Caste and Code: Does India's Tech Industry Export Its Oldest Hierarchy?

The story we tell about technology is that it wipes the slate clean. A résumé is a résumé, code either compiles or it doesn’t, and a machine cannot see the caste name buried in a surname. That promise — that the industry is a pure meritocracy where only skill counts — is why the question of caste in tech makes so many people uncomfortable. It is meant to be the one place hierarchy goes to die. And yet the more time I spend looking at who builds our tools and who trains our machines, the more I see the opposite: technology does not erase old hierarchies so much as give them a faster, cleaner-looking way to travel.

India’s software industry is one of the great success stories of the last three decades. It lifted millions into the middle class, sent engineers across the world, and became the back office and increasingly the brain of global technology. But an industry does not float free of the society that produced it. The people who staff it grew up inside a social order thousands of years old. The data they feed their models was generated by that same unequal world. So the honest question is not whether caste exists in tech — it is whether the industry has exported it along with everything else.

The leveller that wasn’t

Every wave of new technology arrives wearing the same costume: the great equaliser. The printing press, the telephone, the internet, and now artificial intelligence were each sold as forces that would flatten privilege and open the gates to anyone with talent. Sometimes they genuinely do widen access. A kid with a cheap laptop and a good connection really can learn to code and change their family’s trajectory. I have seen it happen.

But a tool is only ever as neutral as the hands that hold it and the world that shaped those hands. This is a pattern I keep returning to: a technology is announced as a leveller, and then the same people who were already ahead find ways to capture it — because they own it, staff it, and decide what it optimises for. If you want the longer version of that argument, I set it out in how technology gets captured. Caste in tech is a specific, painful instance of the general rule. The bias does not live in the silicon. It lives in the people who build the products and in the data those products learn from, and neutrality on the surface is exactly what lets it move undetected.

How caste actually shows up

Ask someone in the industry whether caste matters at work and you will often get a genuinely puzzled answer: we never discuss it, nobody asks. That silence is not the absence of caste. It is frequently the sound of a dominant group that has never had to think about it, because the system already works in its favour.

Caste shows up first in the pipeline. Who gets to the elite engineering colleges that feed the top companies? Access to good schooling, English fluency, private coaching, and the quiet confidence that you belong in a lecture hall are not evenly distributed — they track closely with caste and class. By the time hiring happens, much of the sorting is already done, upstream and invisible, so a company can run a perfectly “merit-based” process and still reproduce the hierarchy it inherited.

Then there are networks. Tech runs on referrals, on the friend who forwards your CV, the senior who vouches for you, the informal circles where opportunities get shared before they are ever posted. When those networks are shaped by who you went to school with, who you eat with, and who your family knows, they quietly recreate old boundaries under a modern name. A referral culture feels meritocratic from the inside and looks like a closed loop from the outside.

And it shows up in the culture of the workplace itself. Workers from marginalised castes describe being asked, subtly or directly, about their background — the surname, the hometown, the college quota that hints at how you got in. They describe slights, exclusion from social circles, and the exhausting calculation of whether to disclose or hide who they are. This is not confined to India. Reporting from global technology hubs, including Silicon Valley, has surfaced worker testimony describing caste-based discrimination among Indian diaspora employees — enough that some US universities, companies, and local bodies have moved to name caste explicitly as a protected category, the way they already protect race or gender. I want to be careful here: these are contested, still-developing matters, and I am describing what reporting and worker accounts suggest, not a proven ledger of who did what. But the fact that institutions felt they had to act at all tells you the problem crossed the ocean.

The bias does not live in the silicon. It lives in the people who build the products and in the data those products learn from.

Can an algorithm be casteist?

Here is where the stakes rise, because the industry is no longer just staffed by people — it is increasingly run by systems that learn. And a system that learns from history will faithfully learn its prejudices too.

Think about how a machine-learning model is built. You give it a large pile of past examples — who got hired, who got a loan, who was flagged as risky, which neighbourhoods generate which outcomes — and it finds the patterns that predict the label you care about. If the society that generated that data was unequal, the “pattern” the model discovers is often just the shape of that inequality. It does not know it is learning caste. It learns proxies: your pin code, your college, the language of your name, the network you belong to. Strip out the caste field entirely and the model will happily reconstruct it from everything correlated with it. Then it applies that reconstruction at machine speed, to millions of decisions, wearing the calm face of an objective score.

That is the quiet danger. A biased human manager discriminates against the people in front of them. A biased model discriminates against everyone it touches, consistently, and hands each rejection a veneer of mathematical fairness. The disadvantage does not just persist — it scales, and it launders itself in the process. I have written more about who these systems ultimately serve in who owns AI, because the answer to “whose bias gets encoded” usually turns out to be the same as “who owns the thing.”

The second half of the problem is the team. A model’s blind spots mirror the blind spots of the people who built it. If everyone in the room shares roughly the same background, the same schools, the same assumptions about what a “normal” user looks like, then whole categories of harm simply never get raised. Nobody tests whether the résumé screener penalises candidates from a reservation-quota college. Nobody asks whether the fraud model treats a particular community as inherently riskier. A non-representative team does not need to be malicious to ship a biased product. It only needs to not notice — and homogeneous rooms are very good at not noticing. This is one of the sharpest arguments Ambedkar made about machines almost a century ago, that a technology’s liberating potential depends entirely on the social arrangement it is dropped into; I unpack it in what Ambedkar said about technology.

Why this is a global problem, not an Indian one

It would be comfortable to file all of this under “India’s internal issue.” It isn’t, and the reason is structural. India supplies an enormous share of the world’s technology workforce — engineers, managers, and founders threaded through nearly every major company on earth. It also supplies a vast, largely invisible layer of the labour that makes modern AI work: the people who label training data, moderate content, and rate model outputs so that the systems you use feel smart and safe.

Both of those channels carry more than skills. When teams hire, promote, and set the culture, the social assumptions they grew up with travel with them into the room. When workers label data — deciding what counts as toxic, what a “good” answer looks like, which image matches which word — their judgments, and the guidelines they are given, get baked into models used by billions. A hierarchy embedded at the point where data is created or a team is staffed does not stay local. It ships inside the product, to every market that product reaches. That is how something as specific as caste ends up shaping tools used by people who have never heard the word.

A hierarchy embedded at the point where data is created or a team is staffed does not stay local. It ships inside the product, to every market that product reaches.

This is the through-line I keep chasing across everything I write: follow the technology and ask who takes, who pays, and who gets to fight back. In the caste story, the people who take are those the old order already favoured, now with a meritocratic alibi. The people who pay are those the system quietly filters out at every stage while insisting the filter is neutral. And the ability to fight back is unevenly distributed too, because you cannot contest a bias that no one will admit is there. When people worry about whether automation will come for their livelihood — a fear I take seriously in will AI take my job — this is the part that gets missed: it is not only whether the machines take the work, but whose prejudices the machines carry while they do it.

What it would actually take to fix

None of this is an argument that technology is doomed to encode caste, or that the industry’s gains are fake. It is an argument that neutrality has to be built, not assumed. A few things matter more than the rest.

• Recognition. Nothing gets fixed while the official position is that caste doesn’t exist at work. Naming it — in company policy, in anti-discrimination rules, in the categories a firm is even willing to measure — is the precondition for everything else. The recent moves abroad to treat caste as a protected category matter mostly because they end the silence.
• Representation. Diverse teams are not a moral garnish; they are a debugging tool. A room that includes people who have felt the sharp end of a system is far more likely to catch the harm before it ships. Representation has to reach the levels where product and hiring decisions are actually made, not just the entry tier.
• Auditing. Models that make consequential decisions — hiring, lending, moderation — should be tested for disparate impact across groups, including proxies for caste, and the results should be legible to someone outside the team that built them. If you cannot show that your system treats comparable people comparably, you do not get to call it neutral.
• Protections. Workers need a real, safe route to report discrimination without torching their careers, and rules with actual teeth behind them. Recognition without recourse is just a nicer-sounding silence.

The reason I keep coming back to this is that the promise itself is worth defending. Technology genuinely can widen the door — I would not have the life I have if it hadn’t. But it only levels when we insist that it does, when we go looking for the hierarchy it is quietly carrying and refuse to accept “the algorithm decided” as the end of the conversation. The alternative is a world where the oldest hierarchy on earth gets a new distribution channel, faster and harder to see than any that came before, dressed in the clean, guiltless language of code. Whether that happens is not up to the machines. It is up to the people who build them, and the questions we are willing to make them answer.