AI, Epistemic Loss and Troubled Visions for Legal Pluralism

I need to talk to a lot of artists and creative practitioners for my PhD. Many of them are based in Bangladesh; some are in other countries. They come with diverse socio-economic backgrounds and training, and their artistic practices, of course, vary a lot. We usually ask them whether and how they use generative AI in their works and what their concerns are around these technologies.

For one such project, our team met Jamal, a 29-year-old rickshaw painter, at his usual working place in Hatirpul Bazar, Dhaka ¹⁾. For readers unfamiliar with the context: rickshaw is a three-wheeler vehicle, very much common in urban Bangladesh and in other parts of South Asia, and rickshaw painting involves colourful oil-based illustrations on the vehicles’ bodies and roofs which can be easily considered as a unique, vernacular visual artistic practice in the region.

Anyway, it is something that Jamal told us which stayed with me for long:

“If someone uploads my artwork to AI, no matter how much I protest, they will continue because they think it’s right. But using our artwork without asking us is exploitation and this has now become a business. If you want original work, commission us for it. Why steal? They make a bit of money this way, but rickshaw pullers and craftsmen lose out.”²⁾

I realised he was talking about copyright laws, but he could not articulate it. Or, perhaps, our legal languages could not reach a so-called “illiterate” painter in one of the margins of the world, who wants to protect his craft from Big Tech in Silicon Valley.

Jamal perhaps did not know the names of those companies. He also did not know whether his country has some laws in place to protect him in situations like this. And, for those who hold power – either those companies or the governments – his concerns rarely count.

This was one of those moments that help you grow perspectives as a scholar. I re-discovered what writer Arundhati Roy, whose work I started to admire growing up, meant when she said, “there’s really no such thing as the ‘voiceless’. There are only the deliberately silenced, or the preferably unheard.”³⁾ Epistemic Injustice. It explains a lot.

Whose knowledge? Whose Ethics?

According to philosopher Miranda Fricker, epistemic injustice occurs when someone is specifically wronged in their practice as a knower.⁴⁾ She theorises two major types of epistemic injustice: hermeneutical and testimonial. Hermeneutical injustice happens when collective interpretive frameworks in a society are inadequate for comprehending marginalised realities of individuals, whereas testimonial injustice entails unjustly undermining a speaker’s credibility because of systemic bias towards them. In addition to these two “discriminatory” types of injustice⁵⁾, distributive epistemic injustice can arise from the unequal distribution of epistemic goods (for example, resources and capacity to produce knowledge) in society⁶⁾.

In the current trajectory of global AI development, these critical conversations have not been at the forefront with the right pace. Whether or how the historically marginalised communities, particularly in the Global South, can understand what AI is and how it affects them? Does their knowledge matter in AI? Can their experience shape ethics around this technology? How can we ensure epistemic justice when we design sociotechnical interventions with and for these communities? I have been looking for these answers in my research over the years.

These are indeed difficult questions, and solutions are hardly straightforward.

Researchers are now increasingly highlighting that generative AI can perpetuate epistemic injustice by misrepresenting, marginalising, or undermining people’s knowledge making capabilities and their ability to contribute as “knowers”⁷⁾. Despite recognising cultural and linguistic inequalities in AI, these debates largely rely on Western social and political philosophy rather than developing its account from non-Western epistemological traditions or community-defined understandings of knowledge and justice.

This is a common pattern that we have identified in contemporary AI Ethics literature. In a study, we had analysed nearly 6,000 publications in the field and revealed how Global South voices are often sidelined, perpetuating epistemic inequality in the field. Therefore, by continuing to marginalise the global majority populations as “others”, the project of AI Ethics falls short of its promise to be universally beneficial.⁸⁾

Thinking beyond the data problem in AI

In practice, epistemic stakes for the Global South in AI are quite high. Earlier this year, we wrote for Patterns and tried to highlight these issues. We defined “epistemic loss” in the context of AI image generation in our piece: “the erosion of worldviews, ontologies, and ways of knowing under the pressure of AI’s largely Western-centric assumptions.”⁹⁾

We found that large datasets used for training AI image models such as LAION-5B had inherited geopolitical, linguistic, and platform biases from their web sources, disproportionately representing Global North contexts. This inequality was also intensified by the dominance of English in image captions and ALT text, poor indexing of underrepresented languages and cultures, and automated filtering methods that reproduce existing biases.¹⁰⁾

It is a valid argument that data around non-Western imagery and artistic practices on the internet are scarce and that shapes their representation in AI. However, we argue that such epistemic biases are extensions of a long-standing historical problem. AI continues to remain Western value-laden by justifying colonial hegemony of knowledge and domination like its predecessor concept of “intelligence”.¹¹⁾

By now, we have spent numerous hours ethnographically engaging with Bangladeshi artists and creatives, and they keep sharing with us how vulnerable they feel, finding their cultural and artistic imaginations constrained by the AI tools they use. Western art remains the normative standard for AI, and any art from non-Western contexts stays exotic, backward, or inferior, gaining recognition primarily when aligned with or shaped by Western aesthetic traditions.¹²⁾

For language models, the concerns are similar, if not more intense.

Deepak Varuvel Dennison, a fellow PhD researcher at Cornell, writes for The Guardian that 97% of the world’s languages are considered “low-resource” in computing, despite many having millions of speakers and rich cultural and linguistic traditions. For example, English is vastly overrepresented in major training sources such as Common Crawl, while widely spoken languages like Hindi and Tamil occupy only a tiny fraction of available data (0.2 and 0.4% respectively) and such patterns are reflected in AI’s linguistic capabilities.¹³⁾

What we are trying to argue in our research is that language, creativity and imaginations are conditions that shape what it means to be human. Beneath these practices, remains vast, limitless human knowledge in personal, social and material forms that helps us in world-making. Now AI is intervening in such knowledge practices – not only for artists and writers, but for ordinary people in their everyday lives. Every individual, no matter where they are geographically located, should receive fair treatment as “knowers” when they interact with these tools. The way forward is to increasingly recognise plural epistemologies in AI system developments: social, collective and indigenous knowledge systems and situated expertise of marginalised groups.

Why does it matter for legal pluralism?

Now, as this piece may reach legal professionals and experts in different parts of the world, it needs to be clarified that we are approaching justice here as a social, cultural and political practice in AI. But why not legally? Because, I do not have the correct expertise for that, or perhaps more accurately, I subscribe to what Gayatri Spivak has suggested, “it is Just that there be Law, but Law is not Justice”.¹⁴⁾ It was more important for her whether the legal system can hear and recognise the worlds from which their claims arise.

In her famous essay “Can the subaltern speak?”¹⁵⁾, she talked about the political context of abolition of sati (widow immolation) in 1829 in colonial India. The British wanted to pass a law to “rescue” Indian women from oppressive Hindu tradition, as goes Spivak’s famous formulation, “white men are saving brown women from brown men”. She argued that colonial law tried to determine what justice could possibly mean for a widow, while patriarchal tradition claimed to speak on her behalf, and in between the two, she lost her epistemic subjectivity.

Spivak is one of those scholars whom I would put before Fricker in a discussion of epistemic justice. She denied being apolitical while talking about marginalisation of knowledge of certain groups, including women, in postcolonial societies, focusing on the “harms” caused by epistemic regimes, rather than merely the “wrongs” done to them.

Gayatri Spivak’s works point out that the pursuit of justice sometimes requires confronting the epistemic limits of the legal categories themselves¹⁶⁾. This is an ideal entry point to discuss how and why epistemic implications of AI become important concerns for legal pluralism, and that is the point this article slowly builds up.

Evidence from the ground

More recently, Salwa Tabassum Hoque’s work at New York University highlights epistemic implications of legal language and practices, particularly in the digital and AI age¹⁷⁾. She conducts 14 months of ethnographic fieldwork in Bangladesh to understand what happens to legal knowledge as disputes travel from communities to rural Islamic/customary non-state courts (shalish), and then to state courts, and eventually into digital databases: essentially, it distorts and erases alternate frameworks of law and rural women’s unique socio-legal experiences.

Hoque particularly studies uthai nawa (“taking” or “picking up”) cases. The Bangla expression ostensibly signifies coercive abduction. However, her observations show that rural communities may also use it as coded language for consensual elopement in a social context where women cannot easily communicate their romantic and sexual agency. Shalish actors can interpret this contextual meaning and address issues such as abandonment, economic precarity, compensation, family reconciliation, and social stigma. Formal Supreme Court proceedings, by contrast, tend to translate these disputes into the formal binary of abduction versus consent, stripping away much of these situated knowledges.

Importantly, Hoque finds that digital legal databases in prominent research software such as HeinOnline, Manuport, Westlaw Next, and LexisNexis increasingly constitute a new form of digital archive through which legal knowledge is preserved and accessed, marginalise and exclude such plural, alternate and community-centric legal systems, and eventually they supply data for training AI systems, meaning that their existing biases can shape subsequent automated outputs. She introduces the concept of “neocolonial digitality” to explain such phenomena and draws attention to how historical and emerging structures of power intersect within digital systems, with particularly consequential effects for marginalised communities in the Global South.

I found these accounts fascinating. It helped me have an awakening that it was me whose knowing was at fault, not Jamal’s. I should not have attempted to restrict his opinion of how “exploitation have become a business” to be coded in mere copyright laws. It was, albeit, more than that, perhaps closer to the idea of justice to rickshaw painters like him.

Concluding notes

The aim of this article was not to simplify, but rather to establish that introducing AI in legal contexts is complicated, since epistemic injustice is entangled with the very design of the technology, its ethics and the socio-political contexts of its possible applications. How AI will change the practice of law through modelling reasons, generating texts, developing arguments for and against a case, predicting and explaining legal outcomes – I leave that for technologists and legal professionals to evaluate for themselves.

But working with the ethics of this technology named AI, a knowledge machine, but perhaps broken, I can suggest a simple question for them to start with: when we are depending on AI to make or understand laws, whom do we want to serve justice to with such laws in the first place?

A machine that erases knowledge may produce laws but cannot provide justice.