Generative AI and Multilingual Digital Democracies

Today, digital democracy is becoming an important complement – and some would say even an alternative – to offline forms of political participation. Increasingly more people use social media as a news source, debates regularly unfold in comment sections, and protests often start in encrypted chats. The digitalization of public life has upended traditional ways of political engagement.

Digital platforms have brought clear democratic advantages: nowadays, information is readily available and communities can self-organize more easily than ever before. In the past, organizing political movements was a monumental task, as the only way to engage in politics were in-person interactions. Dissent, even if widespread, would often dissipate because it was not channeled into organized initiatives. Modern digital networks, on the other hand, allow for political opinions to quickly grow into organized movements, frequently capable of unprecedented democratic successes. The Arab Spring, an online-organized, student-led uprising that toppled multiple authoritarian regimes in North Africa and the Middle East in the 2010s, is a prime example of the transformative thrust of digital democracies.

At the same time, however, digital platforms have introduced democracy to new challenges, especially regarding participation, inclusion…and language! Digital democracies can only be truly democratic if everyone has equal rights and access to its spaces, and language should enable this process, not inhibit it. Yet when online spaces turn monolingual, typically adopting the language of the majority, linguistic minorities may end at the margins of political discourse. It is therefore imperative for digital democracies that aspire to be such to address the question of linguistic inclusion.

In this context, generative AI offers a positive answer. By enabling real-time multilingual interaction, it can help turn linguistic diversity from a barrier to communication into a democratic asset. This way, not only multiple people, but also multiple languages can coexist.

The democratic challenge of multilingualism

Historically, modern democracies assumed monolingualism — i.e. having one single language for exchange—as a precondition for their existence, while they saw multilingualism as an obstacle to democracy. This understanding of language and democracy viewed language as a mere instrument to communication. Language, nevertheless, transcends communication. It defines our personal identity and sorts individuals in categories of “self” and “others”, of “us” and “them”, depending on the use people make of it. Language can promote equality and inequality, and has the potential to create powerful hierarchies that can either include or exclude people. Linguistic minorities, migrant voices, and everyone not fluent in the dominant language can have a much harder time communicating their stances and phrasing compelling arguments. As a result, their positions might be overshadowed by that of the majority. Beyond mere fluency, individuals might also resist speaking a language because they might not adhere with the cultural and linguistic norms of the native speakers. 

Online political spaces are not exempt from these dynamics; if anything, the risk is higher. Digital platforms theoretically support diverse languages, but because digital political space spans across national borders, in practice, the majority of content and interaction takes place in lingua franca (e.g. English). This marginalizes minority and indigenous languages. Hence, the value of multilingualism: with research showing that people are more likely to participate in politics if they can use their preferred language, multilingualism is a golden chance to make our democracies more participatory and inclusive.

How generative AI can help

Real-time communication in different languages is expensive, detractors of multilingualism might say. Historically, this is a compelling argument, since humanity long lacked practical means to realize simultaneous multilingual communication except the individual linguistic talent of its members. It is here, however, that the last frontier of technological development, generative AI challenges, comes to the rescue, calling into question the limits of the unprecedented. Generative AI, in fact, can not only enable real-time multilingual dialogue in politics, but also do so at a fraction of the cost.

To begin with, generative AI in the form of Neural Machine Translation (NMT) systems, such as DeepL or Meta’s No Language Left Behind, enables real-time, relatively high-quality translation. Combining these with Large Language Models (LLMs), another generative AI technology, can further enhance the quality and contextual accuracy of translation. This approach allows citizens to contribute and engage in politics using their native languages, removing the need for a lingua franca and thereby averting language-based discrimination. In the European context, initiatives such as the Decidim, a free and open-source web-based digital platform for citizen participation, could integrate generative AI-powered translation tools to enable multilingual dialogue among Europeans.

Along similar lines, generative AI can be used to summarize and synthesize discussion written in multiple languages. LLMs have proven to be exceptionally strong tools for tasks of summarization and comprehension enhancement. Furthermore, the large datasets they employ make them fluent in a multitude of languages, including minority tongues and dialects. Political participants can use LLMs to break down inputs and political discourse in digestible content while reading it in the language they are most comfortable with. This way we ensure ideas are judged by their content and not by linguistic fluency, and dominant languages are not privileged over less widely spoken ones. Reverse the process and generative AI still rises to the task. LLMs can help participants compose clearer, more formal and persuasive messages in non-native languages, thereby reducing linguistic disadvantage.

Content and logical coherence is surely foundational to effective communication, but we often tend to overlook the role of finer shades of meaning and cultural nuances—the undertones of our languages. These devices are often deciding factors in determining how compelling arguments are, especially in heated political forums, and happen to be the part of language non-native speakers struggle with picking up the most. But by now LLMs have been trained to detect idioms, tone, and cultural references, too. They can suggest to non-native users wordings and phrasings that foster inter-lingual understanding and mediation.

Lastly, LLMs can support institutions to scrutinize multilingual inputs online and identify which minority languages are underrepresented compared to dominant ones. This can inform policy makers on where to increase linguistic content to ensure that speakers of different languages have access to equitable amounts of information and political material.

And the risks?

So far so good, you may say. But deploying generative AI in political communication is a slipperier slope than it may appear. It must come with some caveats. To begin with, instilling human biases in AI models is a tangible risk. If fed biased inputs, AI reproduces the bias, amplifying and entrenching it in its functional algorithms. From the start, multilingual inputs should not only converge as much as possible on definitions, but also employ similar tone and style, so that translation models can perform at similar standards across all languages.

Algorithm training should be bias-free as much as practically possible, too. Having larger databases to train algorithms is the first step to reducing the risk of ultimate bias. In the case of smaller languages, the political content available online might not be enough to train algorithms while avoiding bias. Human translators have a responsibility here. They should work to expand databases as much as possible by defining fundamental political notions and translating existing political material from larger to smaller languages. Native minority speakers should also be encouraged to write blog posts, Wikipedia articles, or social media posts as a way to increase online political content in smaller languages. In this case, however, creating boards consisting of language and communication experts is crucial to approve the content. Only in this way can AI algorithms guarantee effective multilingual communication and normalize political communication in smaller languages.

Lastly, humans should always keep a careful eye over the entire process, from algorithm training to any potential side effect of using generative AI in facilitating multilingual political dialogue. AI should assist, not replace human facilitators, in alignment with the EU AI Act principles. Democratic engagement is something precious that not all societies can enjoy: we should welcome multilingualism only as long as it contributes to a more democratic and linguistically inclusive society.