Constitutional-Legal Aspect of Creating Large Language Models: the Problem of Digital Inequality and Linguistic Discrimination

Introduction

Large Language Models (LLM) are generative artificial intelligence models used in natural language processing technology (NLP). They allow a computer to efficiently process text data, demonstrating the ability to “understand” text at a deep level, create coherent and contextually relevant responses to queries, translate from one language to another, and generate texts that meet certain stylistic and content requirements (Glauner, 2024). Examples of large language models include BERT¹, GPT-3², and related digital products such as Google Assistant and ChatGPT.

Large language models are trained on vast arrays of linguistic data, including structured linguistic corpora: databases containing a variety of texts (books, text transcriptions, translations, etc.) and audio files (audiobooks, broadcast recordings, podcasts, and other audio content). The structure and representativeness of such data, their volume and format determine the learning process and the accuracy of understanding the context (Ilin, 2024), while defects³ or insufficient data can lead to incorrect model functioning and generally hinder the technology development (Hacker, 2021). Thus, the possibility of creating a high-quality language model directly depends on the volume, representativeness and other qualitative characteristics of the training data for a particular language.

At the same time, the levels of digitalization of languages – the volume of existing linguistic corpora and the data for their creation – differ significantly. For some languages or dialects, data may be extremely limited or non-existent. This hinders the development of accurate and effective language models, slowing down their digital development and limiting their integration into modern technologies. For example, if a data set is not comprehensive enough and does not cover all variants of a particular language, the model may process incoming requests incorrectly or inaccurately, and in some cases may not function at all. Differences in pronunciation, vocabulary, and grammar can lead to errors in recognizing and analyzing text or speech, and reduce the quality of results.

The inability to create a full-fledged language model for particular languages or dialects makes unavailable many digital products for speakers of these languages or significantly worsens the quality of their functioning compared to how the same technologies function for speakers of languages with a high level of digitalization. As a result, digital inequality arises, when access to modern technologies is unevenly distributed among different linguistic communities, which, in turn, increases the risk of discrimination.

The article aims to analyze the constitutional and legal aspects of creating large language models in the context of digital inequality and linguistic discrimination. To achieve this goal, we will investigate how digital inequality affects constitutional human rights, as well as analyze the risks of linguistic discrimination associated with the creation of large language models.

The article contains the main results of the corresponding research, as well as directions for further study. The research paper is divided into three thematic sections, supplemented by an introduction and conclusion. The first section analyzes the problem of digital inequality given the different levels of digitalization of languages – the volume and representativeness of linguistic data. The second section examines linguistic discrimination as a potential form of digital inequality, with an emphasis on unequal access to natural language processing (NLP). In the third section, the problem of linguistic discrimination is conceptualized in relation to other human rights and in the context of the digital technologies development.

1. Digitalization of languages as a source of digital inequality: a technical and legal analysis

Digital inequality is a form of social inequality characterized by unequal access to information technologies and by varying levels of skills in using them among individuals and social groups (Mushakov, 2022). This phenomenon encompasses a wide range of factors, including differences in technical equipment, access to Internet, digital literacy, and educational opportunities, which in turn leads to social and economic division (Rogers, 2016). The need to address gaps in access to digital technologies in order to achieve a more equal and inclusive society has been repeatedly highlighted both at the national⁴ and international⁵ levels.

In the context of creating large language models, the problem of digital inequality manifests itself in the limited ability of speakers of low-digitalization languages to use digital products in their language. This leads to unequal access of individuals or social groups to natural language processing (NLP) technologies. As a result, there may appear restrictions on access to information, education, and social services for native speakers of such languages. For example, the ability to understand text contextually and generate appropriate responses contributes to the active use of this technology in areas such as education and healthcare (Jiang et al., 2023; Sohail & Zhang, 2024). The lack of support for particular languages in these areas may negatively affect the exercising of the corresponding constitutionally guaranteed rights: the right to access education⁶ and medical care⁷. This may limit the availability and quality of these services. In this regard, it seems logical to consider the problem of digital inequality from the viewpoint of constitutional and legal relations, i.e. the concepts of “equality” and prohibition of discrimination⁸.

We can also agree with some researchers that constitutional norms ensuring equality before the law and access to services should take into account and eliminate inequality in access to digital resources, as this directly affects the ability of citizens to exercise their rights and freedoms in the digital era (Mushakov, 2022).

To create an effective language model, a set of training data is needed that must meet criteria such as volume, representativeness⁹, and other qualitative characteristics. These parameters directly depend on the level of digitalization of a particular language, since the higher the degree of digitalization, the more diverse and high-quality data can be used to train the model.

Digitalization of a language in a broad sense is the transformation of data into appropriate electronic linguistic corpora. For this purpose, text data (for example, files, transcriptions, abstracts), speech data (for example, audio recordings, phonetic and intonation abstracts) and multimodal data (i.e. data combining several types, for example, video and text, images and text, etc.) are used (Dash & Arulmozi, 2018). It should be noted that this process not only contributes to technological development and digital transformation of society, but also plays an important role in preserving national and cultural identity (Kelli et al., 2016). For example, digitalization of minority languages can significantly contribute to the preservation of the cultural heritage of small nations.

Despite the importance of digitalization for technological progress and the high social significance of this process, the level of digitalization of languages and their dialects remains uneven. There are economic, technical, and legal factors that limit or hinder the digitalization of languages.

The economic factors are related to the fact that languages have different economic potential (Alarcón, 2022; Monteith & Sung, 2023). Hence, digitalization requires significant resources, including time, finance, etc. In this regard, the development of linguistic corpora for some languages may be economically unfeasible. The technical factors are directly related to creating linguistic corpora. Such factors may include errors in data collection, flaws in the corpora design and limitations of existing datasets, errors in metadata, etc. (Solovyev & Akhtyamova, 2019; Doğruöz et al., 2023; Li et al., 2024). The legal factors are related to the presence of regulatory restrictions on access to training data and the need to comply with the relevant legal regime when using them for training.

In previous works, the author discussed in detail the issues of regulating access to training data (Ilin, 2024), as well as compliance with their legal regimes, such as the personal data regime (Ilin, 2020) and the intellectual property regime (Ilin, 2022; Ilin & Kelli, 2019, 2024). The central topic of those studies was the conflict between equally protected human rights when using training data, such as the right to non-discrimination¹⁰ and the right to privacy, to protection of personal and family secrets¹¹. Overcoming this problem is necessary both at the conceptual level (removing regulatory barriers to data access, taking into account the balance of private and public interests) and in practical terms (creating conditions for the dissemination and exchange of linguistic data, for example, by developing an institution of reusing the data accumulated in government information systems (hereinafter referred to as GIS) or involving higher educational institutions to create linguistic corpora and digitalize the language).

According to the analytical report of the Accounts Chamber of the Russian Federation¹², by 2020, more than 800 federal state information systems were operating in Russia, providing data exchange between government agencies in various areas of public life. These systems cover a wide range of information, including statistics, as well as information on healthcare, education, and other key sectors. In this context, the use of GIS data to create linguistic corpora seems to be particularly promising. Despite the varying levels of development of these systems, one may expect that the data collected in them will have the necessary qualitative characteristics, and their diversity can provide the necessary representativeness and volume (Ilin, 2024). However, given the risks associated with legal restrictions on the use of data, their reuse should be carried out in accordance with uniform principles and regulations. These should include legislative standards and control mechanisms that take into account the specifics of each data type and its compliance with the purposes of its initial collection.

Another possible solution to the problem of access and lack of linguistic data is to involve higher education institutions to creating and subsequently disseminating linguistic corpora. The participation of universities in language digitalization can also be justified by the social significance of this activity. As an example of successful cooperation between commercial organizations and higher education institutions in the field of natural language processing, we can mention the joint academic program of the Center for Speech Technologies Group of Companies with the ITMO National Research University (Ilin & Dedova, 2019).

However, although this solves the problem of creating linguistic corpora, the issue of their further distribution remains open. For example, for various reasons, a university may not be interested in further dissemination of the linguistic corpus or may not have the necessary resources for this and, accordingly, may not distribute it. If a university operates as an entrepreneurial university and commercializes its results, for example, via a spin-off company, one may also question the possibility to employ the doctrine of the free use of works¹³ when processing linguistic data. All these questions require further careful analysis, both from a legal and other points of view.

2. Linguistic discrimination as a form of digital inequality

Since, in relation to large language models development, digital inequality leads to unequal access of individuals or social groups to natural language processing (NLP) technologies (the inability to fully use this technology in their language), the problem of digital inequality should primarily be considered in the context of linguistic discrimination.

The problem of discrimination by artificial intelligence systems, although not new, remains relevant today. The development and active implementation of artificial intelligence systems in various spheres of life opens up new areas for discussion of this problem. Examples include discrimination by artificial intelligence systems in the field of labor relations (Morin, 2024), the impact of profiling¹⁴ on human dignity (Orwat, 2024), the potential impact of artificial intelligence on discrimination based on ethnicity, religion and gender (Ozkul, 2024), etc.

In addition, with the increasing autonomy of artificial intelligence systems and the development of generative models, discrimination begins to take on an implicit character, which allows classifying its manifestations into direct and indirect ones. For example, unlike the obvious cases of discrimination observed in predictive crime analytics systems such as those based on PredPol¹⁵ and COMPAS¹⁶ algorithms, discrimination in generative artificial intelligence systems may be less apparent. For example, these systems may preferentially create images of white men in response to repeated requests for examples of people employed in important professions, potentially leading to cumulative discriminatory effects (Hacker et al., 2024). In such cases, discrimination becomes difficult to detect, as it may not be explicit or obvious, but nevertheless has a significant impact on the representation and perception of various groups in society.

The National Strategy for the Development of Artificial Intelligence up to 2030¹⁷ (hereinafter referred to as the Strategy) emphasizes that the protection of human rights and freedoms is one of the main principles of the development and use of artificial intelligence technology¹⁸, while “non-discrimination” is highlighted as one of the main principles of the development of legal regulation of public relations in the sphere of development and use of artificial intelligence technology¹⁹.

Article 2 of the Universal Declaration of Human Rights (1948)²⁰ prohibits discrimination, including on the basis of language. A similar provision is contained in Article 1 (3) of the UN Charter²¹, and is also reflected in paragraph 2 of Article 19 of the Constitution of the Russian Federation, according to which the state guarantees equality of human and civil rights and freedoms, regardless of language.

In the field of language discrimination, several key aspects are identified, related to its recognition, legal protection and public perception. One of the main problems is the lack of recognition of linguistic discrimination at the international level. For example, discrimination based on voice often goes unnoticed (Baugh, 2023), which can be critical when interacting with speech and voice recognition technology and related digital products: interactive response systems and voice assistants.

The UN Human Rights Committee²² has repeatedly addressed the issue of linguistic discrimination, but its judicial practice is underdeveloped and does not provide reliable protection for linguistic minorities (Möller, 2011). The regulatory framework at various levels also often does not take into account all the nuances of linguistic discrimination. Legislation at the international, regional and national levels, as a rule, does not provide sufficient protection for the rights of linguistic minorities, which leads to gaps in the legal protection of crime victims (Chilingaryan et al., 2020).

Discrimination based on language can be defined as any unjustified distinction or restriction that weakens or excludes the possibility of exercising rights enshrined in international or national regulations based on language affiliation. At the same time, it should be added that states also have positive obligations to protect and promote linguistic rights as part of their obligation to respect human rights²³. Therefore, in the context of creating large language models, it seems necessary to expand the definition of linguistic discrimination to include actions aimed at hindering the preservation or development of minority languages. The essence of the first part of the definition is that linguistic discrimination occurs when a person experiences worse treatment than others in a similar situation due to insufficient or complete lack of proficiency in the official language established in a given state or region. The second part refers to a deeper aspect of this problem – the states’ fulfillment of legal obligations, established by international conventions and national legislation, to protect and promote minority languages. This said, it should be noted that the expansion of the linguistic discrimination concept is more likely to reflect the perspective sought by judicial practice and scientific discussion, rather than the current perception of the problem by law enforcers and lawyers.

3. Qualification problem and criteria for assessing linguistic discrimination

Ambiguity in the definition of linguistic discrimination makes law enforcement difficult and raises questions about the criteria used in assessing these situations. As noted earlier, linguistic discrimination occurs when people are treated differently because of their language proficiency or accent, which often leads to limited access to opportunities and rights (Mironova, 2019). However, linguistic discrimination is a multifaceted problem that differs from other forms of discrimination, such as racial or religious, and depends on various factors. An analysis of existing practice allows us to identify a number of key factors for determining linguistic discrimination. The first factor is the number of native speakers: the level of discrimination is often determined by the prevalence of language in society. For example, in Cameroon, the English-speaking minority faces systemic discrimination due to its small size compared to the French-speaking majority (Donard, 2023).

Another important factor is the ability of the state to support multilingualism. The more actively the state creates conditions for learning and using multiple languages, the lower the likelihood of linguistic discrimination. For example, research shows that support for multilingualism in educational institutions helps to reduce discrimination based on language (Page, 2023).

The use of minority languages in public life is also of great importance. When these languages do not receive institutional support, their speakers are often marginalized, which reinforces existing social inequalities.

In addition, it should be borne in mind that linguistic discrimination may overlap with other forms of discrimination, such as racial, religious, or ethnic. In such cases, people are subjected to complex forms of discrimination, which significantly exacerbates the problem (Drożdżowicz & Peled, 2024). In order to illustrate the complexity of the problem, let us briefly consider some of these intersections.

Failure to provide equal access to services in the mother tongue may violate the right to equality, creating barriers that hinder full participation in social life²⁴. These barriers, for example, can affect the right to education²⁵ by limiting access to educational resources and materials in the mother tongue, which can reduce the quality of education and limit educational opportunities.

In addition, linguistic discrimination affects the right to freedom of expression²⁶. People should be able to express their opinions freely in the language they prefer, and restrictions on this may be seen as a violation of this fundamental right. Linguistic discrimination also affects cultural rights, as language is a key element of cultural identity and expression. Restricting the use of a minority language in cultural and social contexts can undermine the cultural rights of these communities and their ability to preserve and develop their cultural identity.

Access to justice can also be hampered by language barriers, as the need to understand and participate in court proceedings in one’s native language is critical to ensuring fair justice²⁷. Language barriers may prevent the correct understanding of charges, court procedure, or legal decisions, which can lead to unfair outcomes.

Thus, although it is possible to identify factors for assessing linguistic discrimination, the legal qualification of such cases in the context of digital technologies causes certain difficulties. For example, it is necessary to find out whether errors in the language model can be considered a manifestation of discrimination. Such errors are often difficult to detect, as discrimination may be hidden, which makes it less obvious for analysis. Discrimination in models can be the result of algorithmic or human bias. Algorithmic bias occurs due to limitations or distortions in the data on which the model is trained, whereas human bias can manifest itself in the developing and configuring algorithms (Kharitonova et al., 2021). Both forms of bias can not only affect the accuracy and fairness of decisions, but also maintain or exacerbate existing social inequalities, ultimately leading to discrimination. The distinction between errors and discrimination requires in-depth analysis, as errors may be accidental or may result from systemic biases. It is important to understand how bias, both algorithmic and human, affects the decision-making process and how it is integrated into algorithms and models. This understanding is necessary to develop more equitable and inclusive digital systems.

Conclusions

This article aims to analyze the constitutional-legal aspects of creating large language models in the context of digital inequality and linguistic discrimination. The study found that digital inequality in the context of large language models is due to the uneven digitalization of languages and manifests itself in limited access to natural language processing technologies. Such unequal access can negatively affect the implementation of constitutionally guaranteed rights and requires consideration through the prism of such concepts as “equality” and prohibition of discrimination. In turn, the identification and legal qualification of linguistic discrimination when creating large language models is a difficult task, since biases in models can be hidden and have a cumulative discriminatory effect. Discrimination may be caused by both algorithmic and human bias. Algorithmic bias occurs due to limitations or distortions in the data on which the model is trained, while human bias can manifest itself in developing and configuring algorithms. Distinguishing between these categories and assessing their impact on decision-making are becoming important areas for future research aimed at developing mechanisms to ensure equal access to digital technologies and the protection of language rights.