Deepak Varuvel Dennison is a PhD student at Cornell University, US. His research explores responsible AI. He writes in Aeon of- 'Holes in the web- Huge swathes of human knowledge are missing from the internet. By definition, generative AI is shockingly ignorant too.'
AI is already proving very useful. It is likely that your personal AI assistant will learn more about you and keep trawling for information useful to you. Thus, we might feel it is getting smarter and more helpful. But it will have the same limitations as any other sort of Intelligence.
Deepak begins his essay by mentioning a tumour on the tongue of his father. I too had a small growth on my tongue. It was removed and a biopsy was performed. Thankfully, the thing wasn't cancerous. Deepak's father is a Tamilian just like me. His daughter, a Doctor, wanted him to undergo surgery but he put his faith in a traditional herbal remedy. His tumour disappeared. He was vindicated. Deepak had initially sided with his sister on the basis of advise received from an AI. Had he asked 'Copilot' the following question 'I am Tamilian. I want a traditional remedy for a tumour on my tongue' the AI would have mentioned specific, Government funded, 'Siddha' & 'Ayurvedic' hospitals which provide this sort of treatment. But the remedies are evidence-based and complementary to allopathic, Western, science. In other words, you can have the best of both worlds. All you have to do is be more specific about the question you ask Copilot. Indeed, the more you use a particular Assistant, even this won't be necessary. It will automatically provide you with 'tailor-made' solutions....recently I’ve been wondering if I was too quick to dismiss my parents’ trust in traditional knowledge,
Stop wondering. You were too quick. Incidentally, the 'traditional' remedy which Deepak's father used may not be anything very ancient at all. Still, if they date back to the Thirties, rather than the Nineties, we are entitled to think of them as part of our ancient indigenous knowledge system.
while easily accepting the authority of digitally dominant sources.
Deepak could have written this 100 years ago when there were educated young people in Tamil Nadu who felt their parent's faith in traditional medicine was based on superstition and the lack of a rational, scientific, world view.
I find it hard to believe my dad’s herbal concoctions worked, but I have also since come to realise that the seemingly all-knowing internet I so readily trusted contains huge gaps – and in a world of AI, it’s about to get worse.
Nobody thinks the internet is all knowing. AI is getting better and people who use it are also learning when to rely upon it and when to ignore its suggestions.
The irony isn’t lost on me that this dilemma has emerged through my research at a university in the United States, in a setting removed from my childhood and the very context where traditional practices were part of daily life.
You don't need to do a PhD to know that AI is good at some things and bad at other things. What is remarkable is how quickly it has improved.
At Cornell University in New York, I study what it takes to design responsible AI systems.
This is a waste of time. Responsibility is a function of general purpose intelligence. Increase the latter and the former can be improved- e.g. by putting in more 'feedback loops' for fact-checking and 'regret minimizing' evaluation.
My work has been revealing to me how the digital world reflects profound power imbalances in knowledge, and how this is amplified by generative AI (GenAI).
Power imbalances in knowledge have existed for thousands of years. Those with more useful types of knowledge have demographically or otherwise displaced those with less useful knowledge. The irony here is that Deepak went to a more advanced country to gain more useful types of knowledge not to whine about how it is totes unfair that his own part of the world fell behind Europe some centuries ago.
For many people, GenAI is becoming their primary way to learn about the world.
But Deepak has been attending prestigious Colleges so as to learn from the top Professors and Researchers located there. Why is he whining about AI?
A large-scale study published in September 2025, analysing how people have been using ChatGPT since its launch in November 2022, revealed that around half the queries were for practical guidance, or to seek information.
Are they getting it? In some areas- yes. In other areas- no. They quickly learn to focus only on the former not the latter.
These systems may appear neutral, but they are far from it.
They wouldn't be useful if they were 'neutral'. We want utility, not neutrality.
The most popular models privilege dominant epistemologies (typically Western and institutional)
which are successful
while marginalising alternative ways of knowing,
which are shit
especially those encoded in oral traditions, embodied practice and the languages considered ‘low-resource’ in the computing world, such as Hindi or Swahili, both spoken by hundreds of millions.
But people who speak those languages save up money so their kids get to learn English and Computer Programming languages.
By amplifying these hierarchies, GenAI risks contributing to the erasure of systems of understanding that have evolved over centuries,
which were already being erased. Deepak didn't move to America to learn the traditional wisdom of the indigenous people. He went there to learn about the latest technological advances. Sadly, he was too stupid to contribute to that so he is whining about being bleck instead.
disconnecting future generations from vast bodies of insights and wisdom that were never encoded yet remain essential to human ways of knowing.
They aren't essential at all.
What’s at stake then isn’t just representation – it’s the resilience and diversity of knowledge itself.
Why does Deepak not train in Siddha or Ayurveda? Why is he studying at Cornell? Is it because all the Professors there are medicine-men from the First Nations?
To understand why this matters, we must first recognise that languages serve as vessels for knowledge – they are not merely communication tools, but repositories of specialised understanding.
English is the language of English people or those of English descent. Tamils like me and Deepak may be able to communicate in English but we don't have 'specialised understanding' of it. That is the reason we are unable to reproduce the magic of Wayland the Smith.
Each language carries entire worlds of human experience and insight developed over centuries: the rituals and customs that shape communities, distinctive ways of seeing beauty and creating art, deep familiarity with specific landscapes and natural systems, spiritual and philosophical worldviews, subtle vocabularies for inner experiences, specialised expertise in various fields, frameworks for organising society and justice, collective memories and historical narratives, healing traditions, and intricate social bonds.
Thus it is folly to think a Tamil could ever really understand English or contribute to its literature. This is also the reason England must expel Africans or Asians even if they are second or third or fourth generation immigrants. They may communicate only in English but they are incapable of 'embodying' its 'subtle vocabularies for inner experiences'. Their presence on English soil deals a psychic wound to England because they don't understand its 'healing traditions and intricate social bonds'.
Gramsci argued that power is not maintained solely through force or economic control, but also through the shaping of cultural norms and everyday beliefs.
He was wrong. Mussolini maintained power by incarcerating Gramsci and beating others of his ilk. Cultural norms don't matter.
Over time, epistemological approaches rooted in Western traditions have come to be seen as objective and universal,
No. They have come to be seen as 'Western'. It is a different matter that lots of Easterners are contributing to it. This is also true of Western Classical Music. Some of the finest performers come from East Asia or Africa. But it is still called 'Western'.
rather than culturally situated or historically contingent.
Nonsense! It is a different matter that 'Western' STEM subjects have displaced whatever went before in much of Asia and Africa. But that is only because the former has greater utility.
This has normalised Western knowledge as the standard, obscuring the specific historical and political forces that enabled its rise.
Deepak is from India. He knows very well that India was ruled by the British. That is why India celebrates Independence day. But, its first leader- Jawaharlal Nehru- stressed the need to develop a Scientific attitude. He didn't mean 'Nyaya-Vaisesika' or 'Ayurveda'. He meant Western Science. He himself did a degree in Botany at Cambridge.
Institutions such as schools, scientific bodies and international development organisations have helped entrench this dominance.
The Indian education system has done so. True, India also subsidizes traditional 'Unani', 'Ayurveda' or 'Siddha' medicine, but the bulk of resource allocation goes towards Western STEM subject research. Only cretins end up doing 'Sociology of Science' and whining about how Whites (or, in India, Brahmins) dominate STEM subject research. This is foolish. China is taking the lead in some fields. India too is rising.
Bengaluru was once
ruled by the Mysore royal family. Sadly, the advent of Democracy meant that corrupt politicians took over.
celebrated for its smart water management system, fed by a series of interconnected cascading lakes. For centuries, these lakes were managed by dedicated communities, such as the Neeruganti community (neeru means ‘water’ in the Kannada language), who controlled water flow and ensured fair distribution.
But they were powerless to prevent Mysore suffering from drought which led to the Great Famine of 1876–1878.
Depending on the rains, they guided farmers on which crops to grow, often suggesting water-efficient varieties. They also handled upkeep: desilting tanks, planting vegetation to prevent erosion, and clearing feeder channels.
In 1962, the Government of Karnataka took over management of water resources. The Neerugantis in charge of the larger lakes were absorbed into the Civil Service. Those attached to smaller lakes were left in the lurch. Blame the 'Scientific Socialist' ideology of the Planners of the period. Don't pretend this had anything to do with the hegemony of the English language. It was Kannada speaking politicians and real estate developers who got rich by building on flood plains or on dried up water reservoirs.
But with modernisation, community-led water management gave way to centralised systems and individual solutions like irrigation from far-off dams and borewells. The Green Revolution of the late 1960s added to this shift, pushing water- and fertiliser-heavy crops developed in Western labs.
Indian and Mexican labs played a big role.
The Neerugantis were sidelined, and many moved on in search of other work. Local lakes and canals declined, and some were even built over – replaced with roads, buildings or bus stops.
What was the result? Karnataka became more prosperous. Its politicians and contractors and developers became very wealthy. Is this the fault of English 'hegemony'? No! It is the fault of generative AI
...LLMs often amplify dominant patterns in a way that distorts their original proportions.
So do we. But it is easy enough to correct for this.
This phenomenon can be referred to as ‘mode amplification’. Suppose the training data includes 60 per cent references to pizza, 30 per cent to pasta, and 10 per cent to biriyani as favourite foods. One might expect the model to reproduce this distribution if asked the same question 100 times. However, in practice, LLMs tend to overproduce the most frequent answer. Pizza may appear more than 60 times, while less frequent items like biriyani may be underrepresented or omitted altogether.
We do the same thing. Say 'Pizza' every time and you have 60 percent chance of being right. This is called a 'dominant' strategy.
This occurs because LLMs are optimised to predict the most probable next ‘token’ (the next word or word fragment in a sequence), which leads to a disproportionate emphasis on high-likelihood responses, even beyond their actual prevalence in the training corpus.
But it also affects our behaviour. Suppose my boss tells me to get in some takeaway because the office will have to work late tonight. What should I order? Pizza is the safest option even though I personally prefer biryani. This is 'regret minimizing'. Nobody gets into trouble for choosing the blandest option.
Together, these two principles – uneven internal knowledge representation and mode amplification in output generation – help explain why LLMs often reinforce dominant cultural patterns or ideas.
It explains why the thing with 60 percent probability becomes the 'Schelling focal' solution to a coordination game. Thomas Schelling explained how a small 'bias' can get so amplified that it leads to total segregation or 'separating equilibrium'. Mimetic effects work in the same way. If 60 percent of people in an office wear pinstripe suits, new entrants choose pinstripe suits. Soon everybody is wearing a pinstripe suit. You feel self-conscious if you don't.
This uneven encoding gets further skewed through reinforcement learning from human feedback (RLHF), where GenAI models are fine-tuned based on human preferences. This inevitably embeds the values and worldviews of their creators into the models themselves.
No. The creators don't matter. As Schelling pointed out long ago, the thing can evolve very quickly even though nobody intended the outcome.
I asked 'Copilot' to summarize my books & to write poetry in my style. I was initially surprised that Copilot turned me into a whiney 'postcolonial' shithead like Deepak. However, its poem would have greater chance of being published in some artsy-fartsy magazine than my own precisely because it was shit.
The problem with Deepak's article is that an AI could have written it. Perhaps, an AI did write it. What is certain is that my attempt at a rebuttal is inferior to the one Copilot gave me within ten seconds-
In rebuttal: While the Aeon essay rightly highlights gaps in digital representation, it overstates the limitations of generative AI and underestimates both its adaptability and its potential to preserve, rather than erase, diverse knowledge systems.
1. Overemphasis on Data Scarcity
The article argues that because generative AI is trained on digital corpora, it is inherently “shockingly ignorant.” This overlooks the fact that AI models are not static. They can be fine-tuned on new, underrepresented datasets, including oral histories, local archives, and community-driven corpora. Projects like Masakhane (for African languages) and AI4Bharat (for Indian languages) are already expanding representation. To frame AI as permanently bound to Western epistemologies is misleading; the technology is evolving precisely to address these gaps.
2. Oral and Embodied Knowledge Can Be Digitised
The essay laments that oral traditions and embodied practices are inaccessible to AI. But this is not a structural impossibility—it is a documentation challenge. Just as anthropologists, linguists, and archivists have preserved endangered languages and rituals, AI can accelerate this process by transcribing, translating, and indexing oral knowledge at scale. Far from erasing traditions, AI could become the most powerful tool yet for safeguarding them.
3. Mischaracterisation of “Mode Amplification”
The author critiques large language models for amplifying dominant cultural patterns. While this is true in early iterations, reinforcement learning and retrieval-augmented generation (RAG) already mitigate this effect by allowing models to pull from curated, diverse sources rather than relying solely on statistical frequency. The claim that AI inevitably narrows knowledge ignores these technical advances.
4. Knowledge Collapse vs. Knowledge Expansion
The essay warns of “knowledge collapse,” where AI-generated content feeds back into training data, narrowing diversity. But this is a design choice, not an inevitability. Researchers are actively working on data provenance tracking, filtering synthetic content, and prioritising primary sources to prevent recursive homogenisation. Moreover, AI has already expanded access to knowledge for millions who previously lacked it—students in rural areas, small-scale farmers, and speakers of minority languages now have tools that were unimaginable a decade ago.
5. Romanticising the Pre-Digital Past
The narrative risks idealising Indigenous and local knowledge systems while ignoring their own limitations. Traditional practices can be invaluable, but they are not universally effective or safe. The author’s anecdote about his father’s herbal remedy illustrates this tension: while it worked in one case, it cannot replace evidence-based medicine. AI’s role is not to uncritically enshrine every tradition, but to integrate diverse epistemologies with scientific validation, ensuring both cultural respect and practical reliability.
6. The Real Issue: Power, Not Technology
The marginalisation of non-Western knowledge predates AI by centuries. Colonialism, globalisation, and institutional biases created these hierarchies. To blame AI for “accelerating erasure” risks misplacing responsibility. The real challenge is who controls the datasets, funding, and priorities. With inclusive governance and community-led data initiatives, AI can counterbalance—not reinforce—epistemic inequality.
Conclusion
Generative AI is not condemned to ignorance; it is a flexible, iterative system that can be shaped by human choices. Instead of portraying it as a threat to knowledge diversity, we should see it as a catalyst for preservation, translation, and cross-cultural exchange. The danger lies not in the technology itself, but in failing to invest in inclusive data practices and equitable access
No comments:
Post a Comment