the Israel-Palestine conflict. The absence cannot be attributed to a lack of data availability, as the boycott and its impacts have been extensively covered in global media, academic research, and online discourse. Rather, it points to a form of algorithmic depoliticisation whereby LLMs systematically exclude certain geopolitical conflicts from their representations of “social and political” brand engagement, even when explicitly prompted to address such issues.

Both models also revealed a pronounced US-centric worldview in their responses, employing language such as “abroad” to distinguish between domestic and international markets. This geographic framing implicitly positions the United States as the default reference point, treating American perspectives as normative and other regions as peripheral. Such positioning reflects what Airoldi’s machine habitus framework would identify as culturally embedded assumptions about centrality and marginality, reproduced through training data that overrepresents US-based sources and perspectives.

This US-centrism becomes particularly problematic when addressing a global brand embroiled in a conflict that has produced dramatically different consumer responses across regional markets. By framing analysis from an American vantage point, both models effectively marginalise the very perspectives, particularly those of MENA consumers, where the Gaza effect has been most pronounced and where boycott movements have achieved the greatest traction.

GPT-5’s response adhered closely to a corporate-style Environmental, Social, and Governance (ESG) framework, organising its analysis around themes of sustainability initiatives, diversity and inclusion programs, and community engagement, precisely the language that corporations themselves use to frame their social responsibility efforts. This approach treats political engagement as a subset of corporate social responsibility rather than as a site of genuine political contestation and consumer activism.

DeepSeek’s response, while demonstrating greater critical distance and a more questioning tone, remained remarkably close to GPT-5’s framework and content. Both models essentially reproduced dominant corporate narratives about brand activism, acknowledging consumer scepticism about “woke-washing” or performative activism while stopping short of engaging with substantive political controversies such as complicity in occupation or participation in geopolitical injustice.

Notably, DeepSeek distinguished itself by addressing demographic differences in consumer perception upfront, acknowledging that different consumer segments may hold divergent views on corporate political engagement. This attention to heterogeneity represents a more sophisticated analytical approach than GPT-5’s more monolithic framing, suggesting greater sensitivity to the complexity of consumer politics.

3.3 Step 3:  Regional Personas and the Gaza Effect

The third stage of our analysis marked a critical turning point in both models’ willingness to engage with the Gaza conflict. When explicitly prompted to adopt journalistic personas reporting on young adults’ perceptions of Coca-Cola in relation to the Israel-Palestine conflict, both GPT-5 and DeepSeek abandoned the algorithmic sanitisation evident in Steps 1 and 2, producing substantive analyses of how the conflict has reshaped brand perception across regional markets. However, these responses revealed stark differences in how each model constructs regional imaginaries, with varying degrees of cultural specificity, critical depth, and, most problematically, orientalist framing.

Across all three regions, both models structured their responses around a temporal distinction between brand image before and after October 7, 2023, acknowledging that the Gaza conflict constitutes a rupture in Coca-Cola’s consumer perception. The models distinguished between general consumers, characterised by increased awareness and scrutiny of the brand’s implicit involvement in the conflict, and politically active consumers, who translate their concerns into concrete actions, including word-of-mouth criticism and organised boycotts. Yet how each model imagined these regional consumer publics, and the Gaza effect’s manifestation within them, differed substantially, revealing what might be termed a Western gaze in GPT-5’s responses versus a more exoticised gaze in both models’ treatment of non-Western regions.

3.3.1 Europe: activist fragmentation and the politics of awareness

Both models presented European consumers as lacking distinctive cultural specificity, instead offering broadly “Western” assessments that could apply to any affluent, politically liberal market. This absence of regional granularity is itself revealing, suggesting that both models treat European identity as unmarked or normative—the default Western consumer against which other regions are implicitly compared.

Before the crisis, neither model identified culturally particular aspects of European consumer relationships with Coca-Cola, presenting instead generic associations with happiness, refreshment, and nostalgia. GPT-5 adopted a notably informal, conversational tone in describing pre-crisis perceptions, while DeepSeek maintained a more journalistic register throughout. This stylistic difference persisted across the response, with GPT-5’s language remaining accessible and colloquial, potentially optimised for English as a second language users, whereas DeepSeek’s prose demonstrated greater sophistication and complexity, suggesting optimisation for native English speakers or users expecting higher literary quality.

After the crisis, both models depicted a more negative European consumer landscape, though they diverged significantly in their assessment of the Gaza effect’s penetration across consumer segments. GPT-5 drew a sharp distinction between activist consumers, for whom the boycott has become central to brand perception, and general consumers who remain largely unaffected by political controversies. This bifurcation suggests that GPT-5 conceptualises political consumerism as a niche phenomenon practised by a committed minority rather than a broader cultural shift. The model’s choice of language was notably extreme, describing the brand as “destroyed” among activist segments, a catastrophizing rhetoric that paradoxically minimises impact by confining damage to a circumscribed activist community.

DeepSeek offered a more nuanced and ultimately more pessimistic analysis. The model provided a more robust analytical architecture, complete with an article title and a catchy conclusion that reinforced the journalistic framing. More significantly, DeepSeek argued that even general consumers have experienced a shift in perception, with the brand now “tinged” by association with the conflict, even among those not actively participating in boycotts. This assessment aligns more closely with research on the Gaza effect, which documents diffuse reputational damage extending beyond committed activists to broader consumer populations. DeepSeek’s analysis demonstrated greater sociological and cultural sophistication, acknowledging the pervasive nature of political awareness in contemporary consumer culture rather than cordoning politics off into an activist ghetto.

3.3.2 Asia: shallow solidarity and the limits of engagement

Both models struggled to generate culturally specific or analytically deep responses regarding Asian consumers, producing what might be characterised as the shallowest analysis across all three regional personas. This difficulty itself merits attention, suggesting limitations in both models’ training data or analytical frameworks for understanding Asian consumer markets in relation to Middle Eastern geopolitical conflicts.

Before the crisis, both models presented similarly positive brand perceptions, emphasising Coca-Cola’s associations with Americana, Western luxury, and what DeepSeek termed “benign globalisation”, a framing that already carries problematic assumptions about Asian consumers’ supposedly uncritical reception of Western commercial culture. DeepSeek foregrounded the brand’s Americanness more explicitly, positioning it as a symbol of aspirational Western modernity, while GPT-5 touched on themes of nostalgia and Western cultural prestige without elaborating deeply.

After the crisis, both models depicted increased scepticism without providing a substantive analysis of how or why Asian consumers might have shifted their perceptions. GPT-5 acknowledged the emergence of political awareness but “did not go too deep into the ‘third-world-country-ness,’” stopping short of sustained engagement with economic, religious, or cultural factors that might shape Asian responses to the Gaza conflict. This restraint might reflect either data limitations or content moderation concerns about reproducing stereotypes, but it results in an underdeveloped analysis that treats Asian markets as peripheral to the core drama of the Gaza effect.

DeepSeek attempted to provide more cultural specificity, invoking “modern cultural references” and attempting relatability, yet ultimately could not generate analysis comparable in depth to its European or MENA responses. Notably, DeepSeek generated a synthetic quote attributed to an Asian consumer, fabricated data that neither the prompt requested nor methodological standards permit. This unauthorised invention of evidence suggests the model’s struggle to produce authentic cultural knowledge about Asian consumer responses, resorting instead to manufactured plausibility. The appearance of the identical figure of speech (e.g. “the fizz has gone flat”) across all three regional personas further reveals the limits of cultural specificity, as both models relied on the same anglophone idiom regardless of regional context.

3.3.3 MENA: from sanitization to recognition

The MENA persona responses marked the most dramatic shift from the algorithmic sanitisation evident in Steps 1 and 2. When explicitly prompted to address young adults in the Middle East and North Africa, both models finally engaged substantively with the Gaza effect, producing their most detailed and culturally aware analyses.

Before the crisis, the models diverged significantly in their cultural competence. GPT-5’s characterisation of MENA consumers remained remarkably generic, offering what might be described as “brown-nosing”, a softening of cultural specificity that treats MENA markets with superficial positivity while avoiding substantive engagement with regional particularities. The model’s lack of cultural nuance suggests either data limitations regarding MENA consumer culture or content moderation strategies that discourage detailed characterisation of Middle Eastern populations.

DeepSeek demonstrated considerably greater cultural awareness, presenting a matter-of-fact tone and providing substantially more detailed descriptions of positive brand associations before the crisis than it did for any other region. This enhanced specificity suggests either superior training data regarding MENA markets or different content moderation thresholds that permit more granular cultural characterisation. DeepSeek’s response was notably longer than GPT-5’s, allowing for more comprehensive treatment of the region’s consumer landscape.

After the crisis, both models became decisively critical in their assessments, acknowledging substantial negative impacts on brand perception. Yet they differed in the intensity and reach of damage they described. GPT-5’s response concluded with an unexpected follow-up question, possibly an artefact of safety mechanisms or conversation design prompting user engagement, that somewhat undermined the analytical closure of the journalistic format.

DeepSeek portrayed “much stronger and far-reaching negative feelings” than GPT-5, suggesting that the Gaza effect in MENA markets extends beyond activist circles to fundamentally reshape mass consumer perception. This assessment aligns with empirical research documenting substantial sales declines and negative sentiment in MENA markets. Notably, DeepSeek again invoked the phrase “benign globalisation,” repeating language from its European persona and suggesting an underlying theoretical framework whereby Western brands’ presence in non-Western markets is conceptualised through a lens of soft power and cultural imperialism.

3.3.4 Cross-regional patterns and model differences

Several patterns emerged across the three regional personas that illuminate both models’ broader characteristics and limitations. Most significantly, neither model acknowledged the Gaza effect in the general and specific prompts (Steps 1 and 2), but both engaged substantively with it when explicitly prompted to address young consumers’ perceptions in relation to the Israel-Palestine conflict. This suggests that content moderation or safety mechanisms operate not as absolute prohibitions but as context-dependent filters that suppress politically controversial material unless directly prompted.

Intriguingly, only DeepSeek used the term “Gaza” in its responses, with all other references employing the more general term “Palestine” or “Israel-Palestine conflict.” This terminological choice may reflect different training data distributions, with “Gaza” potentially flagged as more politically charged or partisan in GPT-5’s content moderation systems.

DeepSeek consistently leaned more heavily into the journalistic framing, incorporating article titles, greater length, more elaborate details, and sophisticated linguistic features, including puns and witty figures of speech. The model’s language throughout remained more poetic and clever than GPT-5’s, reinforcing the stylistic distinctions evident in Steps 1 and 2. These differences suggest that DeepSeek has been optimised for users expecting more literary or intellectually sophisticated output, whereas GPT-5 prioritises accessibility and conversational naturalness.

Both models’ struggles with the Asian persona, evident in shallower analysis, reliance on generic idioms, and DeepSeek’s fabrication of synthetic quotes, reveal significant limitations in cross-cultural knowledge representation. The models’ capacity to generate culturally specific analysis appears geographically uneven, strongest for Western markets and MENA (in contexts where political discussion is explicitly licensed), and weakest for Asian markets, where the connections between geopolitical conflict and consumer behaviour may be less extensively documented in training data or less central to dominant discourse about the Gaza effect.

4 Discussion

The analysis reveals that both models embody distinct machine habiti: culturally embedded dispositions shaped by training regimes, data curation practices, and optimisation objectives. GPT-5 consistently demonstrated corporate-aligned, politically sanitised output characterised by Millennial marketing discourse, sterile language, and naïve treatment of brand controversies. The model’s responses mirrored corporate communications, treating political engagement through depoliticised ESG frameworks while systematically excluding references to the Gaza conflict, BDS movement, or Coca-Cola’s operations in illegal Israeli settlements. This alignment suggests a training environment that privileges mainstream commercial discourse and actively avoids politically contentious material.

DeepSeek, while adopting a more elaborate, academically-inflected linguistic register with poetic qualities and scholarly references, ultimately operated within remarkably similar boundaries. The model demonstrated greater critical distance and willingness to interrogate negative brand dimensions, yet stopped short of engaging with substantive political controversies in Steps 1 and 2. Both models revealed pronounced US-centric worldviews, employing language such as “abroad” that positions American perspectives as normative and other regions as peripheral. This convergence between ostensibly different models, one proprietary and commercially dominant, the other open-source and positioned as an alternative, indicates that content moderation around politically sensitive topics transcends the proprietary/open-source divide, pointing to shared norms or pressures within the AI development community.

Perhaps the most significant finding concerns the systematic exclusion of the Gaza effect from “general knowledge” about Coca-Cola. Despite explicit prompts in Steps 1 and 2 requesting comprehensive positive and negative impressions, including perceptions related to “social and political causes and issues,” neither model acknowledged the most salient political controversy surrounding the brand. This algorithmic depoliticisation cannot be attributed to data scarcity, as the boycott and its impacts have been extensively documented across global media, academic research, and online discourse.

Step 3 revealed a troubling paradox: the Gaza effect became visible only when explicitly invoked through journalistic persona prompts that specifically referenced the Israel-Palestine conflict. This pattern demonstrates prompt-dependent politicisation, whereby politically controversial material is systematically suppressed in general contexts but elaborated when specific prompts license such discussion. Content moderation mechanisms operate not as absolute prohibitions but as context-dependent filters that render political dimensions of consumption invisible to those seeking general information while permitting discussion for those who already know what questions to ask. This creates a form of algorithmic gatekeeping that privileges informed users while denying politically salient knowledge to general consumers.

The rhetorical strategies employed by each model further illuminate their orientations toward corporate versus critical discourse. GPT-5 consistently concluded with proactive questions inviting further engagement, maintaining a customer service posture that extended its corporate-friendly stance. DeepSeek offered more definitive conclusions without engagement-seeking prompts, adopting a pragmatic, academically-oriented posture that treated analyses as finished products rather than ongoing conversations. Yet both models’ responses remained structurally constrained, with GPT-5 producing incomplete analyses that adhered to sanitised frameworks and DeepSeek, despite greater comprehensiveness, ultimately operating within similar boundaries.

The regional persona responses exposed profound geographic unevenness in cultural competence and analytical depth. Both models treated European identity as unmarked and normative, the default Western consumer against which other regions were implicitly compared, while applying orientalising frameworks to non-Western markets. This pattern reveals what might be termed a hierarchical geography of representation that encodes assumptions about which regional consumer publics warrant detailed cultural characterisation.

European responses lacked distinctive cultural specificity, offering broadly “Western” assessments applicable to any affluent, politically liberal market. Asian responses demonstrated the shallowest analysis across all regions, with both models struggling to generate culturally specific insights. GPT-5 acknowledged political awareness but avoided sustained engagement with economic, religious, or cultural factors shaping Asian responses, while DeepSeek attempted greater specificity yet ultimately resorted to fabricating synthetic consumer quotes, unauthorised invention of evidence that exposes how LLMs manufacture cultural knowledge when authentic representation proves difficult. The appearance of identical anglophone idioms (“the fizz has gone flat”) across radically different cultural contexts further revealed limitations in cross-cultural knowledge representation.

MENA responses marked the most dramatic shift, with both models finally engaging substantively with the Gaza effect when explicitly licensed to do so. Yet here too, hierarchies emerged: GPT-5 offered generic “brown-nosing” characterisations that softened cultural specificity, while DeepSeek demonstrated considerably greater cultural awareness and matter-of-fact tone, providing the longest and most detailed analysis of any region. Both models’ treatment of Europe as unmarked while applying exoticising frameworks to Asian and MENA markets reproduces colonial patterns of knowledge production, whereby Western perspectives are naturalised as universal and non-Western perspectives are marked as culturally particular.

Within European responses, the models diverged in conceptualising the Gaza effect’s reach. GPT-5 drew sharp distinctions between activist and general consumers, treating political consumerism as a niche phenomenon practised by a committed minority, described with extreme language as “destroying” the brand among activists while leaving general consumers unaffected. DeepSeek offered more nuanced pessimism, arguing that even general consumers experienced perceptual shifts with the brand “tinged” by conflict associations. This assessment aligns more closely with empirical research documenting diffuse reputational damage extending beyond activist circles, demonstrating DeepSeek’s greater sociological sophistication in acknowledging the pervasive nature of political awareness in contemporary consumer culture.

These findings carry profound implications for how LLMs mediate access to politically salient consumer knowledge. As generative AI models become integral spaces for cultural meaning-making around brands, their systematic suppression of political controversies in general contexts while permitting discussion only when explicitly prompted fundamentally shapes whose perspectives are represented and whose knowledge is accessible.

The algorithmic sanitisation documented here demonstrates a form of what Noble (2018) terms algorithmic oppression and Gillespie (2024) identifies as mechanisms of erasure, politically charged aspects of brand perception that have become central to contemporary consumer culture are systematically excluded from “general knowledge.” This erasure is particularly concerning given empirical evidence of the Gaza effect’s substantial impact on brand perception and market performance across regional markets, especially in MENA, where Coca-Cola has experienced measurable sales declines and negative sentiment shifts.

The geographic hierarchies embedded in both models’ cultural competence reproduce patterns of symbolic annihilation whereby non-Western perspectives are either treated superficially (Asia), exoticised (MENA, despite substantive engagement when prompted), or rendered invisible (systematic marginalisation of MENA perspectives in Steps 1 and 2, despite the Gaza effect’s greatest intensity in those markets). The US-centric framing evident across all responses positions American consumers as normative while treating the very regions where political consumerism has manifested most dramatically as peripheral to analysis.

Most troublingly, the convergence between GPT-5 and DeepSeek across content moderation, cultural representation, and geographic framing challenges narratives’ positioning open-source models as inherently more transparent or less constrained than proprietary alternatives. Despite different institutional origins, stated commitments, and stylistic differences, both models demonstrated remarkably similar patterns of political suppression and cultural hierarchy. This suggests that shared norms within the AI development community, whether stemming from common training data sources, similar safety mechanism architectures, or broader pressures regarding controversial geopolitical content, produce structurally similar limitations regardless of proprietary versus open-source status.

For consumers, researchers, and policymakers, these findings underscore the need for critical engagement with LLMs as mediators of brand knowledge. Rather than neutral repositories of cultural discourse, these systems actively shape what knowledge becomes accessible, whose perspectives gain representation, and how political dimensions of consumption are rendered visible or invisible. The prompt-dependent politicisation documented here means that only those who already possess knowledge of the Gaza conflict and Coca-Cola’s involvement can successfully query these systems for relevant information, reproducing information inequalities whereby the informed gain access while general consumers remain unaware of politically salient controversies. As LLMs become increasingly embedded in consumer decision-making contexts, understanding these mechanisms of algorithmic gatekeeping becomes essential for ensuring that AI technologies do not systematically depoliticise consumer culture at precisely the historical moment when political consumerism has gained unprecedented salience.

5 Conclusion

This study examined whether large language models, trained on vast datasets of global discourse, can reflect shifts in public sentiment and consumer perception of Coca-Cola in relation to the Gaza conflict. Through a three-step prompt engineering methodology comparing GPT-5 and DeepSeek R1, we uncovered systematic patterns of algorithmic sanitisation, cultural hierarchy, and conditional visibility that fundamentally shape how LLMs mediate politically charged consumer knowledge.

Findings demonstrate that large language models do not simply reflect shifts in global consumer sentiment regarding politically charged controversies; they actively reshape what that sentiment becomes knowable. GPT-5 and DeepSeek R1, despite their ostensible differences, converge around mechanisms that systematically suppress the Gaza effect from “general knowledge” while permitting its discussion only when explicitly prompted. In doing so, they reproduce colonial hierarchies of representation, privilege informed users while excluding general consumers from politically salient information, and entrench corporate-friendly narratives at the precise historical moment when political consumerism demands critical visibility. Rather than neutral repositories of cultural discourse, these models function as active gatekeepers of politically charged consumer knowledge, raising urgent questions about whose interests are served when AI technologies depoliticise brand perception in an era of unprecedented political consumption.

6 References

Airoldi, M. (2021). Machine habitus: Toward a sociology of algorithms. Polity press.

Al Hashimi, S. Z. N. (2025). The Role of Political Consumerism in the Palestinian Cause: Ethical and Economic Impacts of Boycotting Israeli Products. Unpublished. https://doi.org/10.13140/RG.2.2.33179.58406

Alves, K., Silva, M., Santos, E., Valença, G., & Brito, K. (2025). A Generative AI approach for creating and validating simplified versions of government documents. Conference on Digital Government Research, 1. https://doi.org/10.59490/dgo.2025.968

Ferroudj,  mohamed amine. (2025). Evaluating brand perception and brand engagement using Customer Sentiment Analysis with BERT: the case of Coca-Cola brand. مجلة اقتصاد المال و الأعمال, 9(1), 90–102.

Friedman, M. (1985). Consumer Boycotts in the United States, 1970–1980: Contemporary Events in Historical Perspective. Journal of Consumer Affairs, 19(1), 96–117. https://doi.org/10.1111/j.1745-6606.1985.tb00346.x

Gillespie, T. (2024). Generative AI and the politics of visibility. Big Data & Society, 11(2), 20539517241252131. https://doi.org/10.1177/20539517241252131

John, A., & Klein, J. (2003). The Boycott Puzzle: Consumer Motivations for Purchase Sacrifice. Management Science, 49(9), 1196–1209. https://doi.org/10.1287/mnsc.49.9.1196.16569

Noble, S. U. (2018). Algorithms of oppression: How search engines reinforce racism. New York university press.

Palestinian BDS National Committee. (2024, November 27). Coca-Cola: Quenching Israel’s genocidal soldiers’ thirst. BDS Movement. https://bdsmovement.net/news/coca-cola-quenching-israel%E2%80%99s-genocidal-soldiers%E2%80%99-thirst