When Linguistic Hierarchy Becomes Infrastructure: AI, Language, and the Reconfiguration of Linguistic Imperialism

The concept of linguistic imperialism has traditionally been used to explain how language hierarchies are reproduced through institutional and geopolitical structures. Phillipson (1992) argues that the global dominance of English was neither accidental nor purely communicative in nature; rather, it was historically sustained through colonial administration, educational policy, international organizations, and the global publishing industry. Language expansion, in this account, is closely tied to asymmetrical power relations that extend beyond linguistic form into political economy.

Within this framework, English became embedded in educational curricula, academic accreditation systems, and international diplomacy. Its dominance was visible, codified, and often explicitly legitimized. Institutional arrangements conferred authority upon English, while other languages were positioned as peripheral or local. Linguistic hierarchy was therefore reproduced through formal governance structures.

Bourdieu’s notion of linguistic capital deepens this analysis by emphasizing how languages acquire symbolic value within specific social fields (Bourdieu 1991). Linguistic competence is not merely communicative capacity; it is a resource convertible into institutional legitimacy and social mobility. Speakers of dominant languages gain access to opportunities that are less available to speakers of subordinated languages. Importantly, such capital is not inherent to the language itself but emerges from its position within structured power relations.

In classical formulations, then, linguistic imperialism is institutional: It operates through schools, policies, accreditation systems, and cultural industries. Authority is stabilized through visible structures.

Recent scholarship has revisited linguistic imperialism in the context of globalization and digital media, noting that English dominance persists in varied forms even where explicit policy enforcement has receded. These analyses extend Phillipson’s framework beyond formal governance structures and suggest that linguistic hierarchy may persist even when explicit policy enforcement recedes.

While such work successfully demonstrates that linguistic power circulates through global media systems and digital platforms, it primarily examines the visibility and distribution of English across communicative networks. It leaves comparatively underexplored the internal architectures through which language is computationally processed, ranked, and generated.

The contemporary digital environment complicates the classical institutional model not only because communication now occurs online, but because linguistic mediation increasingly takes place within machine learning systems themselves. AI-mediated communication introduces a qualitatively different site at which hierarchy may be reproduced: computational infrastructure.

The rise of large language models and machine translation systems marks a further shift in the location of linguistic mediation. Rather than relying primarily on educational institutions, publishing industries, or even media platforms, language interaction increasingly occurs within algorithmic systems trained on large-scale digital corpora. These systems are not neutral conduits; they are structured environments shaped by dataset composition, annotation practices, model architecture, and optimization objectives.

Research on multilingual NLP has documented the uneven distribution of linguistic resources across languages. Joshi et al. (2020) demonstrate that the majority of NLP research attention and dataset availability is concentrated in a small subset of high-resource languages. This imbalance reflects broader inequalities in digital language presence, including disparities in web content, linguistic resources, and online textual production. Similarly, Ruder, Vulić, and Søgaard (2019) note that cross-lingual transfer effectiveness is closely linked to the availability and structure of multilingual training data.

Large multilingual models attempt to mitigate these imbalances by training on large-scale multilingual corpora. However, their training environments remain shaped by substantial differences in data availability across languages.

Documentation of large web-scale datasets further reveals how language representation is unevenly distributed across crawled corpora (Dodge et al. 2021). These asymmetries are not simply technical artifacts but reflect broader structural inequalities in the production and distribution of digital text.

This shift suggests that linguistic dominance may increasingly operate through infrastructural default rather than institutional mandate. English need not be formally declared the preferred language of AI systems in order to function as such. When training data, evaluation benchmarks, and alignment datasets are disproportionately concentrated in English, the resulting systems may systematically exhibit stronger performance in that language. Authority emerges not from explicit policy but from architectural embedding.

The relocation of power from institutions to infrastructure does not eliminate earlier forms of linguistic imperialism. Instead, it rearticulates them. What was once stabilized through curricula and accreditation may now be stabilized through optimization gradients and data density. The site of hierarchy shifts from classrooms and policy documents to model architectures and datasets.

To conceptualize this shift, this paper introduces the notion of algorithmic linguistic privilege. Drawing on Bourdieu’s concept of linguistic capital (Bourdieu 1991), algorithmic linguistic privilege refers to the patterned advantage certain languages enjoy within AI systems due to their infrastructural embedding.

Where linguistic capital describes symbolic authority within social fields, algorithmic linguistic privilege describes infrastructural reinforcement within computational systems. Languages that occupy dense positions in training corpora, annotation pipelines, and evaluation benchmarks are more likely to generate outputs perceived as coherent, contextually nuanced, or authoritative. This advantage does not stem from intrinsic linguistic properties; it arises from differential representational saturation.

The concept of infrastructural amplification clarifies this dynamic. Historically accumulated linguistic capital intersects with computational infrastructure, producing a feedback loop between symbolic authority and model performance. English, already globally prestigious, is also densely represented in digital corpora. As Bommasani et al. (2021) observe, foundation models inherit the properties of their training data. When data concentration aligns with historical prestige, symbolic capital becomes computationally amplified.

Importantly, this process need not involve intentional bias. As Bender et al. (2021) caution in their critique of large-scale language models, training on vast but unevenly curated corpora can reproduce structural inequalities even in the absence of explicit design preference. The neutrality of scale does not guarantee equity of representation.

Algorithmic linguistic privilege thus reframes linguistic dominance in infrastructural terms. It highlights how computational systems can stabilize performance hierarchies through inherited asymmetries in textual production.

The persistence of linguistic hierarchy in AI systems can therefore be understood as a form of structural asymmetry. Structural asymmetry refers to historically accumulated inequalities in data production, curation, evaluation, and alignment that become embedded within computational architectures.

These asymmetries are neither random nor purely technical. They reflect the geopolitical concentration of publishing industries, the dominance of English-language academic indexing, and the centrality of Anglophone media in the development of the modern internet. The digital sphere did not emerge in a linguistic vacuum; it inherited global power relations that continue to shape patterns of textual production and circulation.

Couldry and Mejias (2019) describe data colonialism as the appropriation and structuring of social life through large-scale data extraction. While their analysis does not focus specifically on language technologies, it demonstrates how digital infrastructures can reproduce and intensify existing global asymmetries. When certain languages dominate digital textual production, they are more likely to occupy central positions within data extraction processes and large-scale model training. Linguistic hierarchy thus becomes embedded within the data pipelines that sustain contemporary AI systems.

The uneven distribution of linguistic resources in AI systems is often framed as a purely technical consequence of data availability. However, patterns of digital textual production are themselves historically shaped by geopolitical, economic, and linguistic power relations. Data asymmetries, therefore, cannot be understood independently from the broader historical conditions that structure the global production of knowledge and communication.

Understanding AI-mediated communication through structural asymmetry shifts analytical attention away from intrinsic linguistic properties and toward processes of historical sedimentation. The increasing multilingual capacity of contemporary AI systems does not automatically eliminate these asymmetries. While more languages may now be computationally supported, they remain unevenly positioned within the infrastructures that govern data density, model training, and evaluation regimes. Inclusion in principle does not guarantee equivalence in infrastructural depth.

In this sense, linguistic imperialism in the AI era may not appear as overt institutional enforcement. Instead, it manifests as differential embedding within computational systems. The logic of hierarchy persists, but its mechanisms become architectural rather than declarative, operating through data pipelines, model architectures, and optimization processes.

Linguistic hierarchy has therefore not disappeared in the age of AI; it has migrated from institutions into infrastructure.

If algorithmic linguistic privilege operates through computational mediation, its effects become visible in everyday linguistic practice. Multilingual AI systems accept inputs in numerous languages and are often celebrated for this inclusivity (Scao et al. 2022; Xue et al. 2021). Yet users may encounter subtle performance gradients across linguistic pathways.

When certain languages appear to yield more detailed, stable, or contextually responsive outputs, language users may adjust their practices. Bilingual users may reformulate queries in English; others may translate prompts into English before submission and translate responses back into their primary language.

Empirical research complicates the assumption that such adaptation necessarily improves outcomes. A comparative benchmark study of Swahili-language AI systems shows that translating prompts into English does not consistently enhance performance and may introduce additional errors (Jaffer and Sayer 2025). Perceptions of English as an optimization pathway, therefore, do not always align with measurable system behavior.

Such adjustments may appear individually rational but collectively reproduce linguistic hierarchy. Linguistic accommodation itself is unremarkable in multilingual settings. However, when one language repeatedly functions—or is widely perceived to function—as a more reliable interface to computational authority, adaptation acquires structural significance. English becomes not merely a communicative option but a pathway through which linguistic value appears more readily recognized.

The issue, then, is not formal freedom of choice but the patterned incentives shaping linguistic practice. Even absent explicit coercion, differential processing conditions may condition how users position their own languages in AI-mediated interaction.

To clarify this dynamic, it is useful to distinguish between tactical adaptation and structural expectation. Tactical adaptation refers to occasional strategic adjustments; structural expectation emerges when such adjustments become normalized as necessary for effective participation.

Earlier forms of linguistic imperialism made dominance explicit. Academic publication, for example, privileged English as the language of international visibility, embedding translation within formal accreditation systems. In AI-mediated communication, by contrast, no policy mandates English input. Yet repeated encounters with performance differentials—or circulating assumptions about them—can generate a tacit expectation that higher-quality output requires operating through English.

Here, the distinction between coercion and conditioning becomes central. Classical linguistic imperialism operated through institutional structures that privileged English in education, publishing, and international communication (Phillipson 1992). In AI-mediated contexts, hierarchy operates through perceived optimization pressure. Users are not compelled to abandon their primary language, but they may infer advantages in temporarily doing so. Hierarchy is encountered as a gradient rather than a rule.

Even when empirical findings complicate claims of English superiority, symbolic associations between English and computational robustness may persist. Structural expectation can thus stabilize independently of consistent performance differences. In this way, tactical adaptation can gradually crystallize into a structural expectation about how AI systems should be approached linguistically.

This paper conceptualizes these adaptive practices as compensatory linguistic labor: the additional cognitive, temporal, and translational effort undertaken by users whose languages occupy less privileged positions within computational mediation.

Compensatory linguistic labor may include:

Each step entails interpretive responsibility. Unlike routine multilingual negotiation between speakers, this labor responds to infrastructural asymmetry. The aim is not interpersonal understanding but alignment with unevenly distributed representational density.

The term labor is not metaphorical. Digital systems frequently redistribute productive effort onto users while presenting such effort as voluntary participation (Scholz 2013; Terranova 2000). In multilingual AI interaction, language users absorb the burden of negotiating structural asymmetries embedded in model design. Compensatory linguistic labor does not eliminate inequality; it redistributes its cost. Rather than correcting representational imbalance at the infrastructural level, the system permits users to approximate parity through additional effort.

Beyond effort, compensatory linguistic labor also entails epistemic risk. Translation involves interpretation and abstraction; reformulating knowledge claims within a linguistically privileged pathway may alter nuance or reshape culturally embedded concepts. Fricker’s (2007) account of epistemic injustice clarifies this risk. When structural conditions undermine an individual’s capacity to participate as a knower, injustice occurs. In AI-mediated interaction, the need to reformulate claims in a different linguistic register can shift interpretive burden onto the user. Distortion in translation may then be attributed to phrasing rather than to asymmetrical mediation.

Although AI systems appear linguistically inclusive, performance gradients can individualize responsibility. Structural asymmetry is reframed as a matter of user optimization. The system remains unchanged; linguistic adjustment becomes the user’s task.

Compensatory linguistic labor also has symbolic consequences. When English repeatedly appears—or is widely believed—to yield more authoritative responses, associations between English and epistemic reliability may be reinforced through routine interaction. Bourdieu’s (1991) framework is instructive here. Linguistic capital is sustained not only through formal recognition but through everyday reinforcement. In computational mediation, differential embedding may stabilize perceptions that certain languages are inherently more precise or analytically robust, even when such perceptions reflect infrastructural positioning rather than intrinsic superiority.

Over time, hierarchy can become naturalized. Linguistic dominance is experienced as pragmatic efficiency rather than historical contingency. English is not declared superior; it is encountered—or assumed to be—more effective. This shift from explicit assertion to experiential inference marks a central feature of the infrastructural reconfiguration of linguistic imperialism. Compensatory linguistic labor, therefore, represents the everyday mechanism through which infrastructural asymmetries are translated into lived linguistic practice in AI-mediated communication.

These dynamics suggest that linguistic hierarchy in AI-mediated communication is not maintained primarily through explicit institutional enforcement but through the everyday organization of interaction within computational environments. Algorithmic linguistic privilege does not simply shape system performance; it also reshapes linguistic practice by redistributing the effort required to achieve epistemic recognition. Compensatory linguistic labor, therefore, becomes a routine mechanism through which infrastructural asymmetries are negotiated in practice. What appears as individual linguistic choice is thus conditioned by the structural organization of computational mediation. In this sense, the persistence of linguistic hierarchy in AI systems is not imposed through overt linguistic policy but reproduced through the subtle alignment of user behavior with unevenly distributed linguistic infrastructures.

The preceding analysis suggests that linguistic hierarchy in the AI era operates through mechanisms distinct from those described in classical accounts of linguistic imperialism. In Phillipson’s (1992) formulation, language dominance was sustained through visible institutional governance—education systems, colonial administration, and international organizations. English gained authority through educational policy, institutional prestige, and geopolitical consolidation.

In AI-mediated contexts, linguistic authority is no longer stabilized primarily through ministries or diplomatic mandates. Instead, it is mediated through platforms, data pipelines, and model architectures. The site of linguistic hierarchy shifts from institutional governance to computational mediation.

This shift does not displace earlier dominance; it overlays it. English remains globally prominent in academia and commerce, but its authority is further reinforced through training data density, benchmark design, and optimization processes. Hierarchy is not confined to curricula or policy documents; it is embedded in systems that process and generate language at scale.

Such relocation alters the visibility of power. Institutional mandates are explicit and contestable. Infrastructural defaults are often encountered as technical features. Linguistic privilege may thus appear as an emergent property of scale rather than as the continuation of historical asymmetry.

A defining feature of this configuration is optimization. Foundation models are trained to maximize predictive performance across vast corpora (Bommasani et al. 2021). Because evaluation benchmarks often reflect the distribution of available annotated data, uneven corpora can reinforce representational density in already dominant languages.

Optimization does not require explicit preference. When English-language data are more abundant and diverse, performance gradients may systematically favor English without a formal declaration of priority. Linguistic authority is thereby reproduced through statistical reinforcement rather than institutional decree.

Comparative analyses of multilingual LLMs suggest that model responses to moral evaluation tasks can reflect culturally specific value orientations associated with dominant training contexts (Aksoy 2024). Infrastructural asymmetry may therefore shape not only coherence but normative framing.

This marks a transformation in linguistic governance. Whereas classical linguistic imperialism relied on institutional exclusion (Phillipson 1992), AI-mediated hierarchy operates through incentive structures embedded in optimization. Users are not prohibited from using non-dominant languages, but they may perceive advantages in aligning with the infrastructurally privileged pathway. Optimization shapes linguistic practice through differential reward rather than prohibition.

As Couldry and Mejias (2019) argue in their analysis of data colonialism, digital infrastructures structure possibilities through design. When such design conditions the recognition of linguistic value and even normative authority, governance becomes embedded in mediation itself.

The movement from coercion to optimization entails a redistribution of responsibility. When performance disparities are navigated through compensatory linguistic labor, the burden of achieving parity shifts from infrastructural design to individual language users.

This redistribution parallels analyses of digital labor, in which user activity generates value while appearing voluntary and participatory (Scholz 2013; Terranova 2000). In multilingual AI interaction, cognitive and translational work may be reframed as strategic competence rather than as a response to asymmetrical mediation.

Such framing individualizes inequality. If improved outcomes appear available through English input, the rational response becomes linguistic adjustment. Structural asymmetry is rendered manageable through personal effort, while infrastructural conditions remain intact.

When normative clarity or argumentative coherence is also associated with dominant linguistic pathways, this individualization deepens. Users may internalize not only performance expectations but epistemic hierarchies, aligning their practices with perceived centers of rationality. Because infrastructural conditioning is experienced as technical design rather than policy, it becomes more difficult to contest.

This transformation should not be reduced to simple continuity. The AI era does not merely replicate colonial language policy in computational form. Multilingual capacity represents a genuine expansion of participation (Scao et al. 2022; Xue et al. 2021). Languages previously excluded from digital systems now engage with them.

The term reconfiguration, therefore, signals both continuity and change. Historical concentrations of linguistic capital intersect with computational mediation in new ways. Institutional linguistic imperialism operated through explicit prestige hierarchies; infrastructural reconfiguration operates through routinized performance gradients and subtle normative calibrations.

AI systems neither abolish multilingual inclusion nor impose uniform conformity. Rather, they reorganize the terrain on which linguistic inequality unfolds. The central question is not whether AI is inherently imperial, but how inherited asymmetries interact with computational mediation to produce differentiated linguistic outcomes.

Viewing linguistic imperialism as infrastructural reconfiguration shifts the normative focus. Linguistic justice cannot be assessed solely through metrics such as the number of supported languages or translation accuracy. It requires examining how languages are positioned within training datasets, evaluation benchmarks, alignment regimes, and normative calibration processes.

If structural asymmetry is embedded in data pipelines (Dodge et al. 2021; Joshi et al. 2020), redressing inequality demands infrastructural intervention: balanced dataset curation, multilingual evaluation standards, diversified annotation practices, and culturally plural alignment strategies.

More fundamentally, the discussion raises a design question. If linguistic hierarchy is mediated through computational environments, can linguistic justice be pursued through deliberate architectural choices? Historical experiments with auxiliary and constructed languages demonstrate that linguistic systems can be structured toward equity. In the AI context, this does not imply replacing natural languages, but interrogating the defaults through which linguistic authority is computationally stabilized.

The task is not to reject technological development but to render visible the conditions under which languages are differentially recognized. Only by making infrastructural asymmetry legible can linguistic justice become a matter of design rather than aspiration.