India is once again arguing about whether to count caste. In April 2025, the Cabinet Committee on Political Affairs decided that India’s next census—Census 2027, our first digital headcount—would include the first comprehensive caste enumeration since 1931, with its questions notified before enumeration begins.
The arguments made are usually dressed as technical: classification, data protection, administrative capacity. Beneath this caste count is an old question in Indian democracy: does naming an inequality weaken it or make it permanent?
It is a serious debate that involves judges, dissent notes and oped space. What nobody has said aloud is that the state is already late. A shadow caste census has been running in India for years—unofficial, unnotified and unchallenged in court. It’s conducted by banks, matrimonial platforms and delivery apps, none of which need Parliament’s permission to make inferences from data.
Ask an Indian technologist what AI will do next, and you’ll hear about jobs, GDP and a race to keep pace with San Francisco before Shenzhen gets there first. Ask a sociologist and she may pause. The duller answer is the truer one. AI has not arrived as foreign intelligence that could overturn the social order. It has learnt from what is already embedded in that order—what can be made out from someone’s surname, residence neighbourhood, school, marriage, etc.
AI, thus, will not disrupt Indian society. It will formalize its codes and has already been doing so while the country debates whether the government should.
Take a credit-scoring model. It never asks your caste and may never need to. A 2021 Reserve Bank of India report flagged bias and discrimination risk in credit availability and pricing among unregulated lenders, urging transparent, non-discriminatory and auditable AI models.
The mechanism reveals itself plainly: in 2025, researcher Dhiraj Singha asked ChatGPT to refine an academic cover letter and watched it change his surname to Sharma (common among dominant-caste academics), apparently following patterns absorbed from its training text.
Nobody asked; it guessed. Add pin code, college, employer, phone brand—it reconstructs what it wasn’t designed to ask. Remove one proxy, and the pattern returns through another. You cannot make a society identity-blind by making it form blind. The form was not what AI training picked up.
The score that comes back feels neutral, like all numbers do. But AI is not neutral, but merely relieved of the burden of explaining itself. A prejudiced clerk can be accused of prejudice. An automated score spreads the blame so evenly among the bank, vendor, model and the employee reciting “the system won’t let me” that nobody can be blamed. A village headman has a face, but an algorithm has a help line that doesn’t answer.
Or take matrimonial platforms, which have rebranded old family filters—caste, gotra (lineage), complexion, horoscope— as compatibility science. Digitization did not invent prejudice; it reduced its cost. A family that once vetted 20 candidates can now exclude 20,000 before breakfast, exclusion resembling consumer choice.
A 2026 preprint tested five large language models (LLMs), including India’s BharatGPT, on real matrimonial profiles, varying only caste and income. Every model rated same-caste matches up to 25% higher than inter-caste ones, ranking the rest by the traditional caste order. Nobody explicitly taught the machine how to slot or pair castes; it worked that out on its own.
Something similar has happened to the labour market. The point isn’t that an app can guess or ‘predict’ someone’s caste, but that a system built on a caste-shaped market reproduces its distribution of risk and power without storing caste anywhere. Algorithms can assign a gig worker’s route, set his pay, time his pauses and even suspend him by SMS—with no one to appeal to. The hierarchy has become harder to spot but not easier to escape.
This does not make AI uniquely sinister. Used well, it flattens discretion and makes rules auditable as a clerk’s judgement never was. There is thus a genuine case for automated governance that is more honest than either side of the country’s ongoing census debate makes.
But consistency is not the same as justice. A system can apply an unjust rule perfectly, repeating errors at a scale no clerk ever can. India has built a spine of digital infrastructure—identity, payments, benefit transfers —and will now have an official caste record.
Census records are confidential by law. The exposure lies in what future governments and vendors might link or what private systems may infer. If errors get wired into the digital infrastructure used by a sixth of humanity, an unjust decision could injure those who can least afford to spend time and money appealing it.
I should say: I am writing this in English, on a laptop assembled somewhere I have never visited, for an audience inclined to agree with most of it. That is not a disclaimer, but an argument that demonstrates itself.
Those best placed to worry publicly about caste and AI are usually best placed to escape both. They could switch platforms, hire a lawyer or write an essay. So while we debate census forms, someone else’s credit score, entered as evidence by no court, will have already returned its verdict.
The algorithm already knows your caste. The only open question is whether the Republic will catch up with what the market has been inferring for years.
The author is a senior advisory professional.