Platform

Throughout this Element, platform will be used to refer to digital services such as Facebook and YouTube. According to digital media researcher Tarleton Gillespie, platforms are online sites and services that

1. a) host, organize, and circulate users’ shared content or social interactions for them,

2. b) without having produced or commissioned that content,

3. c) built on an infrastructure, beneath that circulation of information, for processing data for customer service, advertising, and profit, and

4. d) moderate content and activity of users, using some logistics of detection, review, and enforcement.Footnote ⁶

These services are alternatively known as digital/social/new media or social networking services (SNS). The term platform is widely used, including by companies that operate them. Platforms are distinguished from traditional media in that they do not produce content. The term platform positions the companies as a neutral carrier of information, downplaying their control over the conditions under which content is produced and circulated. This is in line with their effort to claim legal immunity as intermediaries.

Global North and South

The North–South divide is a dominant yet contested framework for understanding global inequalities, replacing the three-world system and the developed-developing dichotomy. The United Nations Conference on Trade and Development uses the Global North to refer to affluent economies in Europe and North America, Japan, South Korea, Australia, New Zealand, and Israel. The Global South covers poorer economies in Latin America, Africa, the Middle East, Asia, and Oceania, excluding the aforementioned countries. Despite its geospatial connotation, the North–South divide does not correspond neatly to the hemispheres: India is considered part of the Global South despite being in the Northern hemisphere; Australia and New Zealand lie south of the equator yet belong to the Global North. The distinction is better understood as rooted in power asymmetries structured through historical colonial relations and economic dependency under global capitalism.

Critics argue that the concept oversimplifies today’s complex realities, overlooking internal inequalities and marginalization within countries (the ‘North in the South’ and the ‘South in the North’),Footnote ⁷ as well as the vast differences in population size, economic capacity, and geopolitical influence among Global South countries, which together account for two-thirds of the world’s population.Footnote ⁸ In the digital context, the North–South dichotomy also fails to capture the technological rivalry between the United States and China, the latter still being sometimes classified as a developing country from the Global South despite its sizable global economic and technological power. The terms Global North and Global South will be adopted in this Element with awareness of their limitations.

Language Community

Language community will be used loosely to describe a group of people using the same language. Sociolinguists recognize that the concept abstracts and idealizes by glossing over the internal variation of language practices and the fuzzy boundary between members and non-members. The term, along with its variant speech community, has sometimes been defined using subjective instead of objective criteria, for example, whether speakers feel that they belong together, which is not the definition adopted in this Element.

Reference to minoritized language communities in this Element denotes speakers of languages that occupy a marginalized position nationally and globally. It therefore includes not only national linguistic minorities, whose languages typically do not enjoy official status and may or may not have developed a literary tradition, but also speakers of majority languages in the Global South whose languages may enjoy institutional support at the national level. It therefore entails heterogeneous populations that constitute more than half of the world’s population.

1.1 Global Digital Divides

Early discussion on digital inequalities was centred on the term digital divide. Coined to describe the connectivity gap in rural and urban America, it first appeared in newspaper articles and subsequently in a report by the National Telecommunications and Information Administration in 1995.Footnote ¹³ Following the rapid diffusion of the internet in the United States, scholarly interest in the concept peaked in the early 2000s, but it remains relevant in many parts of the world today. Although more people come online every year, the digital divide is still a lived reality globally. According to the International Telecommunication Union, roughly 74% of the world’s population use the internet, while a quarter of the world’s population are still offline as of 2025.Footnote ¹⁴ The penetration rate is only at 34% in least developed countries. Those who are not using the internet mostly live in rural areas, with the urban–rural divide being particularly prominent in low- to middle-income countries. Connectivity is also divided by gender: men are more likely to be online than women, especially in low- to middle-income countries.

While material access remains a global challenge, as the digital landscape evolves, the complex realities of digital divide have extended beyond questions of accessibility and affordability. In the Web 2.0 environment, where participatory media thrive, internet users actively generate content and interact with one another. They go online to not only receive information but also to have a voice, form relationships, join communities, and gain access to economic opportunities. Digital inequality does not only involve access and consumption but also participation and production. With digitalization penetrating every aspect of life, the consequences of such divides are more far reaching than ever: certain populations could enjoy ‘lower prices for products, greater chances of securing a job or finding a party to vote for, better education opportunities, more and better health information and treatment, opportunities to acquire new friends, and even more chance of forming romantic relationships’.Footnote ¹⁵

Many countries in the Global South have leapfrogged in their adoption of information and communications technology, enjoying wireless internet without ever-experiencing wired internet, and using mobile devices without having had landlines and personal computers. While mobile technologies are instrumental in improving digital access, those who go online only through slower network and basic mobile devices, as is typical in the developing world, may not be able to take advantage of connectivity in the same way as those who are connected through a personal computer and high-speed internet. For one thing, mobile devices are designed more for consumption than creation, more for leisure than professional use; these tasks also require different levels of digital literacy skills.Footnote ¹⁶ Tellingly, only 1% of changes to Wikipedia are made through a mobile phone.Footnote ¹⁷ The digital divide, therefore, is not a simple dichotomy but is multi-layered,Footnote ¹⁸ encompassing device divide and usage divide.

1.2 Digital Multilingualism

Language is both a tool of communication and a marker of identity. While identity carries a non-market value that resists quantification, the communicative value of the world’s languages is hierarchically structured. Sociologist Abram De Swaan captures this in a global language systemFootnote ³⁵ that places the vast majority of languages – around 95% – in a peripheral, marginal position at the bottom of the pyramid. These languages have limited demographic weight and enjoy minimal institutional support. Above them are approximately a hundred central languages that are used by governments or in commerce, including languages such as Finnish and Dutch. Higher still are about a dozen supercentral languages that facilitate transnational communication, such as Arabic, Chinese, French, Hindi, Malay, Portuguese, Russian, Spanish, and Swahili. At the very top sits English, the global lingua franca of our present world and the only hypercentral language. A similar hierarchical stratification exists among varieties of a single language. For example, among global English varieties, Standard American English occupies the hypercentral position, followed by supercentral, central, and peripheral varieties.Footnote ³⁶ While linguistic vitality online may not fully mirror offline use, it is unsurprising that supercentral and hypercentral languages dominate digital communication.

1.3 Language and Inequalities in the Information Society

Language is a well-known proxy for social inequalities. Being key to accessing resources and opportunities, it intersects with economic inequality, cultural domination, and imparity of political participation.Footnote ⁶⁶ It also serves as a basis for exclusion, with the language practices of marginalized groups often becoming racialized and stigmatized.

A plethora of terms have been invoked in recent decades to describe different kinds of language-based inequalities that people experience. Linguistic oppression refers to gross linguistic injustice that is usually associated with a policy of forced assimilation or enforced language loss, through the prohibition or punishment of the use of a language in settings such as the classroom, or forcibly removing children from their parents and depriving them of the opportunity to learn the linguistic and cultural practices of their community, as experienced by indigenous communities in different parts of the world. This leads to what is sometimes emphatically called linguicide or linguistic genocide.Footnote ⁶⁷ The agent of oppression is normally the state. In contrast, linguistic discrimination or linguicismFootnote ⁶⁸ refers to everyday linguistic injustices based on the language someone speaks or other speech characteristics they have, such as the accent they speak with in a second language. Such discrimination may happen at an interpersonal level, such as between landlords and tenants, or within institutions. The interdisciplinary field of language and law, sometimes also known as forensic linguistics, deals with linguistic disadvantage, linguistic barriers to justice, injustice arising from miscommunication, mistranslation, or misinterpretation of linguistic evidence, etc.

Advancements in digital technologies have brought many positive benefits to minoritized languages and their speakers around the world. In the era of mass media, these communities rarely had control over media production, distribution, or consumption.Footnote ⁶⁹ By contrast, the participatory nature of Web 2.0 has opened new possibilities for language revitalization, allowing minoritized languages to gain visibility on a global scale without the high costs traditionally associated with content creation and dissemination.Footnote ⁷⁰ Multimodal affordances, such as audio and video, cater to the cultural needs of communities that have not adopted a written tradition, enabling them to engage in digital communication in ways that reflect their linguistic practices. Mobile technologies have also helped overcome geographic isolation and structural disadvantage, supporting community cohesion and self-determination.Footnote ⁷¹ Social media now plays a vital role in language learning, maintenance, and revitalization, especially by connecting speakers across diasporas. These technological affordances have contributed to increased online linguistic diversity and are sometimes seen as a force for linguistic justice. New technologies may save old languages, or so the narrative goes.

However, linguistic inequalities still plague the digital sphere. The issue extends beyond mere access to technology to encompass how these technologies are designed and deployed. In fact, the concentration of private powers in Web 2.0 and their mediation and commodification of our public and private discourse might have exacerbated global inequalities. The next section will demonstrate how dominant technology platforms make language-related decisions that profoundly affect people’s lives, reaching far beyond the conventional understanding of the digital language divide. To fully grasp the disparate impacts social media platforms have on diverse language users and communities, this study examines not only technological barriers and biases but also corporate policies and practices that produce unequal treatment. In fact, an excessive focus on technology and the belief that technology is always the solution have distracted us from more immediate remedies grounded in policy setting and equitable resource allocation.

2.1 Accessing Platforms: Interface Languages

Efforts to localize interface design vary significantly across platforms. At the time of writing, Facebook’s website setting contains 114 language options, and X is available in 48. Reddit is available in eleven languages, all of which are European. Although the number of interface languages might be one indicator of a platform’s linguistic inclusivity, the reality is more complex and nuanced. For one thing, the list of interface languages often includes closely related linguistic varieties, such as Spanish (Spain) and Spanish (Mexico). Moreover, although all interface languages are presented as fully equivalent options, their usability varies, with some language versions containing significant content gaps. For instance, if a user selects Inuktitut or Hausa as their interface language, Facebook (as of 2025) prompts them to choose a backup language from a list of fifty options for situations where the preferred language is not supported.

Across platforms, from search engines to social networking sites, a common strategy for localizing interfaces is to crowdsource translation services from volunteers, enabling companies to expand their markets worldwide.Footnote ⁷⁶ Facebook localized its platform primarily through community efforts, which may be considered an example of the gift economy and digital free labour.Footnote ⁷⁷ In 2008, it launched an application called ‘Translate Facebook’, allowing community translators to submit their work in languages chosen by Facebook, which was then reviewed and voted on by other volunteers. Although any user can volunteer, they must translate from and navigate the application in US English. The first two years of the translation application brought 108 interface languages to Facebook.Footnote ⁷⁸ However, participation in community translation dwindled over time and the application was retired in 2022. This decline highlights the limitations of unwaged digital labour, as localization efforts depend heavily on volunteers’ motivation, as well as their access to technology, language skills, and leisure time.

Platforms’ investment in interface languages is clearly shaped by market considerations. Languages selected tend to be widely spoken or serve as the dominant language of a given country. The largest transnational platforms have gone beyond national and official languages to include some regional and even artificial ones, such as Frisian or Esperanto. However, decisions about which languages to include, or exclude, can be politically charged. In December 2022, Signal, an encrypted messaging app popular among political activists, disabled its Uyghur and Cantonese interfaces in favour of using ‘the official language of a given region’.Footnote ⁷⁹ The decision drew criticism from activists who pointed out that Uyghur and Cantonese are the primary languages used in Xinjiang and Hong Kong – regions where China’s human rights practices have faced international scrutiny. Signal ultimately reversed the change after facing public pressure.

Platforms exhibit a monolingual ideology in the language options they provide. Users are typically required to select a single interface language, while platforms privilege standardized orthographies and disregard variation. For example, although K’ichee speakers in Guatemala regularly codemix with Spanish, Facebook reinforces rigid linguistic boundaries by releasing a version of its app in pure K’ichee, avoiding even nativized Spanish loanwords.Footnote ⁸⁰

When a user’s preferred language is not supported as an interface option, they can still access and contribute content, as long as they have sufficient proficiency in another language to navigate the platform. Content can be created in any language that has a digital input method, so across all major platforms, the number of content languages far exceeds the number of interface languages. For example, while YouTube has localized its interface into 83 languages, it anticipates content in as many as 230 languages, as indicated by its video language settings. In practice, many multilingual users are so accustomed to navigating the internet in a dominant language that they may not even switch to a local language interface when it is available.

For example, a survey found that Zimbabwean users of Google often do not select African language options on the webpage, even when they are aware of their availability.Footnote ⁸¹ One explanation for this behaviour lies in prevailing language ideologies: English is perceived to have more functional power than local languages when it comes to accessing information.Footnote ⁸² By contrast, Zimbabweans do use African languages on social media, reflecting a broader pattern in which English is preferred for instrumental purposes, while local languages are reserved for socializing. Similar trends have been observed elsewhere. A 2011 study found that although nearly all Tunisian Facebook users surveyed had selected the French interface, more than half used Arabic to communicate with others on the platform.Footnote ⁸³ Likewise, in Egypt, while users were evenly split between Arabic and English interface preferences, roughly three-quarters of them used Arabic in their interactions.Footnote ⁸⁴ In addition to language ideologies, users’ preferences may also reflect the quality and completeness of localization work, which can affect the usability and appeal of the local-language interface.Footnote ⁸⁵

Sometimes, digital diglossia arises out of necessity, as minoritized language speakers have no choice but to acquire some level of foreign language proficiency in order to engage with digital technologies. For example, kigay, an Aboriginal community in northern Australia, developed functionally specific digital literacy – learning just enough English to navigate menu options and file names on mobile devices, but not enough to read or write extended English texts.Footnote ⁸⁶ This has significant implications: users may accept terms and conditions in English without fully understanding what they are agreeing to, and may struggle to comprehend platform rules, policies, or enforcement mechanisms.

2.2 Availability and Quality of Machine Translation

Major social media platforms have developed machine translation capability to help users engage with content originally shared in other languages. Unlike a public service or a utility website, where users typically choose one preferred language for all website functions, social media companies are financially motivated to accommodate multilingual users across their full linguistic repertoires, as doing so increases user engagement and time spent on the platform. Both Facebook and X, for instance, allow users to select interface and content languages separately. Facebook enables automatic translation of posts from any of ninety-two languages into a user’s chosen language. On X, users can click a ‘translate post’ button to view content in their preferred language; this function relies on the Google Translate API, which supports 249 languages.Footnote ⁸⁷ X also lets users specify, from a list of sixty-seven languages, which languages they would like to see in recommended posts. YouTube offers caption translation in 125 languages, while TikTok currently provides automatic captions and translation in nine.Footnote ⁸⁸

Of the approximately 160 languages known to be used on Facebook, its machine translation capacity covers fewer than 60% of them. Machine translation models consistently perform well only for high-resource languages that have a vast bilingual corpus. The great majority of the world’s languages, which lack tens of millions of parallel sentences, are considered resource-poor.Footnote ⁸⁹ In 2022, Meta launched its ‘No Language Left Behind’ project, claiming to have developed a model capable of translating among 200 languages by using AI to construct bilingual datasets from monolingual data, thereby addressing the data scarcity for low-resource languages.Footnote ⁹⁰ Google has also expressed the ambition to extend its machine translation coverage to the next thousand languages.Footnote ⁹¹ Although the development of tools like Google’s Transformer and OpenAI’s GPT (Generative Pre-trained Transformer), based on large language models that are trained using internet content, promises major breakthroughs in machine translation, the performance gap across languages remains substantial, especially for low-resource languages with non-Latin scripts.Footnote ⁹²

Machine translation reduces the cost of accessing information and enables instant cross-cultural communication. It helps break down communication barriers in social interactions and tourist encounters. However, its limited reliability in low-resource languages poses risks, particularly in high-stakes situations such as seeking healthcare advice or receiving emergency alerts in politically volatile regions.

Content moderation on platforms often relies partially on machine translation. The accuracy of these translations can significantly impact moderation outcomes, as demonstrated in cases reviewed by the Oversight Board. In two instances involving hate speech, the Board overturned Facebook’s decisions to remove content when discrepancies were found between Facebook’s English translation and that of the Board’s translators.Footnote ⁹³ Erroneous translations can lead to unnecessary content removal, account suspensions, or other serious consequences. In 2017, Israeli police arrested a Palestinian man for posting ‘يصبحهم’ (‘yusbihuhum’, meaning ‘good morning’) on Facebook. The platform’s machine translation rendered the phrase as ‘hurt them’ in English or ‘attack them’ in Hebrew, leading to the arrest without an Arabic-speaking officer verifying the original post.Footnote ⁹⁴ This incident highlights how the questionable quality of machine translation is often not apparent to users, including law enforcement, because language options are presented as equivalent, and no warnings are provided for machine-translated texts, especially in low-resource languages.

Beyond the risk of inaccurate translation, machine translation can also perpetuate stereotypes and prejudices. Since output is generated based on the probability of word sequences from the training data, biased patterns present in the source texts can be reflected in the translations. For example, when input texts contain gender stereotypes, machine translation may reproduce these biases, such as translating a gender-neutral phrase into English with the pronoun ‘he’ for a doctor, reinforcing gender stereotypes.Footnote ⁹⁵

The expansion of machine translation could have long-term effects on language communities. Conceivably, being able to access foreign language content in one’s first language can disincentivize language shift towards dominant languages. However, some critics worry that the replacement of foreign language learning with commercial, ‘industrially prosthetic’ multilingualism ultimately encourages individual monolingualism, as users might feel less compelled to learn foreign languages when they can depend on machine translation.Footnote ⁹⁶

2.3 Interacting with Platforms: Warnings and Notifications

Platforms often do not consistently communicate with users in their chosen interface language, which is especially problematic for warning labels and other important notifications. This issue hampers efforts to combat misinformation, as many social media sites apply labels to content that may be false, misleading, disturbing, sensitive, or dangerous.

A study comparing Facebook, X (formerly Twitter), YouTube, and TikTok found that for users browsing the platforms in nine of the most widely used languages – Spanish, Indonesian, Portuguese, Hindi, Chinese, Arabic, French, Russian, and Bengali, respectively – content labels are not consistently translated into users’ preferred languages when they shared the same English content that has been flagged as containing misinformation.Footnote ⁹⁷ TikTok performed best, translating the labels across all nine languages. Facebook translated some labels, but only partially in Spanish, Russian, and Bengali. X offered some translated labels but displayed others in English. Most concerningly, YouTube’s untranslated labels were not shown in English either – they were simply hidden, meaning that users received no warning at all about the nature of the content.

One might assume that if misinformation is shared in English, then providing content labels in English would suffice. However, this overlooks the fact that English-language misinformation is often shared by and circulated among non-English speakers. In fact, limited English proficiency might affect one’s ability to assess the credibility of a post, heightening the need for warning labels in the user’s preferred language.

For instance, although Spanish is officially supported by X, a study of the 2020 US election found that warning labels on misinformation were not automatically translated into Spanish, even when users had set it as their default interface language. The research team notes that this ‘follows a broader trend observed throughout the election season, in which non-English language policy enforcement fell distinctively behind even when the narrative themselves were the same across languages’.Footnote ⁹⁸ They further noted that fact-checking measures were significantly less effective for Spanish- and Chinese-speaking communities compared to English-speaking users.

A different study that examined over 100 pieces of coronavirus-related misinformation content posted on Facebook in six different languages has similarly found that speakers of some languages are at a greater risk of exposure to misinformation than others. These include English and five supercentral languages: Arabic, French, Italian, Portuguese, and Spanish. The study suggests that the platform’s effectiveness in issuing warning to users about the quality of content does not only hinge on its machine translation capacity but also on its ability to detect violating content in different languages. Even for content flagged by third-party fact checkers, Facebook struggled to apply warning labels and reduce visibility. This issue was evident for English-language posts, where 29% of verified misinformation lacked warning labels. The failure was even more pronounced in other languages: 68% of such content in Italian and 70% in Spanish went unlabelled.Footnote ⁹⁹ These lapses meant that false claims, such as using hairdryers or high doses of vitamin C for coronavirus prevention, reached tens of millions of Facebook users without any indication that the content was misleading or dangerous.

Taken together, these studies show that even for widely used languages, content detection and user notifications do not function nearly as effectively as they do in English. Since machine learning systems perform best on resource-rich languages, the situation is likely far worse for less commonly used ones. The disproportionate risk of exposure to misinformation that non-English language speakers face can have direct impact on their health, safety, and wellbeing. Misinformation also has cumulative effects on the society by affecting elections, fuelling conspiracy theories, inciting inter-group conflicts, and undermining trust in institutions, experts, and the media.

Another type of notification that users receive informs them that their post has been removed for violating platform policy. The Oversight Board has explicitly recommended that Facebook notify users of the specific rules their content has breached. This guidance stemmed from concern over Facebook’s vague and inconsistent application of its Hate Speech Community Standard across multiple cases.Footnote ¹⁰⁰ Yet as of 2021, Facebook had implemented detailed violation notices only for English-language users, with vague assurances that the feature would be extended to other languages ‘in the future’.Footnote ¹⁰¹ In the meantime, non-English speakers are often left in the dark about why their content was removed. Some may assume that their content was removed because the platform disagreed with their views.

2.4 Access to Platform Rules

Platforms usually have an external facing set of content rules they make available to its users, and a much more elaborate set of rules available only internally. Facebook’s external facing rules, called Community Standards, are published in various language versions, but they are not created to be equal: the authoritative version is in US English. Its Transparency Center notes that ‘the US English version of the Community Standards reflects the most up-to-date set of the policies and should be used as the primary document’.Footnote ¹⁰² Inconsistencies across language versions have been noted;Footnote ¹⁰³ in one case, the English version of the Standards was already confusing, and its translation into Urdu became incomprehensible.Footnote ¹⁰⁴ More fundamentally, the Community Standards are not available in all the languages Facebook otherwise claims to support. Even where translations exist, updates could be slow. The Oversight Board has observed, for example, that updated versions in Urdu and UK English were still unavailable five months after the US English version had changed.Footnote ¹⁰⁵

As of 2025, Community Standards on Facebook were published in ninety-seven languages – about 85% of the languages the platform officially supports and 60% of the content languages known to the platform. This leaves a substantial number of users without easy access to the rules that govern their behaviour on the site. Users who cannot read the Community Standards in their preferred language receive no clear notice of what is permitted or prohibited. Their content may be removed or their accounts suspended without their understanding of what rule was broken. Nor can they easily determine whether they have grounds to report content that may be harmful or offensive to them. For example, Facebook has not translated its Community Standards into Pashto or Dari, the two official languages in Afghanistan, leaving the platform’s five million Afghan users unable to access the rules in a language they speak. Considering the importance of these platforms for social networking, public discourse, and political expression, especially in countries with limited independent media, inadequate access to platform rules has significant implications for free speech and due process.

The Oversight Board has flagged the lack of translated Community Standards in several of its decisions. In 2021-003-FB-UA, for example, the Board expressed concern that Community Standards were not available in Punjabi, a language with 118 million speakers globally, including 30 million in India, and one of the fifteen most widely spoken languages in the world. Although the Board recommended that Community Standards be translated into Punjabi,Footnote ¹⁰⁶ it used notably softer language when addressing other language groups in similar circumstances. The Board told Facebook that it ‘should also aim to make its Community Standards accessible in all languages widely spoken by its users’. While it rightly noted that this would ‘allow a full understanding of the rules’ and help users ‘engage with Facebook over content that may violate their rights’, it failed to define what qualifies as ‘widely spoken’. The vagueness is concerning, given the Board’s own observation that the absence of Community Standards in Punjabi means that its speakers’ inability to access platform rules may have effects that ‘raise human rights concerns’. Presumably the same human rights concerns apply to other minoritized language speakers. Moreover, these language gaps are systemic rather than mere oversight. The closing of these gaps therefore requires a structured and comprehensive approach, not piecemeal recommendations on a case-by-case basis.

2.5 Multilingual Capacity in Content Moderation

Content moderation on social media platforms is governed by a combination of local laws and the platform’s internal rules, with the latter typically encompassing a broader scope. For instance, Meta restricts speech that involves violent or criminal behaviour, poses safety concerns, or is deemed objectionable or inauthentic. Moderation activities can occur either before content is published, known as ex ante moderation, or after it has been shared, known as ex post moderation. Given the volume of content that is shared on social media every day,Footnote ¹⁰⁷ this moderation work relies heavily on automation. Users can also flag content they find objectionable, prompting platforms to review these reports. Human moderators are employed to identify and remove violating content and to review flagged material. Meta, for example, has 15,000 human content moderators who speak about fifty languages and work in more than twenty outsourced sites worldwide.Footnote ¹⁰⁸ Their automated tools can detect hate speech in about thirty languages and terrorist-related messages in nineteen languages. Despite being an industry leader, Meta’s capacity, both human and technological, is insufficient to handle moderation across all content languages. This leads to a situation where multilingual users may experience significantly different speech environments on the same platform, with content being more effectively moderated in certain languages than others.

Speakers of low-resource languages are at a greater risk of encountering more abusive speech or disinformation when platform rule enforcement in their language is inaccurate, delayed, or entirely absent.Footnote ¹⁰⁹ This lower enforcement accuracy also increases the likelihood that speakers of these languages will be subjected to erroneous censorship, where legitimate content is wrongly removed or suppressed. Currently, platforms do not publish data on the accuracy of content moderation as a factor of language. The Oversight Board has recommended that Meta include error rates broken down by country and language for each of its Community Standards in transparency reports.Footnote ¹¹⁰ However, implementation of this recommendation from 2021 is still “in progress” as of February 2026.Footnote ¹¹¹

The lack of effective content moderation in certain languages has led to dire consequences. In 2021, when Meta CEO Mark Zuckerberg appeared before the United States House of Representatives, Congressman Tony Cárdenas questioned him about Facebook’s ability to handle misinformation in languages other than English.Footnote ¹¹² In particular, he highlighted the problem of Spanish-language disinformation in the United States, which has led to anti-vaccine conspiracies, voter suppression, and hate crimes. During the 2020 election campaign, for example, the Trump campaign ran Spanish-language ads that falsely claimed Joe Biden was endorsed by Venezuelan President Nicolás Maduro. This underscores how gaps in moderation across different languages can have harmful impacts on democratic processes.

As these cases illustrate, harm spawned on platforms can easily lead to real world harm, affecting both platform users and non-users. To help mitigate risks, Meta developed a registry of at-risk locations that helps guide its resource allocation in content moderation. Having deemed Ethiopia as a Tier 1, or the highest, at-risk country since late 2020, Meta developed language classifiers, machine learning tools trained to automatically detect potential violations of the Community Standards, in Amharic and Oromo, two of the most widely used languages in Ethiopia. Despite such efforts, an Amnesty International report found Meta too slow to respond to harmful content. Between 2020 and 2022, armed conflict in the Tigray region resulted in the death of up to 600,000 civilians. Meta currently faces a $1.6 billion lawsuit for allegedly fuelling the Tigray War, which cited language as a key source of inadequate moderation, as the company could only review content in four out of the eighty-five languages spoken in Ethiopia.Footnote ¹¹³

Whether it is done by human or machine, content moderation tends to suffer from pragmatic deficiency, which is often considered a trade-off against efficiency, responsiveness, and scalability.Footnote ¹¹⁴ Machine learning systems have a limited ability to interpret context and are often insensitive to linguistic variation and cultural particularities. Human moderators, meanwhile, operate under significant time pressures and typically have restricted access to broader speech context, leading them to rely on scalable rule-based systems.Footnote ¹¹⁵ These rules, while efficient, are prone to generating false positives – incorrectly flagging or removing content. For example, Instagram’s DeepText identified ‘Mexican’ as a slur because of it frequently co-occurred with the word ‘illegal’ in the dataset. Automated content moderation systems are built upon datasets that contain human biases, which most likely disadvantage vulnerable groups, according to former United Nations Special Rapporteur on the Promotion and Protection of the Right to Freedom of Opinion and Expression David Kaye.Footnote ¹¹⁶

The lack of linguistic and cultural sensitivity is a built-in feature of Meta’s current system of content moderation. Meta, for example, assign ‘language-agnostic’ reviewers to moderate content through machine translation for languages where the company lacks human or algorithmic expertise but are supported by its machine translation capacities. These reviewers may not be familiar with the region where the content was flagged. Language-specific reviews are not supported when Meta deems that the demand for content moderation in a language to be low, but no detailed criteria have been published. In a telling example, three human moderators decided that a video portraying violence against gay men in Nigeria did not violate Community Standards.Footnote ¹¹⁷ Although the video caption was in English, all the verbal exchanges in the video were conducted in Igbo, a major language spoken by 33 million people in Nigeria. It turned out that the human content reviewers might not have been familiar with Igbo. Meta explained that its automated systems identified the language as English because of the caption, and routed the case to English-speaking reviewers. Even if Igbo had been correctly identified, Meta’s policy does not support large-scale content moderation in Igbo for Nigeria. Instead, content in Igbo is processed through machine translation by language-agnostic reviewers unfamiliar with Nigeria’s cultural context. It takes contextual knowledge to appreciate the danger that the video poses to its victims: same-sex relationships are criminalized in Nigeria, and hostility towards homosexuals is widespread. When the case was appealed, Meta again misidentified the language, this time as Kiswahili, further illustrating the system’s deficiencies. The video eventually went viral, amassing over 3.6 million views over five months before the Oversight Board intervened and ordered its removal.

Context sensitivity also depends on how platforms categorize markets and allocate resources. In one case, a Cuban woman’s protest speech was flagged as hate speech by both Meta’s automated ‘hostile speech’ classifier and its human reviewers. While her words may have appeared dehumanizing when read literally, the Oversight Board found that they should have been permitted when understood as a call to action within the context of political protest.Footnote ¹¹⁸ Meta explained that Spanish content from Spain receives country-specific review; content from Venezuela, Honduras, and Nicaragua belong to a single regional queue. In contrast, content from Cuba is lumped into Spanish-language content generally because it is not considered a ‘separate market’. As a result, reviewers may lack the cultural and contextual expertise specific to Cuba. A similar issue arises with Arabic: Meta has developed a content filter targeting bullying and harassment for ‘general language Arabic’, even though Arabic is spoken across a vast geographical area from Middle East to North Africa and regional Arabic varieties are not mutually intelligible. It is doubtful whether Meta’s enforcement tools are sufficiently attuned to these dialectal differences.Footnote ¹¹⁹

Content moderation processes have been shown to exhibit biases that disproportionately affect minoritized communities. A human rights impact assessment conducted in response to the May 2021 crisis in Israel and Palestine revealed significant disparities in how Meta’s moderation systems handled content in different languages. Specifically, there was over-enforcement in Arabic content, meaning a higher rate of erroneous removal of Palestinian voices, and under-enforcement in Hebrew content, allowing more violating content in Hebrew to remain undetected. After adjusting for population size, these disparities indicate that Arabic speakers experienced more frequent misclassification and censorship, raising concerns about fairness and discrimination.Footnote ¹²⁰ While Meta has developed an Arabic hostile speech classifier to address these issues, there is no equivalent tool for Hebrew. The assessment notes that such differential impacts on Arabic speakers have implications for the human rights principle of non-discrimination.

A platform’s internal processes significantly influence the consistency and quality of content moderation across different languages. For example, Facebook’s Internal Implementation Standards, that is, the internal guidelines that are used to enforce Community Standards, are provided in English, even for moderators who work in other languages.Footnote ¹²¹ This reliance on English-language guidelines could increase the likelihood of enforcement errors and raise human rights concerns for minoritized populations.Footnote ¹²² Meta has declined to translate the internal guidelines, asserting that its moderators are fluent in English.Footnote ¹²³

The uneven distribution of content moderation resources across languages reflects broader issues of market prioritization and resource allocation. Notably, the Oversight Board has raised concerns about the limited number of Urdu reviewers – fewer than fifty as of mid 2022 – despite the significant size of the Indian market, which has the highest number of Facebook users, accounting for approximately 12.2% of Facebook’s monthly active users in 2024. This disparity suggests that resource investment is not necessarily proportional to the user base size or the regional risks posed by dangerous groups.Footnote ¹²⁴ The Board’s concern was highlighted in a case where a post flagged for review under the High Impact False Positive Override, which is designed to identify and correct mistaken content removals, was never examined due to language capacity bottlenecks.

Whistleblowers who worked at Meta have revealed significant disparities in the company’s allocation of attention and resources across different markets. According to The Facebook Papers, a set of internal documents leaked to the media in 2021, 87% of Meta’s global budget for classifying misinformation was dedicated to the United States, despite US users constituting only about 10% of the platform’s daily active users worldwide. Conversely, only 13% was allocated to the rest of the world.Footnote ¹²⁵ Another whistleblower revealed that the company would prioritize campaigns in the US or Western Europe and foreign adversaries such as Russian or Iran, while ignoring abuses in smaller nations.Footnote ¹²⁶ For example, she lamented that instances like fake engagement by a former president of Honduras and human rights violations in Azerbaijan were largely ignored. Moreover, moderation errors tend to receive more attention and correction when flagged by high-profile accounts with many followers in the United States.Footnote ¹²⁷ Errors involving non-English content are less likely to be noticed or addressed.Footnote ¹²⁸

2.6 Access to Appeal Processes

No platform has guaranteed users the right to appeal a content moderation decision, even though the opportunity to contest wrongful decisions can enhance procedural fairness. Meta accepts appeals for most types of violations, and appeals may be handled by human or by its automated systems. However, not all appeals will be attended to when the company’s review capacity is exceeded. The Oversight Board has observed that non-English content is far less likely to be noticed, flagged, and given additional attention for escalated review (2022-003-IG-UA). Once users have exhausted Meta’s three-tiered internal appeal system,Footnote ¹²⁹ they can choose to further appeal to the Oversight Board. Unfortunately, the likelihood of their cases being heard is very low. In 2022, the Board received approximately 1.3 million cases but only published decisions on twelve of them.

Members of the Board are selected from different regions of the world, with efforts made to ensure that the cases they review have a balanced geographical distribution that reflects Meta’s user base.Footnote ¹³⁰ According to their bylaws 3.1.3, at least one of the five panel members assigned to each case must originate from the region primarily affected by the content.Footnote ¹³¹ While this approach is commendable, the appeals they receive do not proportionally mirror the demographics of Meta’s global users. In 2023, 38% of appeals originated from the United States and Canada, and 26% from Europe; only 12% came from Asia Pacific and Oceania, another 12% from Latin America and the Caribbean, with just 5% from Central and South Asia, 5% from the Middle East and North Africa, and 2% from sub-Saharan Africa. Yet, as of 2025, only 9% of Meta’s users are based in the United States and Canada. Fewer appeals do not necessarily suggest fewer moderation problems but may instead highlight procedural and outreach shortcomings. As discussed, many minoritized language groups cannot access platform interface and rules in their own language and might not be aware of appeal processes. They may also lack resources such as time and literacy skills needed to navigate the appeals system.

According to the Oversight Board’s 2022 bylaws (Section 4.3), Facebook and Instagram users could submit an appeal to the Oversight Board in any language, with their statements translated into English for review. However, in the most recent updated bylaws (2023, 2024 and 2025), this specific section appears to have been removed for reasons that have not been publicly disclosed. Despite this change, the Oversight Board has confirmed through personal communication that the policy remains the same: users can still appeal in any language.Footnote ¹³² The Board’s spokesperson clarified that appeals submitted in languages other than English are processed via machine translation into English, alongside the original text. This raises the earlier discussed issue that not all languages are equally well resourced when it comes to machine translation.

The Board’s website is currently available in twenty-eight languages, but not all content is translated in all these languages. For instance, announcements regarding new cases are not consistently available in every language, leading to gaps in accessible information for non-English speakers. Similarly, on Meta’s Transparency Centre, the availability of content varies across different interface languages, with many non-English language versions missing certain materials. In cases where translations are absent, content defaults to English.

2.7 Participation in Platform Governance

Platforms are a service provider to its users, who have no opportunity to influence how the service is run when they sign up. However, since online platforms have an outsized impact on public life and free expression, scholars have come to expect that they operate with due process, accountability, and even democratic governance.Footnote ¹³³ Platforms have, to a varying extent, responded to such expectations, not least as a means of gaining user trust. Large social media platforms and search engines have implemented transparency measures, such as sharing information about content moderation, content restrictions, and government requests for data. They have also occasionally offered users opportunities to help shape platform policies or to provide input to content moderation processes.Footnote ¹³⁴ For example, Facebook has conducted participatory governance experiments, including user votes on site policies; however, these efforts often faced limitations, such as low voter turnout, which led the platform to proceed with policy changes opposed by most voters.Footnote ¹³⁵ Platforms also engage experts, civil society organizations, and partner groups in fact-checking and policy development. Incorporating user input into governance structures can enhance perceptions of legitimacy and procedural justice – users are more likely to feel their voices matter when they have opportunities to influence decisions, receive transparent explanations, and see rules applied fairly.

Other than its automated systems and professionalized content moderation workforce, Meta’s content moderation model also relies on user flagging and input from external consultants. During the development of the Oversight Board, Meta conducted extensive global consultations, including twenty-eight workshops and roundtable discussions attended by over 650 individuals from eighty-eight countries.Footnote ¹³⁶ While questions can be raised about how well these participants represented Facebook’s then-2.5 billion users, it is evident that Meta tried to enhance the legitimacy of its content moderation framework through these consultations.

Unlike traditional judicial processes that are open to the public and involve direct participation by the parties involved, the Oversight Board conducts its deliberations privately behind closed doors. However, similar to how courts sometimes consider the opinions of individuals and organizations not directly involved in a case, known as amicus curiae or ‘friend of the court’, the Oversight Board seeks to incorporate broader public input on the cases it reviews. When the Board selects a case, it announces this on its website and social media channels, provides a brief, anonymized description of the case, and invites the public to submit comments within a fourteen-day window. The graph in Figure 1 shows, for the cases that the Board dealt with in its initial years,Footnote ¹³⁷ the regional origins of the public commenters compared to the regions primarily concerned by each case.

As illustrated in Figure 1, regardless of the geographic region a case pertains to, the majority of public comments submitted come from the United States, Canada, and Europe. Many of these submissions originate from organizations engaged in related issues such as disinformation or hate speech. This pattern of responses is likely influenced by how comments are solicited and the accessibility of the process. Notably, most comments, regardless of the commenters’ location, are submitted in English. Douek argues that this reliance on voluntary, often English-language submissions from wealthier, English-speaking regions means the public comment process may not be genuinely democratic or representative.Footnote ¹³⁸ It tends to favour well-resourced or powerful interest groups for two main reasons: First, although the Board has translated its judgements into eighteen languages, its communication channels, including social media, are predominantly in English, and some cases only accept public input in English. Second, since public commentary is voluntary and unpaid, economic inequalities can limit participation from populations with fewer resources. In a roundtable meeting I attended with some of the Board members in September 2023, they described the public comments system as an amazing opportunity to ask Meta questions, but acknowledged that soliciting sufficiently diverse comments remains a challenge.

Figure 1 Where the commenters come from versus which region the case is primarily concerned with in the Board’s published opinions dated between 2020 and 2024. The graph excludes two cases (2021-001-FB-FBR, on Donald Trump’s posts on 6 January 2021, and 2024-004/005/006-FB-UA, on the Palestinian protest slogan From the River to the Sea) which are significant outliers in the data because of the number of comments they attracted. The graph follows how Oversight Board divides the world into regions.

Figure 1Long description

The bubble chart compares case regions and commenter regions. The size of the bubbles indicates the volume of interactions, showing that the majority of comments come from Europe, the U.S., and Canada, regardless of the geographic region to which a case pertains. The other regions are Asia Pacific and Oceania, Central and South Asia, Latin America and Caribbean, Middle East and North Africa, and Sub-Saharan Africa.

An additional challenge to public commentary on or scholarly research into the Board’s cases is that the actual content is usually not disclosed due to privacy concerns. Instead, the Board provides a summary of the post in question, which means that observers and researchers do not have access to the actual language or specific details that may have led to content removal. Finally, there’s the further question of the actual impact of the public commentary. Although the Board briefly summarizes the opinions received in each case, it does not clearly indicate how, or if, these public inputs impact the final rulings.

2.8 Amplification and Demotion of Content

Platforms do not merely transmit what users post; they actively shape what users see.Footnote ¹³⁹ Apart from moderating content, platforms also actively engage in content recommendation. They curate and personalize feeds by selecting posts, advertisements, and out-of-network content that they believe will maximize user engagement. The dissemination, personalization, and amplification of content are driven by algorithms optimized for user engagement, based on datasets including browsing history, user demographics, sentiment analyses, and other behavioural factors.Footnote ¹⁴⁰ The inner workings of these recommendation models are largely opaque to users. Resourceful individuals and companies can exploit these systems by paying for increased engagement or manipulating AI-driven algorithms, such as by using bots, to increase exposure. How content is presented to users has profound implications where people use social media as the main source of news and information. Notably, elections and political outcomes have been influenced, sometimes decisively, by how content is disseminated and amplified on platforms like Facebook.

Social media platforms’ business model is based on advertising revenue and data extraction. As a result, content go viral not only because they resonate with many users, who like, share, or comment on posts, but also because the platform’s algorithms actively promote these posts to garner more attention. Conversely, posts that do not receive attention are not always irrelevant to people’s interests – they might not have the opportunity to receive audience exposure. A well-observed problem with this business model is that the algorithms create an echo chamber or filter bubble that reinforces users’ viewpoints even when they are not supported by facts, making the society more polarized and people less empathetic towards one another.Footnote ¹⁴¹

Meta has not publicly disclosed the specifics of how its content recommendation algorithms function, and it has declined to provide the Oversight Board with detailed insights into these systems. This lack of transparency has raised concerns about how platform design may incentivize the promotion of sensationalist or provocative content (2021-016-FB-FBR). Generally, content predicted to be engaging to a broad audience is boosted, while posts expected to generate low engagement are deprioritized. Such algorithms tend to reward outrage and inflammatory material, often at the expense of minority voices or nuanced discussions.Footnote ¹⁴² The language of a post can influence how widely it is disseminated, although major platforms do not openly reveal the criteria they use to prioritize content, partly out of fear that users could manipulate or ‘game’ the system. But there is evidence of English posts boosted at the cost of other languages. In 2023, components of Twitter’s recommendation algorithm were open-sourced, providing some insights into its content ranking. While these disclosures do not fully clarify how factors are calculated, they reveal a bias toward English-language content: tweets in English tend to be promoted more than those in other languages, and tweeting in a language different from one’s interface language is often viewed unfavourably by the ranking system.Footnote ¹⁴³ Essentially, the algorithm manifested a monolingual English ideology: it did not appear to promote other popular languages on the platform, such as Japanese or Spanish; the distinction is simply English and non-English. Tweets written in a language deemed to be ‘unknown’ to the user are heavily demoted.Footnote ¹⁴⁴ Although this is likely to be a measure to reduce noise on a user’s feed, it also reduces opportunities for exposure to cross-cultural perspectives through machine translation and prevents multilinguals from taking full advantage of their linguistic repertoires.

This analysis indicates that having equal opportunities to participate and post on social media platforms does not necessarily guarantee equal opportunities to be heard. Content creators who produce content in English are more likely to benefit from amplification and wider dissemination due to platform biases in content ranking algorithms. Such disparities in content visibility are especially significant in regions lacking strong independent media, where social media often functions as a vital platform for free expression. During crises and conflicts, when the risk of physical violence is heightened, the ability to share information becomes even more critical. For example, a human rights assessment report on the conflict between Palestine and Israel in May 2021 shows that the ability of Palestinians to share information about their experiences was more heavily curtailed than their Israeli counterparts, not only because of Meta’s differential ability in moderating content in Arabic versus Hebrew but also because of harm-mitigation measures that reduced the visibility of all repeatedly reshared content.Footnote ¹⁴⁵ These ‘break-the-glass’ measures, meant to be emergency responses to extraordinary circumstances, are seemingly content neutral but can disproportionately impact certain geopolitical contexts. The practice of algorithmically reducing the visibility of a user’s posts without notification, sometimes called shadow banning, further limits individuals’ ability to communicate and be seen, effectively silencing or marginalizing voices without their knowledge.

In summary, minoritized voices on social media platforms are often suppressed or erased due to linguistic and cultural biases inherent in content moderation and risk management practices. Additionally, when the social and communicative needs of these communities conflict with the platforms’ corporate goal of maximizing user engagement, their content may be demoted. This combination of structural biases and algorithmic incentives can significantly hinder diverse voices from being heard and represented online.

2.9 Understanding the Impact

The capacity of global populations to express themselves, access information, and engage in the digital economy and public discourse is now largely mediated by private social media platforms. The language-related decisions made by these private entities, regarding policy design, content moderation, resource allocation, and algorithmic prioritization, carry significant political implications. These choices can result in material injustices, affecting not only active platform users but also the broader communities they are part of. Such linguistic injustices manifest as epistemic inequalities (limits on access to knowledge and information), participatory inequalities (barriers to meaningful engagement), representational inequalities (underrepresentation or misrepresentation of cultural identities), and distributive inequalities (unequal distribution of resources and opportunities). Collectively, these disparities undermine the inclusivity and fairness of digital spaces for diverse language communities.

Social media platforms facilitate access to vast amounts of user-generated content, enabling peer-to-peer connectivity among language communities worldwide. However, persistent digital language divides remain, even for languages officially supported by these platforms. While advancements in machine translation improve access to information across languages, they also pose risks of errors and inaccuracies, leaving speakers of low-resource languages more vulnerable to misinformation and miscommunication. The quantity and quality of content available in different languages vary significantly, partly due to uneven platform capacities to filter harmful, dangerous, or misleading material. What starts as an epistemic gap – limited access to reliable information – can quickly escalate into crises threatening the safety and well-being of both platform users and non-users. These divides are not solely the result of uneven information availability, a bottom-up phenomenon, but are also influenced by top-down factors such as disparities in investment for harm mitigation and moderation efforts. Ultimately, the divides are mediated by power relationships, as the next section will explore further.

In the era of interactive media, assessing linguistic inequalities must extend beyond mere consumption to include content production, as active participation is essential for minoritized language speakers to have their voices heard and to meaningfully engage in the digital economy. Structural barriers such as device and connectivity disparities, along with the limitations of zero-rated services, contribute to the production gap between the Global North and the Global South. It has been observed that the great majority of people who use crowdsourcing forums, from mailing lists and blogs to social media, do not produce content actively: around 90% are passive consumers of information, 9% contribute occasionally, and 1% is actively generating content.Footnote ¹⁴⁶ Usage divide is a behavioural phenomenon and does not necessarily point to unfairness, but uneven patterns of digital production may reveal underlying structural barriers. In the US, voices of the working class are marginalized; survey data found a class-based digital production gap among those who are online,Footnote ¹⁴⁷ with education being the most important predictor of production. Since digital participation involves labour, often unpaid, those with limited time and resources are less able to contribute.Footnote ¹⁴⁸ For minoritized language users, these structural barriers, such as difficulties in understanding platform rules, being notified of content removals, or facing algorithmic biases favouring majority languages, intersect with pre-existing inequalities. This participation gap translates into unequal access to economic opportunities, as content creators who reach larger audiences can monetize their work and benefit from the digital economy, further marginalizing those who are unable to do so.

Inequality in digital content production directly fuels representational inequality, which pertains to how different peoples and issues are depicted online. For example, in 2011, the Wikipedia entries for Jerusalem varied significantly: the Arabic version described it as the largest city in occupied Palestine, while the Hebrew version called it the capital of the State of Israel.Footnote ¹⁴⁹ Such conflicting narratives exemplify how representations are contested, especially around sensitive topics like historical events, land claims, and identity. Distorted or incomplete digital representations often reflect global structural inequalities. For example, although Wikipedia exists in 360 languages,Footnote ¹⁵⁰ some editions in smaller languages were created by non-speakers residing in the Global North based on dictionaries and machine translation. An American teenager had produced half of the articles in Scots, a minority language in Europe, by inserting Scots words into English sentences. A large number of articles written in isiXhosa, which is spoken as a first language by 8 million people in South Africa, were found to be created by contributors in the Global North who had very limited knowledge of the language.Footnote ¹⁵¹ Although Wikipedia entries are not permeant and can be corrected through communal efforts, such efforts require sufficient native speakers to be online and be prepared to volunteer their labour. Social media platforms contribute to representational inequality through algorithmic amplification of certain viewpoints over others, often to the detriment of minoritized communities. Input on platform governance, such as public commentaries that the Oversight Board receives, also tends to be dominated by perspectives from high-income countries. Representational inequality reinforces existing economic, political, and cultural domination, and challenges the optimism that digital media offers a more democratic and inclusive marketplace of ideas than traditional mass media.Footnote ¹⁵²

Distributive inequality manifests in the unequal allocation of resources across different platform languages, affecting everything from the translation of rules and policies to content moderation efforts. This uneven distribution means that the transformative opportunities, and the risks, associated with connectivity are not shared equitably among users worldwide. Certain communities face language bottlenecks that limit their ability to access platform information or have their content properly moderated, while others benefit from more robust support. Although platforms naturally prioritize geopolitical contexts that are most relevant to their strategic interests, neglecting less prominent or marginalized communities can have serious human rights implications.

The following section examines how such inequalities might be mitigated, what they reveal about the structure of the global digital order, and why conceptions of linguistic justice must account for digital transformations, particularly with regard to competition in the global linguistic market and the role of private actors in shaping linguistic hierarchies.

3.1 The Goals

Linguistic inequalities both reflect and intersect with deep-seated global divides and digital platforms may have unintentionally amplified them. Addressing these issues is complex. There is no straightforward solution that can fully undo or reverse entrenched global power asymmetries, and social and political challenges extend beyond legal and technological remedies. Nonetheless, many of the negative impacts stemming from platform policies and processes can be mitigated through targeted actions. This section outlines achievable goals for the near future.

3.2 The Levers

The good news is that major transnational platforms possess the resources necessary to implement the recommendations outlined above. The bad news is that private actors do not have the same responsibilities to their platform users as governments have to their citizens. Companies are free to choose which values to prioritize, or to prioritize none at all.Footnote ¹⁶³ The question is then how they may be incentivized or coerced to act in ways that promote responsible governance. This section further focuses on the levers for change, acknowledging that some are clearly more accessible than others and each has its own limitations.

3.3 Linguistic Inequalities under Information Capitalism

It is no revelation that English dominates the digital realm. But this is not only because English is economically privileged and attracts more users in online exchanges. The powerful private actors who shape today’s digital public sphere actively prioritize the interests of English speakers through platform design and governance. As a result, non-English speakers encounter a speech environment that is less safe, struggle more to be heard, and have limited access to platform processes and decision-making. At one end of the spectrum, content from some language communities is disproportionately censored; at the other, content in certain languages is scarcely moderated at all. Such linguistic inequalities are a facet of global structural inequalities in the political economy created by platforms.

In a book confronting biases in technology, data journalism professor Meredith Broussard reminds readers that problems in the tech sector are often dismissed as glitches – temporary and unexpected errors in need of a quick fix. She points out that many of these biases are not accidental but are embedded and structural.Footnote ²⁰⁶ They cannot be solved simply by tweaking code. In the same vein, it is not the advancement of information and communication technology itself, but the business decisions made by private actors that deepen inequalities across language communities. This distinction matters: the former naturalizes inequality as an inevitable by-product of progress and shifts the burden onto those left behind to catch up; the latter opens the door to institutional, political, and regulatory solutions.

There are two common rhetorical strategies that platforms use to publicly respond to content moderation failures. The first is the normalization of deviance: ‘We are sorry, but mistakes happen.’Footnote ²⁰⁷ This narrative obscures the root causes of failure, deflects responsibility, and gradually desensitizes the public to rule violations, recasting them as routine rather than unacceptable. The second is the rhetoric of perpetual improvement: ‘We will do better – our technology is improving.’ This response is rooted in technochauvinism, the belief that technology is always the solution, regardless of the nature or complexity of the problem.Footnote ²⁰⁸

Content moderation is inherently difficult because it involves navigating conflicting values. For instance, there may be a compelling public interest in sharing graphic content that documents human rights abuses, yet the same content can also pose serious risks of harm. Moderation is further complicated by users’ ingenuity in evading controls, such as through the use of Algospeak – coded language designed to bypass automated moderation systems (such as using ‘unalive’ to refer to suicide or death, or the watermelon emoji for Palestinian resistance). However, these challenges should not be conflated with the persistent underperformance of moderation in certain languages, which often stems from chronic underinvestment. While perfection is not expected, platforms should be held to a baseline standard of content moderation across all of their content languages. They should also provide transparent, language-specific reporting that demonstrates how these standards are being upheld, as discussed earlier in this section.

The normalization of deviance has enabled a disregard for marginalized regions and populations who are not ‘big on people’s minds’,Footnote ²⁰⁹ ignoring and invisibilizing their suffering. The way these transnational media corporations concentrate their resources on markets that their stakeholders care about is similar to how attention flows around the world. The public pay more attention to news happening in places that share geographic, linguistic, historical, or cultural similarities with themselves.Footnote ²¹⁰ News value is also related to power relationships in a Euro-American centered world order. For example, the United States is heavily reported on by national media around the world due to its perceived news value, but the media in the United States do not pay nearly as much attention to the rest of the world.

The structure of the digital economy is reflected in its part in the global labour market. Platforms rely heavily on low-wage labour from different parts of the world to support their content moderation system and their AI technologies.Footnote ²¹¹ For example, Facebook employs content moderators working in call centres in the Philippines, Ireland, Mexico, Turkey, India, and Eastern Europe. Similarly, data labelling for training AI models is largely outsourced to annotators in the Global South, who do not only work long hours and receive low pay, but are also often exposed to graphic materials that affect their mental health in the long run.Footnote ²¹² Instead of being asked to contribute their linguistic and cultural knowledge, they are typically required to label data based on Western perspectives and in dominant languages.Footnote ²¹³

The political economy created by platforms contributes to global economic inequalities, which see cheap labour from the Global South powering the world economy,Footnote ²¹⁴ while the benefits of production are reaped disproportionately by the Global North. Dal Yong Jin has called the asymmetrical relationship of interdependence between transnational technology companies and the majority of people and countries platform imperialism.Footnote ²¹⁵ Imperialism has been reinvented through the digital ecosystem, which is primarily controlled by the US corporations, but with increasing competition from China. China, now a major technology powerhouse, has extended its influence by supplying technological infrastructure to developing countries through initiatives like the Digital Silk Road project. Although the rise of China has punctured the North–South divide, most countries in the Global South remain subservient to a digital world order characterized by US–China rivalry.

This new form of imperialism is no longer centred on nation states but transnational corporations, and capital accumulation is no longer based on territory as property but on intangible goods such as data and intellectual property. It is by and large a continuation of the North–South divide created by Western imperialism. Computer security specialist Bruce Schneier has gone as far as describing the relationship between users and platforms as ‘more feudal than commercial’: users provide free labour by generating data, which platforms package and sell for profit.Footnote ²¹⁶ In this conception, users are not customers to platforms; they are the products that get sold to the real customers. Unlike older forms of imperialism, which relied on the coercive force of states, digital capitalism is arguably just as imperialistic, but it operates through soft power, versed in the discourse of consent, participation, and development.

3.4 Dilemmas of Linguistic Inclusion

It is often assumed that the presence of a minority language in a majority-language setting signals recognition and inclusion, and that translation practices advance linguistic justice by enabling cross-cultural communication. However, drawing on linguistic landscape studies of public signage, sociolinguist Philipp Angermeyer argues that inclusion can also be discriminatory, if such inclusion ties deviant behaviour to speakers of certain languages. In a striking example from a regional train in Germany, safety instructions are provided in German, English, French, and Italian, while the passenger bill of rights appears only in German. By contrast, a warning about fines for fare evasion is posted in German, English, French, and Turkish. Turkish speakers are thus implicitly marked as potential fare evaders, yet their rights and safety are not equally addressed. Angermeyer terms this phenomenon punitive multilingualism.Footnote ²¹⁷ A systematic feature of punitive multilingualism is that public order signs written in subordinated languages are often ungrammatical or unidiomatic, as the texts are regularly produced through machine translation without input from actual speakers.

The concept of punitive multilingualism can also be extended to the more complex realm of digital technologies. On social media platforms, making interfaces, rules, and processes available in multiple languages is generally seen as promoting access and improving procedural fairness. But digital inclusion is not an unqualified good. Often, this is because the existing structures into which communities are being included are themselves flawed or harmful. For instance, platforms that algorithmically amplify content based on engagement metrics tend to favour sensationalism over nuance – conditions that are often unfavourable to minoritized communities. In such cases, inclusion may reproduce or even exacerbate inequalities. Inclusion is therefore not always the solution, at least not until the broader conditions of participation are improved. As a former Facebook employee reflected in hindsight, Myanmar might have been ‘far better off’ if Facebook had never entered the country.Footnote ²¹⁸

Linguistic inclusion does not always improve conditions for the communities being included. Although inadequate content moderation threatens user safety, excessive moderation can stifle free speech. While it is necessary to moderate content in different languages, developing algorithms that detect speech transgression in some languages but not others may have the effect of limiting speech unevenly across language communities and amplifying some voices and suppressing others in a public debate. When a language community is made visible to platform management through irregularities, such as political or military conflicts, platforms might mitigate risks through shadow banning the whole community. This may reduce the spread of disinformation, but it can also silence legitimate voices, including those of citizen journalists. Content moderation is a double-edged sword: just as language can be both oppressive and emancipatory, both under-moderation and over-moderation can lead to harm.

Another reason why inclusion is not always emancipatory is the risk of miscommunication when the quality of translation is poor. The misuse of machine translation in legal and medical settings can cause serious harm; for example, an asylum seeker’s credibility was undermined because their application had been generated using machine translation.Footnote ²¹⁹ Applying such technologies to languages where accuracy is low can lead to unacceptable risks, especially when users are not warned about potential inaccuracies. Even when machine translation becomes more reliable, industrialized effort to develop technological ‘solutions’ to the ‘problem’ of multilingualism may ultimately narrow language practice and linguistic freedom in the long term.Footnote ²²⁰ Well-intentioned innovations can have unintended consequences: for example, people may feel less compelled to learn other languages if they believe AI will mediate all cross-linguistic communication. Meanwhile, to ensure effective multilingual transfer, technologies in cross-linguistic information retrieval and machine translation produce translingually controlled meanings at the expense of other variations in meanings that could create fuzzy ontologies for computer scientists. Meanings that are paralinguistic or untranslatable, or ones that are dependent on silence or implicature, are dispreferred.Footnote ²²¹ In short, the desire to use technology to bridge immediate communication challenges may have longer term repercussions.

Moreover, overt inclusion and covert exclusion may happen to the same language community at the same time. For example, a platform may offer an interface in a given language but invest little in content moderation for that language, or it may moderate content without providing accessible appeal processes to its speakers. When users sign up to a platform, the range of language options in the user interface creates the illusion of access and choice. Presenting the language options as if they were equal performatively negates the language hierarchy without actually disrupting it. Visible forms of digital language support can obscure more critical but less visible forms of exclusion. Crucially, inflated claims about what big tech platforms can do for minority languages risk undermining local, community-led efforts to develop contextually appropriate solutions.Footnote ²²²

Digital inclusion of a language is often implemented without involving its language communities in the process. A study of Guarani, an indigenous language co-official with Spanish in Paraguay, found that since 2006, speakers had developed a range of grassroots solutions to digitally encode the language. In 2013, Microsoft released an operating system with Guarani language support, followed by similar offerings from Facebook, Google, and Apple. While these developments may be hailed as milestones, they were introduced unliterally as an act of generosity without the advocacy, involvement, or feedback from speakers of the language.Footnote ²²³ Rather than being treated as collaborators, speakers were positioned as passive consumers of innovation, to whom language technologies could be marketed. Without engagement with language communities, digital inclusion could come across as corporate intrusion. For example, the Mapuche nation in Chile had threatened to sue Microsoft for releasing a version of its Windows operating system in their native language Mapudungun without their permission, and considered it an act of ‘intellectual piracy’ and a violation of their right to self-determination.Footnote ²²⁴

One important reason why the absence of a language online should not be readily equated with its exclusion is the consideration of speaker agency. A study of minority languages in the Nordic countries found that while these languages have a limited presence on websites and in digital archives, interviewed speakers show little desire for their languages to have more extensive presence on the internet.Footnote ²²⁵ Digital tools such as online archives have certainly supported language documentation and revitalization in some contexts, for example, by providing access to learning materials, but not all communities share the same priorities. Some Indigenous cultures, such as the Māori in New Zealand, have refused to participate in digital libraries, because their knowledge traditions do not support the free circulation of information.Footnote ²²⁶ For many communities, digital ‘ascendance’ is not necessarily seen as a goal; instead, language revitalization is often rooted in relationships with people, place, and land.Footnote ²²⁷ As discussed earlier, the same digital tools used to support a language can also be deployed for surveillance, posing serious risks to marginalized groups facing displacement, criminalization, or state violence.Footnote ²²⁸ Ultimately, speakers of a language may choose to preserve, abandon, or develop their languages, even though they may be acting under influence that is outside of their control.Footnote ²²⁹ Outside experts sometimes carry assumptions that clash with the perspectives of language communities, such as by viewing their language as data rather than as a connection to their land and culture, or believing that endangered languages should be documented and maintained in the purest form possible. Conflicting language ideologies may also develop within language communities. Facebook and Microsoft engaged young K’ichee’ speakers to translate its platforms into the Mayan language, without seeking approval from the local governing body or consulting elders, who are traditionally regarded as authoritative custodians of the ancestral language. The young people view their language proficiency as a personal skill that can be commodified in the market, while older generations see the language as a sacred ethnic artefact that belongs to the community as a whole.Footnote ²³⁰ Linguistic inclusion by transnational corporations could therefore become a site of intergenerational ideological struggle.

Finally, it is important to remember that whether linguistic inclusion is experienced as emancipatory or punitive often depends on perception – and may be quite disconnected from the intentions of those who design the technology. Software localization, for example, is typically pursued as a market expansion strategy to increase profit, yet it may be interpreted by end users as a gesture toward bridging the digital divide. This gap between design intent and user interpretation reflects the concept of interpretive flexibility, the idea that technological artefacts are culturally constructed and can be understood differently by those who create them and those who use them.Footnote ²³¹

3.5 Competition in the Global Linguistic Market

Beyond its impact on individual users and non-users, globalized digital media affect the value of languages at a macroscopic level, with significant implications for the sociolinguistic experiences and life chances of their speakers and future generations of speakers. Pierre Bourdieu’s sociology of social stratification, alongside its key concepts of field, capital, and habitus, offers a useful analytical framework. A field, or market, is a space in which actors, whose relations are defined by their access to resources, or capital (which may be economic, social, cultural, political, etc.), compete with one another based on rules and logic specific to the field.Footnote ²³² Their behaviour is guided or constrained by predispositions, or habitus, such as their capacities and preferences. Linguistic practices are thus a product of the interaction between a linguistic habitus and a linguistic market;Footnote ²³³ for example, one may adopt a particular style or language variety in their linguistic repertoire based on its anticipated value in a market. By viewing language as a form of cultural capital and a component of social capital, Bourdieu reframes it as a means of action and power, rather than merely a tool for communication.

In Bourdieusian terms, one can therefore speak of information capital or digital capital that circulates in the digital field and conceptualize the digital divide as the uneven distribution of such capital. The digital field is dominated by the American big five (Google, Apple, Meta, Amazon, and Microsoft) and their Chinese counterparts (Alibaba, Tencent, ByteDance, etc.), who command enormous digital capital. These corporations are not only among the largest holders of personal data globally but also provide the foundational infrastructure on which countless other digital companies and content creators depend. Their dominance reinforces structural inequalities. For example, Google’s AdWords has been described as a global, real-time, multilingual auction system for advertising placement – one that effectively constitutes a global linguistic market. In such a market, advertisers are incentivized to prioritize more profitable languages, reinforcing linguistic hierarchies and contributing to the marginalization of less economically valuable ones.Footnote ²³⁴

Likewise, social media can be understood as a subfield within the broader digital field, where content creators seek to accumulate both economic and symbolic capital, while platforms extract and convert user activity into digital capital through data collection and analysis, in a digital economy financed by advertisers.Footnote ²³⁵ In this subfield, content creators compete for visibility, engagement, and monetization, just as platforms compete for user attention and market dominance. This competitive structure is shaped by the platforms themselves, which design the ‘rules of the game’. As discussed earlier, these rules are often calibrated in ways that disadvantage users from minoritized language communities, thereby further devaluing peripheral languages and reinforcing existing linguistic hierarchies.

Linguistic products, such as literary works, that once circulated within the enclosed boundaries of local markets were to some extent protected from external competition. Similar to how state formation creates the conditions for a unified linguistic market, usually dominated by an official language within the state, digital media platforms have forged a global linguistic market, in which content from different corners of the world are put in much more direct competition with one another than before. Although these platforms have lowered the cost of publishing in minoritized languages, they embed incentives to prioritize language practices that maximize engagement, putting pressure on minoritized language practices.

Revisiting De Swaan’s framework of global language hierarchies (see Section 1.2.1), we are reminded that the communication potential of a language is a strong motivation for learning a foreign language; as a result, language learning occurs mostly upward in the hierarchy. Few peripheral language speakers can afford to remain monolingual, but monolingualism is common among supercentral and hypercentral language speakers. Speakers of peripheral languages tend to have more reasons to be interested in what speakers of more central languages have to say, rather than the other way round. Translation therefore typically flows downward, such as from hypercentral to central;Footnote ²³⁶ when a peripheral language needs to be translated into another, it is sometimes bridged through a more central language. The use of a bridging language is known as pivot translation. For example, the European Union translates across its twenty-four official languages through using English, French, and German as pivot languages.

Users often use vernacular languages on the platforms to connect and socialize with their local or diasporic communities. However, peripheral language speakers may switch to a more central language when accessing platforms not only to be part of a potentially less hostile speech environment but also to access a wider range of content and to improve their chance of success in capital-generating activities. This practice, known as digital diglossia, is most commonly observed among internet users whose first languages occupy relatively low positions in the global language hierarchy (see relevant discussion in Section 1.2.3 and Section 2.1). These users are typically members of the educated elite in their societies, having acquired bilingual proficiency in both their local language and a more dominant, central language. Their command of a central language enables them to navigate platforms more effectively. Yet this linguistic flexibility also presents a dilemma in content production. Using a central language may allow creators to reach wider audiences, but it also places them in competition with a much larger pool of content producers. Conversely, producing content in a peripheral language may limit their reach to a smaller, more niche audience, though with the advantage of less competition for visibility and attention.

Does digital diglossia lead to language shift? Although Ferguson argues that diglossia could remain stable,Footnote ²³⁷ Penelope Eckert contends that, in most circumstances, it continually evolves.Footnote ²³⁸ Using the displacement of Gascon by French as an example, she observes that as domains associated with the low variety become obsolete, the language can be stigmatized as that of those who had been ‘left behind’, while the domains of the high variety are exclusively tied to social and economic survival and success. In the case of digital diglossia, as digitalization permeates our daily lives, the domains that require the adoption of digitally thriving languages continue to expand. Social media platforms also foster language contact, which is the primary driver for language shift. Digital diglossia is therefore likely to incentivize language shift. Rather than remaining stable, individual bilingualism often represents a transitional phase in the process of language shift, especially when the diglossic functions are poorly compartmentalized and when bilingualism is widespread rather than being confined to a small, privileged class.Footnote ²³⁹

Language is a hyper-collective good – ‘the more people use it, the more valuable it becomes’.Footnote ²⁴⁰ Speakers of majority languages benefit from a location rent, profiting whenever an outsider learns their language and pushes its communication value up without effort on their part.Footnote ²⁴¹ Language is a valuable commodity in many former colonial powers and their settler colonies, which export educational and cultural products worldwide. Education-related exports alone were valued at $39.7 billion for the United Kingdom in 2022, $50 billion for the United States in 2024, and $33 billion for Australia in 2024. This has led some to conclude that the global dominance of English is neocolonial or imperial.Footnote ²⁴²

The flip side of the coin is that the fewer people use a language, the less value it is perceived to have. As Bourdieu points out, the dominated are often complicit in their own subjugation, as they are susceptible to market forces. Speakers of peripheral language varieties collaborate in the devaluing of their languages, when they switch to a dominant language to improve their life chances. If language shift does happen with many speakers, which causes the demise of a language, it may be considered a tragedy of the commons, where individuals acting in their own self-interest lead to the depletion of a shared resource.

Bourdieu emphasizes the relationality and conflicts among actors in a field. The fact that linguistic products on social media platforms now compete in a global linguistic market, where infrastructure and material resources are accorded to a small number of dominant languages much more than others, positions minoritized language speakers, who constitute a global majority, as a digital underclass. This is highly significant to our understanding of digital linguistic inequalities, as well as digital inequalities more generally, and of the limitations of prevailing conceptual frameworks. In theory, digital divides can be bridged; the digitally excluded can become included; digital language support can be enhanced. However, neither digital divide nor digital exclusion pay sufficient attention to relationality in the digital field: the digital underclass are unlikely to emerge as winners in a global competition for attention. If they must compete globally for content moderation resources, they are unlikely to enjoy adequate protection from harm.

Under information capitalism, knowledge and information have increasingly become commodified – transformed from public goods into marketable assets. This commodification goes beyond economic transactions, influencing how core social goods such as due process, privacy, and human rights responsibilities are managed. Consequently, these foundational principles risk being subordinated by market logics such as efficiency, profitability, and platform metrics, rather than by legal obligations or democratic accountability.

3.6 Rethinking Linguistic (In)justice

Although many linguists who write about language-based injustices are concerned with the survival of minoritized languages, it is important to note that the consequences of these injustices are not only linguistic in nature. Discrimination based on language can hurt one’s ability to secure housing and employment. Linguistic barriers to justice lead to lack of access to conflict resolution and judicial remedies. As the current study shows, gaps in digital language support and inequalities in content moderation and recommendation can affect not only opportunities to participate in the digital economy but also the safety and well-being of platform users and non-users.

Sociolinguistic approaches to rectifying linguistic injustices is predominantly framed in the discourse of language rights.Footnote ²⁴³ Advocacy in this area often sees entrenchment of rights into domestic law as the ultimate goal; some assert that fundamental language rights should become linguistic human rights (LHR).Footnote ²⁴⁴ Currently, language-related rights embedded in international and regional human rights instruments are mostly tolerance-oriented and focus on individual rights grounded in principles of non-discrimination and fair trial; a narrower set of group-based minority rights also exists.Footnote ²⁴⁵ LHRs present a more radical proposal to address linguistic inequality by pushing for the extension of rights typically enjoyed by majority language speakers, such as the right to receive basic education in one’s mother tongue and the right to use it in official contexts, to minoritized language speakers. These are promotion-oriented rights that require resourcing from the state. Despite its ambitions, advocacy for LHRs has gained limited traction. States are wary that if they supported minoritized language groups, they might also be sponsoring sub-state nationalism. When national law grants official status to minoritized languages, such law tends to emphasize political unity and have symbolic significance for language communities, but rarely translating into substantive rights with emancipatory power.Footnote ²⁴⁶ Fundamentally, the rights model positions the state as the primary actor, whether as aggressor or protector, while often overestimating its political will and economic capacity to fulfil the envisioned rights. This model has limited utility in the digital realm, where private actors have a dominating presence, as rights are primarily claimed against the state.

As an alternative to the rights approach, linguist Christopher Stroud has proposed an idea called linguistic citizenship, which encourages speaker agency and the development of political subjectivities beyond the bounds of existing power structures.Footnote ²⁴⁷ Although the emphasis on the agency of individuals is laudable, the strength of linguistic citizenship is also its weakness: by shifting the arena of contestation to spaces where community stakeholders can exercise full control of their linguistic practices, the framework does not present itself as a candidate that can be used to confront the structures that directly produce and reproduce social inequalities.Footnote ²⁴⁸ A more practical proposal was put forward by Docrat and Kaschula, which does not replace but supplement the rights approach.Footnote ²⁴⁹ Observing the limited success of implementing constitutional language rights in South Africa, Docrat and Kaschula propose applying the meaningful engagement framework, developed in the sphere of socioeconomics rights, to the implementation of language policy. This approach emphasizes effective consultation and mediation with stakeholders and requires that solutions be shaped by the views of those directly affected through democratic processes. Crucially, such engagement must take place in their own languages to ensure genuine participation.

Political philosophers have contemplated the normative question of how public institutions should handle societal multilingualism, particularly how state governments balance the demands of nation-building with language rights and accommodations for minority languages. While establishing a common language for public discourse may foster deliberative democracy, selecting a single language can also be perceived as exclusionary and unjust within a multilingual society. The dilemma is ultimately concerned with the relationship between governments and the governed. Social contract theory presumes that individuals consent to surrender some autonomy and liberty in exchange for security and a cooperative system of resource distribution. However, since social contract theories do not easily apply to private entities, this body of literature offers limited guidance on how private actors should navigate platform multilingualism. What is noteworthy for the purpose of our current discussion is that this literature seems to be moving away from the rights discourse. A volume edited by Will Kymlicka and Alan Patten suggests approaching linguistic justice not from the perspective of outcomes, such as language preservation or nation building, but procedures.Footnote ²⁵⁰ This approach aligns with recommendations from this Element: procedural safeguards that will guarantee a minimum standard of treatment across languages supported by a platform.

Scholars working in intersections between language and economics have attempted to tackle the challenge of balancing between the economic advantages of having a limited number of standardized languages and the undesirable consequences of disfranchising minoritized language communities in developing language policies. Using communication in the EU as a case study, Ginsburg and Weber propose that the selection of core languages for communication should take into account linguistic proximity between groups and between languages.Footnote ²⁵¹

In sum, the linguistic (in)justice literature has largely been developed as a response to the linguistic homogeneity imposed by modern states following 19^th century European nationalism. It has given limited consideration to new forms of linguistic injustices in the digital arena, as outlined in this Element. Nationalist governments often seek to unify a country through promoting a majority language, and may see minoritized language communities as impediments to nation-building. By contrast, corporations typically hold no animosity towards minoritized languages – they simply do not care. Their engagement with language policy is minimal and driven solely by commercial or legal necessity. As such, the linguistic inequalities documented in this Element are not the result of deliberate exclusion, but rather collateral damage inflicted as platforms trample their way to profit.

Some insights from the existing literature could potentially be transposed to the digital context, such as the emphasis on meaningful engagement, the analytical shift towards procedural justice, and prioritization methods in language policy. Overall, however, our conceptual toolkit will need to be expanded to tackle the rise of private powers in mediating public discourse, which is exercised largely through soft power. This task calls for multidisciplinary approaches grounded in the political economy of language, with a willingness to engage with government, law, and markets. Discussions in this Element invite reimagining of what linguistic injustice entails and what meaningful remedies may look like.

3.7 Conclusion

Platforms embody contradictory forces: they at once connect and isolate; they enable as well as suppress; they equalise yet also reinforce hierarchies; they democratize access while concentrating power; they promise freedom but impose control. Their technologies can be both friend and foe to marginalized language communities. With billions of users worldwide, platforms today do not only produce technologies, they also govern who may speak, to whom, and how.

This Element identifies platforms as a key site of contemporary struggles over language choice and linguistic revitalization, global linguistic competition, and language ideologies. It underscores the specific risks and harms arising from how platforms govern language-related matters, and calls for urgent intervention. Platforms like to tout technology such as AI as the catch-all solution to challenges ranging from content moderation to digital language support. However, technological breakthroughs require time, resources, and luck, and greater reliance on AI will likely diminish, rather than enhance, transparency and public accountability. Platform policies and processes can be reconfigured today, if the will is there.

On a global platform like Facebook, the majority of users participate in a language environment where content moderation efforts are heavily under-resourced compared to that available to English speakers. Platforms may not warn them when they come across disinformation or disturbing content; they may not be told what rules they need to follow; their content may be removed or violating content may remain because the content moderation system fails to understand their language; they may not be told why their post is removed or how they may challenge the decision; and their voice may not receive much of an audience when dominant voices are amplified.Footnote ²⁵² These linguistic inequalities are much more complex than what is typically captured by discussions of the digital language divide or digital language support, and have received scarce attention in the literatures of digital inequalities and linguistic injustices.

Despite positioning themselves as conduits, platforms exercise control over content and make governance decisions that shape the conditions for speech. Even though this Element foregrounds linguistic inequalities, it does not posit that linguistic equality is possible. Following Piller,Footnote ²⁵³ it sees linguistic equality not as absolute equality among languages but as the overcoming of linguistic inequalities, by distributing opportunities and resources as fairly as possible. For example, while one cannot expect every language spoken in a country to enjoy official language status, linguistic accommodation through quality interpreting can be provided to mitigate barrier to justice for those who do not speak the language of the court. The Element uses linguistic inequalities to talk about the disproportionate risk, neglect, and inability to access platform processes that some language communities but not others experience on digital platforms. It advocates for a baseline of multilingual management not drawn by the market (see Section 3.1.1), as well as the decentralization of platform governance (see Section 3.1.2), using a multi-pronged approach (as detailed in Section 3.2). These proposals aim to give marginalized language communities a genuine chance to flourish online and to reduce their vulnerability to harm. They should not have to compete for visibility in a global neoliberal economy simply to receive basic protections, due process, and opportunities to be heard.

Linguistic inequalities on platforms are a feature, not a bug, of digital capitalism. Although many of these inequalities can be mitigated to some extent, assuming sufficient goodwill, it is difficult to find comfort in the reality that most of our online interactions, whether with friends and family or in relation to government communication and public debate, now take place on private platforms driven solely by profit; that we practice our civil and political rights through private actors who bear no accountable human rights responsibilities; that digital infrastructure and governance frameworks developed in the Global North are transplanted to the Global South to perpetuate economic dominance over the global majority. This Element has shown that language offers a powerful lens through which we can visibilize these deeper systemic problems, which are pertinent to our fundamental rights and freedoms. Language can provide a point of intervention, given its close ties with knowledge, participation, representation, and global redistributive justice. While certain impacts of linguistic inequalities can be mitigated through improving platform policies and processes, tackling these larger problems requires innovations in international law, the imposition of effective legal constraints on private power, a model of digital governance that gives voice to different stakeholders, and alternative visions of a digital public space.