Empires of Intelligence: What the Colonial Past Can Teach Us About the AI Future

 

The tech superiority in shipping and gunpowder warfare, combined with the appetite for distant resources and captive markets, aided Spain and Portugal in colonizing the Americas from the 1490s. The British and Dutch East India Companies were chartered around 1600, extending this colonization.

Then came the Industrial Revolution in 18^th century: a handful of brilliant machines, the steam engine, the spinning jenny, the power loom, burst out of Britain in the second half of the eighteenth century, raced across Western Europe and the Atlantic, and remade human productivity forever.

What the Industrial Revolution did was not invent extraction — it industrialized it. Railways now hauled raw cotton out of colonized land and finished cloth back in, at a speed and scale no sailing ship could match. The telegraph let a handful of colonial offices in London coordinate an empire on which the sun never set. Steamships and the Maxim gun turned a slow, centuries-long process of colonization into the frantic, decades-long "Scramble for Africa." The machines did not create the logic of empire. They gave an existing logic ‘industrial teeth’.

That distinction matters, because it is also the more useful lens for thinking about artificial intelligence today. The worry is not that AI will cause a new colonization from nothing. It is that AI is industrializing a concentration of power that is already visible, already forming — and that, exactly as happened two centuries ago, the architecture being built right now will determine who spends the next century as a builder of intelligence and who spends it as a tenant.

The New Raw Material, the New Merchant Fleet

Colonial extraction needed three things: a resource worth taking, a fleet capable of moving it, and a captive market for what came back. Frontier artificial intelligence has its own version of all three. The resource is data - the accumulated digital exhaust of human behavior, conversation, and transaction. The fleet is compute: the small number of companies and countries that control advanced semiconductor fabrication, the data centers, and the energy to run them. And the captive market is everyone else, every individual, enterprise, and government that will consume intelligence as an API call rather than build it as sovereign infrastructure.

Look at where frontier foundation models actually get built today. A handful of firms in the United States and China account for nearly all of the models capable of general reasoning at scale. Building one requires not just world-class research talent but hundreds of millions to billions of dollars in compute, energy contracts that rival those of small nations, and access to a semiconductor supply chain concentrated in a handful of firms and geographies. Export controls on advanced chips are, in effect, the modern equivalent of a colonial power controlling who gets gunpowder. None of this is a moral accusation against the companies involved — they are responding rationally to the economics of the technology. But the structural resemblance to the old pattern of resource, fleet, and market is difficult to unsee once you look for it.

The Questions Policymakers Cannot Defer

This raises three questions that deserve to be asked plainly, even though the honest answers are uncomfortable.

How many countries, realistically, will ever have the capital, energy infrastructure, chip access, and research talent needed to build and maintain a frontier foundation model? The number today is small (probably fewer than half a dozen) and the barriers to entry are rising, not falling, as the frontier moves toward ever larger training runs.

What happens to the nations and enterprises that cannot cross that threshold? Do they simply rent intelligence indefinitely, the way a colonized economy once exported raw cotton and imported finished cloth at a price set elsewhere? Renting is not inherently ruinous; nations rent all kinds of capability today, from cloud infrastructure to vaccine manufacturing, without becoming colonies. But renting the layer that increasingly mediates commerce, education, healthcare, and governance is a different order of dependency, because it is not a discrete purchase; it is a permanent tax on every future transaction, and the terms of that tax are set entirely by the renter.

And is "intellectual colonization" too strong a phrase for this, or is it precisely the right one? Colonization was not merely economic dependency, it involved the erasure of local systems of knowledge and their replacement with the colonizer's categories, language, and worldview. A world in which every culture's laws, medicine, and commerce are mediated through a handful of models trained overwhelmingly on the historical and linguistic corpus of a few countries risks something structurally similar: not a flag planted in the ground, but a worldview quietly planted in the model weights that every other nation's citizens interact with daily.

These are not rhetorical questions asked for effect. They are the kind of question that, if left unanswered for another decade, answers itself by default,  in favor of whoever already holds the compute.

Two Ways to Diffuse the Power

If concentration is the risk, diffusion is the countermeasure, and there are two distinct architectural choices policymakers and technologists can push toward, both of which already have working precedents.

The first is pushing intelligence to the edge instead of pooling it at the center. Today's default architecture treats a handful of giant, universally trained models, updated continuously from everyone's interactions, as the intended destination for every query, every business process, every personal decision: an omniscient friend, philosopher, and guide for each individual and enterprise, with the full transcript of the relationship flowing back to a central server. An alternative already exists in embryonic form. Federated learning, a technique in which a model on a device learns from local data and sends back only aggregated, anonymized updates rather than the raw data itself  has been used for years in consumer products like predictive keyboards, precisely because it lets a system improve without every keystroke leaving the phone. Extend that logic further: a model that lives on a phone, a home router, or an enterprise gateway can handle the great majority of everyday reasoning locally, drawing on a general model only when it genuinely needs broader context, and sharing back to any central knowledge base only what the user or enterprise explicitly consents to share. The transaction stays local by default; participation in the global commons becomes an opt-in choice rather than an automatic surrender.

The second is resisting the pull toward one all-purpose model and instead building highly specialized models for individual domains like  health, education, law, financial services,  each trained deeply enough in its own field to outperform a generalist model at the tasks that actually matter to citizens in that domain, and each able to keep improving through use within that domain rather than through indiscriminate absorption into a universal corpus. The risk of leaving specialization there is that it simply recreates ten small walled gardens instead of one large one. The answer is open standards and protocols that let these specialized models interact with each other on a consent basis, a health model calling a financial-inclusion model to check affordability, an education model calling a language model to translate content into a local dialect,  without any of them needing to defer to, or route through, a single dominant global model to reason well.

This is not a hypothetical. India's own experience with open digital protocols is a working demonstration of the underlying principle, even though it was built for commerce rather than AI. Before the Open Network for Digital Commerce, digital commerce in India was consolidating toward the same pattern seen almost everywhere else: a small number of platforms that owned both the buyer relationship and the seller relationship, with every transaction and every unit of pricing power flowing through their walled infrastructure. ONDC instead created an open, interoperable protocol, built on the Beckn protocol, that let any compliant buyer app discover and transact with any compliant seller app, with no single platform sitting in the middle extracting rent from every exchange. The lesson generalizes directly to AI: an open protocol layer for model-to-model interaction could do for intelligence what an open commerce protocol did for retail - letting specialized, smaller players interoperate on equal footing instead of every interaction defaulting to whichever platform happens to be largest.

Both of these architectural choices carry a second, more practical benefit that should appeal to any finance ministry worried about the cost of the AI transition: they are cheaper. Routing the bulk of everyday reasoning through small, local, or narrowly specialized models rather than a giant universal model every single time reduces the number of tokens processed, the compute cycles consumed, and the electricity drawn from the grid. Diffusion is not only a safeguard against concentration of power; it is very plausibly the more economically sustainable path as AI usage scales into billions of daily interactions.

Why Digital Public Infrastructure Is the Precondition, Not an Afterthought

Neither of these architectural interventions works unless the underlying data exists in a usable, trustworthy, and interoperable form in the first place. A local model on a phone in a country where identity, land records, health records, and financial transactions are still paper-based, fragmented, or locked inside proprietary corporate databases has nothing meaningful to reason over. This is precisely the argument for treating digital public infrastructure — the interoperable, open "rails" for identity, payments, and data exchange that countries like India have built through systems such as Aadhaar and the Unified Payments Interface — not as a separate policy agenda from AI, but as its precondition.

The alternative to open, interoperable rails is not the absence of digitization; digitization is happening everywhere regardless. The alternative is digitization captured inside walled gardens controlled by a handful of private platforms, each sitting on a pool of data large enough to train a proprietary model, each with every incentive to prevent that data from ever becoming interoperable with a competitor's, and each able to charge rent on that data's use indefinitely. Digital public infrastructure, built on open standards with consent-based data sharing at its core, is what allows every country, not just the handful that can afford frontier compute,  to accumulate a well-structured, contextually rich pool of its own data, and to let smaller, local, or open-source models be trained meaningfully on that data instead of being permanently dependent on a foreign model's second-hand understanding of local context.

This is also, not incidentally, the strongest antidote to rent-seeking. A market with open, interoperable rails and many interoperating specialized models is a market with real competition, which pushes the cost of intelligence down for everyone. A market of walled gardens converging on two or three global models is a market that, however impressive the technology, will behave like a monopoly,  because eventually, it will be one.

The Choice Is Being Made Now

The colonizing nations of the eighteenth and nineteenth centuries did not sit down and vote on empire; the choice was made, cumulatively, by which ships got built, which trading companies got charters, and which technologies got industrialized first, long before most of the affected societies had any say in the matter. By the time the consequences were fully visible, the architecture was already locked in, and undoing it took centuries.

The architecture of artificial intelligence is being decided now, in this decade, in choices that look small: which protocols become standards, whether edge inference is subsidized or taxed, whether digital public infrastructure is built as an "open commons" or licensed out to whichever platform arrives first with capital. None of these choices individually looks like a decision about empire. Collectively, they are exactly that. The lesson of the last industrial revolution is not that the machines were the problem, it is that by the time everyone understood what the machines had made possible, the terms had already been set by whoever built them first.

Policymakers have a narrower window than they think to make sure that this time, the terms are set by more than a handful of hands.

 “Every empire, however, tells itself and the world that it is unlike all other empires, that its mission is not to plunder and control but to educate and liberate." -Columbia University professor Edward Said.

When AI Climbs the Quadrants of Success

 

What happens to careers, and to education, when machines start solving the problems we used to climb through

For decades, a professional career followed a script so familiar we rarely questioned it. Score well in school. Earn a professional degree that hands you a domain-its concepts, its standard procedures, its tricks, tips and tools. Land a job in an established institution or in government. Then climb, level by level, into steadily more complex problems, until either the ladder ends or you do.

Strip away the detail and every job at every level reduces to the same thing: solving problems, for yourself or for others. What changes as you rise is not the act but the kind of problem you are handed.

A map I drew fifteen years ago

In 2010 I sketched this progression as four quadrants of uncertainty, borrowing a frame from R. Gopalakrishnan and stretching it to fit a career. The two axes were simple: how well you can identify the problem, and how well you can find the solution. (The original post is here.)

In Quadrant 1, the problem is known and so is the solution. Success needs a good repository of knowledge and the discipline to apply it, you know the tune and you know the steps. In Quadrant 2, the solution is known but the problem is not; you earn your keep through analysis and root-cause deduction, figuring out what the crowd wants and then performing it. Quadrant 3 is where the known solutions run out: the problem is clear, but the answer must be invented, which demands curiosity, persistence, and the maturity to abandon an idea you have grown fond of. Quadrant 4 is the leader's terrain, problem undefined, solution unforged, only a vision of a place no one has been and the will to get others to build toward it.

Most careers walk this map slowly, from corner to corner. A few leap quadrants. A rare few begin in the hardest one. But the shape held for a generation, and organisations were built around it, entry rungs where the young learned their craft, senior rungs where judgment lived. The whole edifice assumed a pipeline: you paid your dues in Quadrant 1, and the paying itself was the education. The trainee who reconciled ledgers for three years was not just producing reconciliations; she was, without anyone naming it, absorbing the pattern-sense that would later let her spot the anomaly no procedure flagged. The value flowed in two directions, output for the firm, tacit judgment for the person, and we mostly noticed only the first.

It is worth being precise about what each quadrant rewards, because AI does not treat them equally. Quadrant 1 rewards fidelity: doing the known thing correctly and consistently. Quadrant 2 rewards diagnosis: seeing through symptoms to the actual fault. Quadrant 3 rewards invention under constraint: producing a solution that did not exist, to a problem you can at least name. Quadrant 4 rewards imagination under ambiguity: deciding what is even worth attempting when neither problem nor answer is given. Fidelity is the most codifiable of these, and codifiable is exactly what machines eat first.

Then AI walked onto the map

Every wave of technology made us better inside each quadrant. Better tools, better analysis, better reach. AI is different in kind, not degree: it does not just sharpen the work within a quadrant, it absorbs the quadrant.

Quadrant 1 is already substantially automatable. The junior coder writing from a known pattern, the first-line support agent following a script, the trainee doctor working through a differential, the junior lawyer drafting from a template; this is precisely the known-problem, known-solution work that current systems handle well. Quadrant 2, the diagnostic labour of finding the real problem before applying a known fix, is following fast; increasingly the machine does the root-cause pass and hands a human the shortlist.

Quadrants 3 and 4 will hold out longer. Genuine invention and genuine vision still sit beyond what these systems originate rather than remix. But "longer" is not "forever," and even here the honest word is augment: the innovator and the visionary will work with AI as a collaborator that widens what one mind can attempt.

Two stories, and the one that actually matters

You can tell this two ways. The grim version: AI is eroding the opportunities of a vast section of job-seekers, dissolving exactly the lower rungs where a generation once learned its trade. The expansive version: it multiplies what we can attempt, new problems become solvable, service quality rises, new forms of entertainment and enterprise appear, and human–AI collaboration opens frontiers we could not reach alone. Perhaps, eventually, even new geographies, new planets, a genuine abundance.

Both may turn out to be partly true, and we will argue about the balance for years. History gives ammunition to both camps: earlier automation destroyed categories of work while creating others we could not have named in advance, and the net was, over long horizons, more prosperity. But that reassurance carries an asterisk that is easy to skip past. The transitions were brutal for the people caught in them, they unfolded over generations rather than years, and the new work usually demanded capabilities the displaced did not have and could not quickly acquire. "It worked out eventually" is cold comfort to a cohort whose working life is the transition.

What is genuinely new this time is the direction of travel. Past machines climbed from the bottom, they took the physical, the repetitive, the manual, and human advantage retreated upward into cognition. AI climbs from a different face of the mountain. It is strongest precisely at the codified cognitive work we treated as the entry point to professional life. The rungs being removed are not only the lowest; they are the ones we used as the on-ramp to everything above.

But fixating on the optimist-versus-pessimist argument is a distraction, because it treats the outcome as weather, something that happens to us while we hold an umbrella. The decisive variable is not whether AI expands or erodes opportunity in the aggregate. It is whether we rebuild the one institution that decides which future the next generation is actually equipped for: education.

Here is the uncomfortable mechanism. For a century, education has been optimised to manufacture competent Quadrant-1 workers, people who can absorb a domain and execute known solutions reliably. The examination, the syllabus, the graded problem set, the very architecture of a degree, all reward the reliable reproduction of known answers. That was rational when the economy needed millions of such people and when Quadrant-1 competence was also the on-ramp to everything above it. AI breaks both halves of that logic at once. The economy needs far fewer Quadrant-1 executors, and the traditional path, learn the craft by grinding through its lowest tasks for a decade, no longer works when those tasks are the first to be automated. You cannot apprentice into judgment by doing the drudgery when the drudgery is gone.

This is the trap, stated plainly: the same system that is least useful for producing the capabilities AI cannot replicate is the system we are still running at full capacity. We are optimising harder for the quadrant that is vanishing, because that is the quadrant our institutions know how to measure, fund, and rank. The measurable is crowding out the valuable, and the gap widens with every model release.

What reinventing education would actually mean

If that diagnosis is right, tinkering will not save us. "Add a coding class, add an AI module" leaves the Quadrant-1 factory intact. The harder pointers, offered as arguments rather than certainties:

–    Stop teaching for recall; teach for judgment. When the machine holds the facts and executes the procedures, the human value is knowing which problem to solve, when the machine is wrong, and what to do when the situation falls outside the training data. Curriculum built around memorising and reproducing is training for the quadrant that is disappearing fastest.

–    Move Quadrant 3–4 skills forward by decades. We currently defer invention, ambiguity, and vision to the senior rungs, things you are "ready for" at forty. If AI collapses the lower rungs, an eighteen-year-old must start much closer to the top. That means teaching students to sit with ill-defined problems, run experiments, and tolerate the frustration of not knowing, early and deliberately, rather than as a reward for surviving the grind.

–    Rebuild the apprenticeship model without the drudgery. The old system had a hidden virtue: repetitive junior work quietly built pattern-recognition and tacit judgment. Remove the work and you remove the schooling. Someone has to design the replacement, simulated hard cases, AI as a sparring partner that poses problems rather than solves them, deliberate exposure to failure, so the young still develop instincts they can no longer absorb by osmosis.

–    Teach students to direct AI, not compete with it. The durable skill is orchestration: framing a problem well, interrogating an AI's output, catching its confident errors, and combining several tools toward an end. This is closer to managing a talented, unreliable team than to using a calculator, and almost no curriculum teaches it.

–    Make learning genuinely continuous, and mean it. "Lifelong learning" has been a slogan for thirty years while the actual structure, front-load education, then work, stayed fixed. If the ground shifts every few years, the institutional model of a one-time degree followed by a career is itself obsolete. Re-entry into education has to become normal, cheap, and expected, not a mid-life exception.

–    Credential the things machines can't yet certify. As AI can pass most knowledge tests, the signalling value of exams that measure recall collapses. What still needs a human verdict, originality, judgment under ambiguity, the quality of a question rather than an answer, is exactly what our assessment systems are worst at measuring. That gap is where the reinvention has to happen.

None of these is a finished blueprint, and reasonable people will contest several. The point is the direction: an education system that keeps producing Quadrant-1 graduates for Quadrant-1 jobs that no longer exist is not protecting the young, it is walking them off a cliff with a certificate in hand.

The real question

So the future of work is not, at heart, a question about AI. AI is the forcing function. The real question is whether our schools, universities and training systems can stop preparing people to begin at the bottom of a ladder whose bottom rungs have been sawn off, and start teaching the next generation to begin where the machines, for now, still cannot follow.

We will figure out the world of abundanc
e or scarcity as it comes. What we decide now is whether the coming generation is built to shape that world or to be displaced by it. That decision is being made, mostly by inertia, in classrooms today.

 

“Disruption becomes transformative when it does not merely change what we do, but changes the ground on which we decide what is possible.”

THE UPSIDE WITHIN

 

The morning we launched NSDL, the newspaper called it “a new era for Indian investors.”

I kept that line in my head in the coming days as the purpose of our existence. Because for almost a year, that was all we had. The headline said new era. The ledger said single digits.

Not thousands of accounts. Not hundreds. Single digits, against a country of investors we were supposed to be opening this for.

And every few weeks the phone would ring, from the press, from the markets. “How many today?” And I would give a number you could count on one hand.

Then the same newspapers that had written “new era” printed the verdict. “A costly experiment that isn’t working.” Change the model. It has failed.

I was not the man in charge, I was young, and a part of the founding team. But when you read your project’s obituary in the same paper that announced its birth, seniority doesn’t protect you. It felt like mine.

And I had the choice a lot of you either have had or are going to have. Walk away from something the world has already buried or don’t.

Now you’re waiting for the grit speech. I pushed harder, I never gave up, believe in yourself.

That would be a lie. It wasn’t grit. Grit often is just fear with its teeth clenched, and fear cannot build anything new, a mind in panic can only defend, it cannot imagine.

What got us through was quieter. Two things that our MD and mentor Mr Bhave instilled in us. A sense of gratitude to steady the mind, a grip on what was still standing, even at single digits. And a purpose to point it, a reason worth the year.

Steady, and aimed. That combination will attempt almost anything. We just didn’t have the words for it yet. We only had the account that wouldn’t open, and a reason I couldn’t let go of.

Here’s the thing nobody warns you about hard problems. There’s a kind where you know the problem and you know the answer, you just do the work. There’s a harder kind where you know the problem but not the answer. And then there’s the last kind, where you don’t know the problem and you don’t know the answer. You set out with nothing but the shape of a dream in the dark.

That’s where the depository lived. That’s where anything worth building lives. Not with a plan, with a shape, and the nerve to walk toward it.

There’s an old story. A mouse complains to a lion his life is too dangerous. “Simple,” says the lion. “Become as big as me.” “But how?” asks the mouse. “That’s your problem,” says the lion. “I’m the king. I only make policy.”

Most grand visions are that lion. A magnificent goal, and none of the weight of how. A real purpose is the opposite; it’s a dream you agree to carry the how for. That’s the whole difference between a leader and a slogan.

Picture a metro station that’s been spotless for years. Not one wrapper. Not one mark on the wall. And notice what that does to you, you’d feel ashamed to be the one who drops the first thing.

Now picture one cigarette butt in the corner. Just one. And something in you shifts. Now one more won’t matter. And if there are already ten thousand complaints no one has answered, what harm can one more do?

That’s how a station falls apart. That’s how an institution falls apart. And that’s how a person falls apart. Never in one collapse, always one small permission at a time.

So the hardest discipline of that whole year wasn’t heroic. It was an account that everyone said wasn’t worth chasing. And chasing it anyway. Because the day you stop caring about the one, you’ve already started to fall.

Years later I was asked to build something much larger. An open network, so the smallest shopkeeper in the country could reach a buyer anywhere, without belonging to anyone’s walled garden.

And early on, it came back. The doubt. The “this will never work.” The pressure to quietly change the model. The same cold feeling I’d had years before, alone, at single digits.

But this time I recognised it. I knew exactly what state I was in, and I knew it was lying to me. The steadiness I’d stumbled into by accident the first time, I could now find on purpose.

Same two things. Bigger stage. I won’t walk you through what got built. What matters today is only this, it was portable. Whatever carried a frightened young man through one project carried a national idea through the next. It travelled with me.

And then it did something I didn’t expect. It left.

The idea travelled to another country. There’s an Indonesia Open Network now, and I’m part of it, but not as the founder, and not as the man in charge. As a mentor. My entire job is to help a local team learn from our idea, and our mistakes, and then to make sure they own it, and they deliver it.

Think about what that proves. If this had really been mine, my brilliance, my achievement, it couldn’t move. It would still need me. The fact that it’s growing in Jakarta, in hands that aren’t mine, in a language I don’t speak, that’s the proof it was never mine to begin with.

The idea was never in me. It was in the idea. And the idea only proved its upside the moment it stopped needing me.

So, that headline. “A new era for Indian investors.” It turned out it was never really about investors, or accounts, or a model that worked.

It was a test of what was inside, whether there was something under the failure steady enough, and aimed enough, to chase one stuck account through almost a year of obituaries.

Every one of you is carrying something right now that looks, from the outside, like the odds are against it. And you’ve been told the upside is out there, in a better year, more money, an easier hand.

It isn’t out there. It never was.

And you’ll know it was real not when it makes you look big, but when it outgrows you. When you can put it in someone else’s hands, step back, and watch it stand without you.

That’s the upside within. It was never meant to stay inside you.

Insurance for a Billion: Will AI Make It Fairer, or Just More Precise?

 

[This article is built around thought I shared in the round table titled “Insuretech for India” at Stride Forward 26]

India built the world's most complete public digital infrastructure to include people. The next test is whether it uses AI to protect more of them,  or to price the riskiest ones out.

Insurance is the one financial product designed to work by pooling strangers together. The healthy subsidise the sick, the lucky subsidise the unlucky, and everyone buys protection against a future none of them can predict. That is not a flaw in the model. It is the model. Artificial intelligence is now very good at predicting exactly who will get sick, who will crash, and whose house will flood, and that ability, left to run on its own logic, quietly dismantles the thing that made insurance worth having.

This is the real question hanging over the insurance industry, and it is sharpest in India, not because India is behind, but because India is unusually well equipped to take it either way. Over the last decade the country has assembled the most complete public digital infrastructure in the world: a billion-scale digital identity, real-time payments, consent-based data sharing, digital documents and signatures. That stack was built, deliberately, to include people who markets had left out. The same stack, pointed at insurance and combined with AI, can be used to include far more people,  or to segment them so finely that the ones who most need cover can no longer afford it. India will have to choose. Most countries will not get to make that choice as consciously, because they lack the rails to make either outcome happen at scale.

So it is worth being clear about what is actually at stake, and where the genuinely hard problem lies — because it is not where most of the industry conversation puts it.

The easy part: insurance is about to disappear into everything else

Start with the parts that are, by now, close to consensus. The first wave of insurance technology everywhere,  cheaper distribution, faster underwriting, quicker claims, lower operating cost,  is largely a solved direction of travel. Three shifts follow from it, and they will define the next several years.

Insurance becomes embedded rather than sold. It stops being an annual contract you remember to renew and becomes a service that attaches itself, invisibly, to something else you are already doing. You buy a two-wheeler and accident cover comes with it. You take a home loan and property cover is part of the transaction. You book a trip, finance an MSME invoice, buy farm equipment, or see a doctor on a health platform, and the relevant protection is simply present. India's digital platforms make this possible at a scale few markets can match. The right ambition is for insurance to become, in the phrase I keep coming back to, always present but almost invisible.

AI settles the straightforward claims in near real time. A large share of claims are simple, honest, and slow only because a human has to look at them. Those will be assessed and paid in minutes. This matters less as an efficiency story than as a trust story, which I will come to.

Every citizen carries a portable insurance profile. Identity, verified financial history, health records shared with consent, property records, and past claims can travel with the individual rather than being locked inside one insurer. That lets a person move between insurers without starting from zero each time, and it lets underwriting happen in real time at a fraction of today's cost.

None of this is the hard part. It is expensive and fiddly to build, but the direction is not in doubt and the benefits are real. If this were the whole story, the correct posture would be enthusiasm and patience.

The barrier that technology alone does not fix

The deeper obstacle in India has never been mainly technological. It is trust, and it has four distinct faces. People do not reliably know what to buy, whether a claim will actually be honoured when it matters, whether the premium they are quoted is fair, or whether the whole process is simple enough to be worth attempting. Every one of those is a reason someone who should be insured is not.

Technology helps with each, AI-assisted advice for the first, transparent pricing for the third, paperless onboarding and cashless claims for the fourth. But the second, whether claims are honoured, is where technology and trust actually meet. A claims process that pays honest claims in minutes, visibly and repeatedly, builds the kind of trust that a marketing campaign cannot. That is why real-time claims settlement matters more than its efficiency suggests: it is the mechanism by which an industry with a credibility problem earns credibility back.

Why India can build rails, not just products

Here is where India's position differs from most markets, and it is worth stating precisely rather than triumphantly. Elsewhere, the natural unit of progress is the company: a better insurer, a smarter underwriting model, a slicker app, each building its own private ecosystem. India has the option to build the shared layer underneath all of them,  common digital rails for insurance in the way real-time payments became common rails for money, so that insurers compete on top of shared infrastructure instead of each rebuilding the plumbing.

The reason India can attempt this is that most of the foundation already exists and is public: verifiable identity, a consent architecture for sharing financial and health data, digital documents and signatures. Very few countries have that combination in public hands. The missing layer is insurance itself , the standards and rails that would let a verified individual be underwritten, insured, and served across providers with their consent and without friction. Get that layer right and the cost of issuing and servicing a policy falls far enough that protecting a low-income family becomes commercially viable rather than charitable. This is the genuinely globally significant experiment, and it is why people outside India should be watching it: it is a test of whether insurance can be run as public infrastructure rather than only as a private product.

But this is exactly where the capability turns double-edged, because the same rails that can underwrite a poor family in real time can also price that family out in real time. The infrastructure is neutral. The choice is not.

The tension at the centre: pooling versus prediction

Traditional insurance rests on risk pooling. AI, fed with rich personal data, pushes relentlessly toward risk segmentation,  pricing each individual according to their own predicted risk. Taken to their logical ends, these two ideas are in direct conflict. The better AI becomes at predicting risk, the worse insurance becomes at sharing it.

Follow the logic to its conclusion. Healthy, low-risk people pay very little, as they should on pure actuarial grounds. High-risk people face premiums that climb until cover is effectively out of reach. The result is quietly perverse: the people who most need protection are the ones priced out of it. That does not just produce an unfair market. It hollows out the social purpose of insurance altogether, because a pool that has expelled everyone likely to claim is no longer performing the one function that justified it.

This is not a hypothetical that regulators have failed to notice. It is precisely why many jurisdictions already prohibit insurers from using certain information, genetic test results, disability status, pregnancy, some pre-existing conditions, various protected characteristics. Those prohibitions are not technological limits; the data is often perfectly usable. They are deliberate policy choices to preserve solidarity even when better prediction is available. AI does not create this dilemma. It sharpens it to a point, by making near-perfect prediction cheap and universal rather than partial and expensive.

The better question: predict risk, or prevent it?

There is a way out of the trap, and it comes from asking a different question. What if AI were used not to charge sick people more, but to make them less likely to be sick?

Imagine an insurer that continuously observes, with consent, the early signals,  rising blood sugar, worsening blood pressure, an irregular heart rhythm. The segmentation instinct is to reprice the moment the risk appears. The alternative is to intervene: a teleconsultation, nutrition coaching, subsidised medication, a fitness programme, an early screening. If the intervention works, the person stays healthier, hospitalisations fall, claims drop, the insurer's costs fall with them, and society carries a lighter burden of disease. The insurer stops being a payer of claims and becomes a manager of health outcomes. Prediction is put to work preventing the loss rather than pricing it.

That reframes the whole debate. The question regulators and builders should be asking now, at the start of AI adoption and not after the practices have set, is a simple one with large consequences: should AI be used primarily to predict risk more accurately, or to reduce risk before it materialises? An industry that mostly predicts becomes more exclusionary with every improvement in its models. An industry that mostly prevents becomes more inclusive as its models improve. Same technology; opposite social result. The difference is a design choice, and design choices are easiest to make early.

A question for society, not for the algorithm

Insurance has always balanced two principles that pull against each other: actuarial fairness, which says each person should pay according to their own risk, and social solidarity, which says a pool should absorb risk on behalf of those who draw the unlucky number. For most of the industry's history the balance was set by ignorance,  insurers simply could not price individuals finely enough to fully abandon the pool. AI removes that ignorance. It will make actuarial fairness very nearly perfect.

Which means the balance can no longer be left to accident. The real question is no longer what insurers are able to price, but how much solidarity a society chooses to keep once perfect pricing is possible. That is not a question an algorithm can answer. It is a question for society, and it has to be answered on purpose.

India is unusually well placed to lead that conversation, and not by coincidence. Its digital public infrastructure gives it prediction capabilities that few countries can match, and its public policy has, through that same infrastructure, consistently chosen to expand inclusion rather than optimise markets for their own sake. That combination is rare: the technical capacity to segment perfectly, paired with a stated preference for including everyone. The test for Indian insurance technology — and the reason the rest of the world has a stake in how it goes — is whether it uses AI to expand protection or merely to refine pricing.

India's last financial inclusion story was about opening bank accounts, and it largely succeeded. The next one will not be about accounts at all. It will be about how many lives we choose to protect once we finally have the tools to protect, or to exclude, every one of them.

The better a machine gets at predicting who will suffer, the more deliberately a society must decide to stand with them anyway.

The Queen or the Swarm: Why AI’s Future Depends on Who Gets to Learn

The detailed version of my oped in Transcontinental Times 

Every species that has ever competed for resources on this planet has done so with roughly the same toolkit: strength, speed, camouflage, venom, numbers. Humans are unremarkable on most of these axes. We are slower than a cheetah, weaker than a chimpanzee, blinder in the dark than an owl, and worse at smelling danger than almost anything with a snout. What we do better than any other species, by an order of magnitude no other animal comes close to, is cooperate with strangers at scale.

A wolf pack cooperates. So does a beehive. But a wolf will not lay down its life for a wolf it has never met from a pack three mountains away, coordinated by a shared story none of them witnessed. Humans do this routinely. We show up at war memorials for people we never knew, buy shares in companies run by executives we’ve never met, and hand savings to banks based on nothing but a shared belief that the institution will honor its promises. Yuval Harari’s core insight in Sapiens, that humans are unique in our capacity to organize around shared fictions: nations, currencies, corporations, human rights, describes exactly this. None of these things exist as physical objects. They exist because enough people agree to act as if they do, and that agreement is what lets seven strangers organize a supply chain across four continents.

That capacity has a name in organizational theory: institutionalized trust. And it has a delivery mechanism: communication, first spoken, then written, then encoded into the procedures, contracts, and bureaucracies that let a stranger in Rotterdam trust a signature from a stranger in Chennai. Bureaucracy earns a bad reputation as red tape, but at its root it is a trust technology, a set of standardized procedures that lets unrelated people transact without needing to personally verify each other’s character. Double-entry bookkeeping, the joint-stock company, the postal system, the passport, these are all inventions in the same category as language itself: tools that let cooperation scale past the roughly 150 people (Robin Dunbar’s famous number) that our brains can track through personal relationship alone.

Physical technology has always ridden alongside this social technology, amplifying it. The wheel didn’t just move goods faster, combined with standardized axle widths and road networks, it made regional trust networks viable at distances no courier on foot could sustain. The printing press didn’t just reproduce text,  by making the Bible, and later pamphlets, newspapers, and scientific journals, available to anyone literate, it broke the monopoly that scribal elites held over what counted as agreed-upon truth, and in doing so it re-founded whole religious and political orders. The telegraph, the telephone, the internet: each one is, underneath the marketing, a new trust-and-coordination layer stacked on the ones before it.

Now comes artificial intelligence, and it is not just another rung on that ladder. It is different in kind, because for the first time the tool doesn’t merely transmit human-generated trust signals faster ,  it can generate judgment, synthesis, and decisions on its own. That changes the question. It’s no longer just “how fast can strangers coordinate” but “who gets to do the coordinating, and on whose behalf.”

There are two directions this can go, and they lead to very different civilizations.

The first direction: the Borg model

A small number of frontier labs - right now, realistically, a handful of companies in two countries - train models on a scale of data and compute that nobody else can replicate: the entire searchable internet, increasingly private data through partnerships and acquisitions, and enough proprietary usage logs from hundreds of millions of daily conversations to know how people think, argue, and decide better than the people know themselves. Everyone else becomes a client, querying a central intelligence that has never revealed what it learned about them to anyone but itself. It is not an accident that the Star Trek Borg is the right metaphor: a distributed set of drones, individually unremarkable, whose intelligence is aggregated upward into a Queen who alone sees the whole picture and alone decides. Assimilation doesn’t require malice. It only requires that everyone’s data flows one way  upwards ,  while judgment flows back down as a service.

The second direction: diffusion

Instead of one model trained on everyone’s data and queried by everyone, imagine a nested architecture of intelligence that mirrors the way trust itself has always scaled in human societies - from individual to family to community to nation, each layer adding coordination without fully surrendering what came before it. A personal model that learns primarily from an individual’s own history and stays substantially theirs. A household-level router that reconciles the family’s shared needs,  finances, schedules, health, without exporting the raw data to any central party. Community and institutional layers that pool just enough signal to coordinate like a hospital network sharing anonymized treatment outcomes, a farming cooperative sharing yield data and so on without surrendering the underlying record. National or civilizational layers that federate further still, for the genuinely public-goods problems: pandemic response, climate modeling, financial stability. Intelligence increases with altitude, but so does the friction required to extract raw data upward. This is closer to how evolution actually organizes complexity -  through modular, semi-autonomous units that coordinate without fully centralizing control - than the single-brain model the Borg represents.

The diffusion model is the one worth betting on, but it is worth being honest about why it isn’t automatic. Model weights being “open” or a chatbot running locally on a phone does not, by itself, redistribute power. Underneath even the most local-feeling AI product today sits a stack that is still extremely concentrated: pretraining compute that only a few labs can afford, chip design and fabrication controlled by a handful of firms in a handful of countries, and energy infrastructure that is itself a scarce, geopolitically contested resource. An open-weight model trained on a closed, centrally-scraped corpus is diffusion in name and centralization in substance , a longer, more comfortable road to the same Queen. If this century’s version of the printing press turns out to require a printing press factory that only three governments can build, the diffusion story collapses into the Borg story with better marketing.

There is, however, an answer to the training problem, and it comes from the closest analogy available: how humans themselves acquire capability. Every human being is “pretrained” on a broad common corpus - language, schooling, the accumulated knowledge of a culture - before developing anything distinctive. Universal education does not centralize human intelligence; it equips each mind to then learn recursively from its own experience, in directions no curriculum planned. The base model can play the same role: a common endowment, trained once on broad public data, the way a public education system is funded once for everyone. Diffusion becomes real at the point past that endowment , when each node in the hierarchy, whether an individual, a household, a firm, or a community, holds not merely a copy of the model but the capacity to keep learning from what it alone can see, and when the owner of that node decides what portion of the learning is exposed upward. This mirrors how capability has always worked in human society. A doctor shares her diagnosis, not the decade of pattern recognition behind it; a firm sells its product, not its process knowledge; a family teaches its children things it would never publish. Skill and disclosure have always been separable, and that separability is precisely what a query-everything-through-the-center architecture destroys - the center learns from every interaction, while the individual accumulates nothing that is durably theirs.

Honesty requires admitting that this recursive-learning-at-the-edge capability does not fully exist yet. Most of what is marketed today as personalization is retrieval: the model consults an individual’s documents and history at query time without changing itself, which means the accumulated learning still lives wherever the model lives. Genuine local learning,  models that update themselves from experience, cheaply, on modest hardware, without catastrophically forgetting what they already knew remains a hard, open engineering problem. The technical trajectory, to be fair, is bending in the right direction: models keep getting smaller for a given level of capability, techniques for cheap adaptation keep improving, and consumer chips now ship with dedicated neural hardware as a matter of course. What is not bending is the commercial trajectory. The economics of every frontier lab reward keeping the learning loop at the center, because centrally accumulated learning is the moat, the more the central model learns from everyone, the harder it becomes for anyone to leave. So the two trajectories diverge: feasibility is diffusing outward while deployment keeps concentrating inward, and it is precisely in that gap that policy has work to do. Recursive learning at the edge will not be handed down by incumbents whose business model it undermines; waiting for the market to deliver it is like waiting for the scribes to distribute the printing press. It has to be pulled forward deliberately  by public research funding, by procurement rules that require publicly purchased AI systems to support local learning and owner-controlled disclosure, and by writing the principle that the learning stays with the learner into data protection law, the way purpose limitation was written in a generation ago.

So the honest version of the bet is not “small models will save us.” It is that diffusion has to be built deliberately, at every layer of the stack, the way earlier trust infrastructure was built deliberately, through standards, law, and public investment, not left to emerge on its own from a market that has every incentive to concentrate.

There is precedent for exactly this kind of deliberate construction, and it is worth pointing to because it already exists rather than remaining hypothetical. India’s approach to digital public infrastructure - a unified payments protocol that any bank or fintech can plug into rather than routing transactions through a single dominant platform, a verifiable-credentials system that lets individuals hold and share their own documents rather than surrendering them to a central database, and an open commerce network that lets buyers and sellers transact across competing apps rather than being locked into whichever platform got there first,  is essentially an attempt to build trust infrastructure as a shared, low-lock-in utility rather than as proprietary rails owned by one company. It is not a perfect model and it has real gaps, but it demonstrates something important: that population-scale coordination doesn’t require a single controlling entity if the protocol layer is deliberately kept open and interoperable. The lesson for AI is not “copy this system” so much as “copy the design principle”, build the equivalent of open rails for identity, data portability, and model access, so that intelligence can be composed from below rather than only distributed from above.

What would that take in practice? A few concrete interventions, none of them exotic:

•         Data portability as an enforceable right, not a feature,  so an individual’s interaction history can move with them between AI providers the way a phone number now moves between carriers, preventing lock-in from doing quietly what outright control could not do openly.

•         Public or multilateral investment in compute and energy capacity outside the two or three countries that currently dominate it, on the model of how public investment built highways and rural electrification rather than waiting for private markets to reach unprofitable places on their own schedule.

•         Interoperability standards for model-to-model and agent-to-agent communication, so a household-level or community-level system can coordinate with a national one without needing to be owned by the same company that owns the national one,  the AI equivalent of the postal system agreeing on envelope sizes.

•         Regulatory pressure specifically aimed at the infrastructure layer,  chips, cloud capacity, energy contracts , rather than only at the visible chatbot layer, since that is where real concentration risk is currently accumulating fastest and most invisibly.

•         A cultural shift among the capable middle tier of nations, those with talent, institutions, and ambition but not frontier-lab-scale capital, toward building shared, federated capability with each other rather than each negotiating bilaterally and separately with the handful of dominant labs, which only reproduces a hub-and-spoke Borg structure one client relationship at a time.

Sceptics of the diffusion path will point out, correctly, that some problems genuinely need a Queen, or at least a very large brain. Pandemic modelling, climate prediction, and fundamental scientific discovery benefit from the kind of massive, centralized compute that only a handful of institutions can field, and no household-level router is going to fold a protein or model a hurricane. The diffusion argument is not that centralized capability should not exist. It is that centralized capability should be treated the way we treat other infrastructure with natural concentration risk, nuclear power, undersea cables, the electrical grid,  as a public utility subject to oversight, access rules, and accountability, rather than as the private property of whichever company got there first. The European Union’s AI Act, whatever its flaws in execution, is at least an attempt to draw that line: to say that as models approach systemic scale, the obligations on their operators should scale with them. Export controls on advanced chips are a cruder version of the same instinct, aimed at slowing the concentration of the compute layer rather than the software layer, even if their current form is more about geopolitical rivalry than about distributing power more broadly.

And the early scaffolding for genuine diffusion is already visible. Federated learning in healthcare, hospitals training shared diagnostic models by exchanging model updates rather than patient records,  demonstrates that collective intelligence does not require pooling raw data in one place. On-device inference, now standard on flagship phones, means a growing share of everyday AI use never has to leave the device at all. Neither fully solves the concentration problems described above, but both show that the direction is technically viable; what is missing is the institutional will to deploy such architectures at population scale rather than leaving them as premium features for those who can already afford to ask.

Humanity’s edge was never raw intelligence. It was the invention of trust technologies that let intelligence combine across strangers without requiring a single mind to hold it all. Writing, law, currency, and bureaucracy did this by distributing judgment while standardizing the interface between people. Whether AI becomes a fifth trust technology in that lineage, or the tool that finally lets a single mind hold it all, is not a question that resolves itself as models get better. It resolves according to who builds the rails underneath them, and how deliberately the rest of us insist that those rails stay open. That is a choice still being made, right now, mostly in rooms far from public view — which is exactly why it needs to be argued for in public.

 

"We taught the whole species to read. We did not hand every book to one reader."

India’s Innovation Strategy and the China Misread

India is assembling an industrial-policy toolkit that includes production-linked incentives, the IndiaAI Mission, semiconductor subsidies, and lessons drawn from global innovation systems. The instinct is understandable. Governments want to compress technological catch-up through coordination and capital. Yet the lesson India appears to be drawing is more complicated than either its admirers or critics suggest.

The most consequential Chinese technology outcome of this decade was not produced by the Chinese state in the way it is often assumed. DeepSeek, the AI firm whose low-cost frontier models unsettled Silicon Valley and reshaped assumptions about the price of intelligence, did not emerge from a national champion program. It was not a state-picked winner under a five-year plan. It did not originate inside China’s formal industrial policy machinery.

India’s Innovation Strategy and the China Misread

Click to read on the full article published in Transcontinental Times

AI Governance and Future of Work

 

My Speech at AI-DPI – 26 Conference organised by NCEAR

Let me begin with a simple observation that I think frames everything we're about to discuss.

In the last few centuries, we witnessed multiple technological disruptions ranging from printing press to industrial revolution to computers to internet. It restructured society what work meant, where people lived, what skills had value, what governments needed to do. The economic and social ripple effects played out over decades.

Today, we are living through a transformation that is much more profound but the ripples are moving in months, not decades. And that gap between the speed of technological change and the capacity of our institutions to respond is precisely why conversations like this one matter.

Welcome to what I hope will be a frank, insightful, and perhaps uncomfortable conversation about AI, governance, and the future of work.

 Let me set the scene.

In the last three years, artificial intelligence has crossed a threshold that surprised even its creators. Large language models can now draft contracts, write code, analyse medical scans, counsel customers, generate creative content, and conduct research, tasks that, until recently, defined the upper tier of knowledge work.

We are no longer talking about AI that automates the routine. We are talking about AI that can perform the cognitive. That is a qualitatively different kind of disruption, and it demands a qualitatively different kind of response.

Three tensions sit at the heart of today's discussion.

First tension is on Governance

When it comes to governance of AI key questions that arise are

Who governs AI? Who benefits? Who bears the cost of disruption? These are political and moral questions, not just technical ones.

There lies the tension between speed and safety. AI development is moving at a pace that regulatory frameworks were simply not designed to match. The EU AI Act took years to negotiate and is already facing questions about whether its risk categories reflect the technology as it exists today, let alone as it will exist in three years. India is developing its own digital governance frameworks, and the choices made here, given the scale of this country's workforce and its digital ambitions, will matter not just domestically but globally.

The core challenge for governance is this: if you regulate too slowly, you cede the field to actors, corporate or national, who face no constraints. If you regulate too quickly, you risk encoding today's assumptions into law and stifling the innovation that could actually solve problems. There is no comfortable middle ground. There is only the hard work of trying to get it roughly right, fast enough to matter.

Here the tension is also between innovation and accountability. The companies building the most powerful AI systems are, understandably, advocates for an environment that allows rapid development. Many also, to their credit, genuinely grapple with questions of safety and responsibility. But the incentive structures of competitive markets are not naturally aligned with the kind of careful, transparent, accountable development that the stakes of this technology require.

Governance, at its best, creates the conditions under which accountability becomes not a constraint on innovation but a foundation for the trust that allows innovation to scale. We do not have that governance architecture yet. Building it nationally and internationally is one of the defining challenges of this decade.

Next tension in in he "Future of Work" that is Already Here

There are three competing narratives on this paradigm shift

• Displacement: AI replaces human jobs at scale
• Augmentation: AI makes workers more productive and valuable
• Transformation: New categories of work emerge that we can't yet name

Here lies the tension   between productivity and dignity. Every study that examines AI's impact on knowledge work shows significant productivity gains. Legal researchers, coders, financial analysts, writers when well-supported by AI tools, they produce more, faster, and often at higher quality. This is genuinely good news.

But productivity gains do not automatically translate into widely shared prosperity. The history of technological disruption is also a history of transition costs  borne disproportionately by workers who lack the resources, the retraining opportunities, or the institutional support to adapt. The question is not whether AI will transform work. It will. The question is whether that transformation will be something we navigate together or something that happens to millions of people who had no voice in shaping it.

 What does this mean for India, specifically?

India is not a passive observer in this story. It is one of the central actors.

This country has one of the world's youngest and most rapidly digitising workforces. It has a technology sector that has spent decades building the global knowledge economy's operational backbone. It has a government that has shown real ambition in digital public infrastructure, from UPI to Aadhaar to the Open Network for Digital Commerce.

And it faces a specific, urgent challenge. A significant proportion of India's IT and BPO workforce, millions of skilled, educated, middle-class workers are employed in precisely the categories of knowledge work that generative AI most directly disrupts. Customer support, document processing, software testing, data annotation, back-office operations. These jobs are not going away tomorrow. But the trajectory is clear, and the window for preparation is not infinite.

At the same time, India has something that not every country has in this moment: scale as an asset. The diversity and volume of India's linguistic, cultural, and domain-specific data; the depth of its technical talent; its position as a potential standard-setter for the Global South in AI governance, these are genuine opportunities, if they are seized with intention.

 So what do we actually need?

I'll offer three propositions to anchor our panel discussion.

First: governance must be adaptive, not just reactive. We need regulatory frameworks that are designed to evolve, that build in review cycles, that involve multistakeholder input, that distinguish between the risks of different applications rather than treating AI as a monolithic category. A diagnostic AI in a hospital has different risk parameters than a recommendation algorithm on a social platform. Governance that treats them identically will either over-regulate the beneficial or under-regulate the harmful.

Second: the future of work requires active investment, not just passive optimism. It is not enough to say that new technologies create new jobs, historically, they often do. What matters is the transition: whether workers have access to retraining, whether institutions like schools and universities adapt their curricula in time, whether social safety nets are designed for an economy where the nature of employment is changing. This is a policy challenge, not just a market one.

Third: the voices in the room must expand. The conversations that shape AI governance tend to happen in a relatively small number of rooms, boardrooms, regulatory agencies, international standards bodies, academic conferences. The people whose working lives will be most directly transformed are rarely in those rooms. That needs to change , not as a matter of procedural fairness, but because the decisions will be better if the inputs are broader.

Let me close with this.

I am neither a pessimist nor an optimist about AI. I am a realist who believes that the outcomes of this transformation are genuinely open that they will be determined not by the technology alone, but by the choices we make about how to develop it, deploy it, govern it, and distribute its benefits.

The future of work is not written. It is being written right now, in the decisions being made in companies, in legislatures, in classrooms, and in conversations like this one.

My hope for today is that we leave this room with sharper questions, not just comfortable answers and perhaps with a clearer sense of where action, not just analysis, is required.

 "AI gives us leverage, ethics gives us direction."