Last week, Anthropic announced that its latest artificial intelligence (AI) model, Claude Mythos, was too dangerous to release. In testing, the company discovered that the model could unearth thousands of hitherto unknown security vulnerabilities in many of the software applications, operating systems and web browsers that the world depends on. Until it could be sure that these capabilities of the model would not be misused, said Anthropic, it believed it was too risky to let the model loose on the world.

Lurking Bugs

What was particularly disconcerting was that since some of the bugs had been around for decades, they are deeply embedded in many of the critical systems we rely on. This includes a 27-year-old vulnerability in OpenBSD, an operating system believed to be unhackable, and a 16-year-old flaw in FFmpeg, a video library used by billions of devices and that has passed millions of security tests. The model also demonstrated how attackers could assume complete control of a machine by chaining together vulnerabilities in the Linux kernel; when asked to try to escape a sandbox and contact a researcher, the model succeeded effortlessly, posting details of its actions on public-facing websites without being asked.

These are just the bugs Anthropic was willing to talk about. Over 99% of the vulnerabilities the AI firm discovered are yet to be patched, and so details about them have been withheld. The question is not whether these bugs will be fixed, but who gets to decide when, and for whom.

Given the "substantial leap" in the model's cybersecurity capabilities, the company has granted a small number of organisations (several of the world's top tech companies) access to its capabilities so they can scan and patch their systems before these vulnerabilities are exploited. This is, without a doubt, the responsible thing to do. But even as I applaud Anthropic for its restraint, I cannot help but reflect on what this means for everyone else. The small group of organisations with access to Mythos will likely address vulnerabilities in their own systems. But there is a long tail of smaller developers that will not have access to these capabilities, whose software is just as likely to have critical bugs that affect a disproportionately large number of people.

The Race to Exploit

The bigger risk is what happens next. Now that thousands of bugs have been identified in testing, it is a matter of time before these vulnerabilities fall into the hands of those looking to misuse them. The moment it announced that it had identified all these vulnerabilities, Anthropic painted a big target on its back. We must assume that hackers and malevolent non-state actors alike are already doing all they can to access this information now that they know that this trove of vulnerabilities exists.

One might think that, as a leading tech company, Anthropic is probably better equipped than most to keep this information secure. But if there is one thing we know, it is that even the best among us have their moments of weakness. Just weeks before the official Mythos announcement, a misconfiguration in its own content management system exposed nearly 3,000 internal documents to the open internet, including draft blog posts that in hindsight appear to describe Mythos itself. That leak was not the work of a sophisticated attacker but the result of a toggle left in the wrong position. Days later, a second lapse exposed over half a million lines of source code from Claude Code, Anthropic's AI coding tool, to the public for several hours.

To be clear, I am not pointing this out to suggest that Anthropic is a careless company—what happened to it could happen to anyone. The point I am trying to make is that once knowledge of this information enters the public domain, every minute that passes without these bugs being fixed compounds the risk that they will be used to inflict damage.

The Sovereignty Question

But what I worry about the most is what happens when this information reaches rogue actors, who we must presume are already doing everything in their power to gain access to it. Zero-day vulnerabilities have long been a favoured tool in the geopolitical militarisation of technology, and many countries have dedicated considerable resources to acquiring them for deployment against their adversaries.

In 2010, the US and Israel used a piece of malware called Stuxnet to destroy roughly a thousand Iranian nuclear centrifuges by exploiting vulnerabilities they had discovered in the Siemens industrial control software that was being used in its operation. The US government operates an active Vulnerabilities Equities Process to decide which software flaws should be disclosed and which should be retained for intelligence and military use. Stuxnet is just one product created by using these vulnerabilities. It is unlikely to be the last. And while Stuxnet emerged as a result of years of effort by elite teams to identify a handful of exploitable flaws, Mythos can find thousands within weeks.

AI has reached a point of strategic consequence. While cybersecurity may be the first domain where this asymmetry manifests, similar capabilities will be aimed at health systems and military infrastructure before long. For decades, we have built our critical infrastructure on software that we do not control, and it is distributed by companies we do not influence. Now that their vulnerabilities are being catalogued by tools we cannot access, this is no longer a cybersecurity problem. It is an issue of sovereignty.

This article first appeared in the Mint. You can read the original here. For the full archive of all my Ex Machina articles, you can visit my website.

On 30 March, the Ministry of Electronics and Information Technology (MeitY) published draft amendments to the Information Technology (Intermediary Guidelines) Rules, 2021, which, if passed, will have far-reaching consequences for internet users in India. While each of the three primary amendment proposals does something different, together they will transform the internet from a rules-based regime into one governed by discretion. Though they may seem benign, their bite could be worse than their bark.

The new Rule 14(2) would expand the powers of the Inter-Departmental Committee (IDC) from hearing just "complaints and grievances" to "any matter" referred directly by the ministry. Rule 8 takes things further by not just expanding adjudicatory power, but who is subject to it—from 'publishers' to anyone who posts about news online. But both these pale in relation to Rule 3(4), which reshapes the source of legal authority.

A More Powerful IDC

Let's go through each provision to better understand how they will all affect us.

Let us start with Rule 14. Through a set of innocuous adjustments, sub-rule (2) will significantly expand the powers of the IDC. From being an apex appellate body in a three-tier complaint-resolution mechanism for violations of the Code of Ethics, it could become a super-regulator with the power to hear just about "any matter" that the ministry refers to it.

There are two problems with this. First, since the IDC would have the power to rule on anything the ministry sends it, the recommendations it issues could well exceed the statutory authority under which it was created. But what's far worse is that by accepting references directly from the ministry, parties could be denied the two levels of appeal they are currently entitled to, thus collapsing the current three-tier mechanism into a single executive-controlled process. With this one amendment alone, the government could end up violating principles of both administrative law and natural justice.

It is worth noting that in August 2021, the Bombay High Court stayed the Code of Ethics framework as prima facie violative of Article 19(1)(a) of the Constitution of India—a prohibition subsequently extended by the Madras High Court. In this backdrop, for the inter-departmental panel to operate within this now tenuous structure would expand its powers even as the constitutionality of the framework itself remains under litigation and is questionable.

Expansive Code

This brings us to Rule 8(1). As originally drafted, this provision applied only to "publishers of news and current affairs content" and "publishers of online curated content." In other words, regular news outlets as well as their online equivalents. The proposed amendment will extend the applicability of certain provisions to ordinary users who share news or current affairs content. Since sharing posts about news is something we all do, this means that a provision earlier designed for news organisations will now apply to you and me. In effect, the Code of Ethics, a standard of accuracy, fairness and impartiality with which news organizations must comply, could get extended to ordinary internet users who have neither the resources nor the inclination to vet everything they post at this level of rigour.

But it is the proposed addition of a new sub-Rule (4) to Rule 3 that in many ways is the most disappointing. This new provision seeks to legally elevate clarifications, advisories, orders, directions, standard operating procedures, codes of practice and guidelines to the level of rules and regulations (the only forms of subordinate legislation that the executive branch is constitutionally permitted to enact). If brought into force, it will allow the executive to make a new law without calling it a law, and, in doing so, arrogate upon itself the power and authority that ought to vest with the legislature. Or at the very least be subject to Parliamentary supervision.

The power of the executive branch to make rules and regulations is always subject to an obligation to place them before Parliament as soon after they have been made as possible. By elevating various informal executive instruments of this kind to the level of enforceable law—none of which need to be laid before the legislative branch—the government seems to be trying to evade a necessary constitutional check on its power.

Deregulation

What makes these regressive proposals particularly disappointing is that they were released for public consultation just three days after the Jan Vishwas Bill was introduced in the Lok Sabha. The latter, as many have noted, is the single largest horizontal decriminalization effort anywhere in the world, significantly improving the ease of doing business across sectors. Central to this deregulation effort is the principle of regulatory certainty it embodies—that the state should not impose compliance obligations on citizens through instruments that do not carry the weight of the law. For the government to just a few days later introduce amendments that do just that shows how little the left hand appears to know of what the right is doing.

I am the first to admit how hard it is to govern modern online spaces. These proposed amendments are an attempt to walk a well-known tightrope between allowing free expression online and preventing the harms that can result. But if we cannot find a more effective way to strike that balance, we will replace a system governed by rules with one ruled by executive discretion.

Even though the Strait of Hormuz is just about 33km wide at its narrowest point, about 20 million barrels of crude oil and refinery products would pass through it every day before the war in West Asia. This was roughly a quarter of all seaborne oil trade. Since late February, tanker traffic through the strait has fallen sharply, driving up the price of Brent crude and triggering an oil shock. But while our attention has been focused on oil, there are other consequences—equally, if not more significant—that could threaten the global economy in far more dangerous ways.

The Hidden Cascade

Amid the freight sitting idle in the Gulf are 200 containers of liquid helium waiting for Iran's permission to pass through the strait. These containers were chilled to near absolute zero when loaded, but have been slowly warming up with every day's delay. If they are not allowed to pass within the next 40 days, the liquid helium will eventually revert to gas and be vented into the atmosphere, where it will be lost forever.

Helium is produced as a byproduct of processing natural gas into its liquefied form, LNG. Qatar accounts for roughly a third of the world's commercial helium output. When QatarEnergy halted production after Iran attacked its Ras Laffan refinery on 2 March, helium output ceased. This leaves us staring at a global helium shortage both in the short and medium term.

Chipmakers in the semiconductor industry use liquid helium to cool silicon wafers during the plasma etching process. South Korea, home to Samsung and SK Hynix, sources 65% of its helium from Qatar, while Taiwan sources 69%. SK Hynix produces 62% of the world's high-bandwidth memory (HBM) units contained in every Nvidia H100 and B200 GPU. Without HBM units, GPUs cannot be made. Without SK Hynix, HBM units run scarce.

The Hormuz squeeze has not just affected the global energy trade, but several other elements of the deep, multilayered supply chains that hold the world together. While our immediate focus is on energy, the current conflict is a stark reminder of just how tightly interlinked the global economy is.

Our instinctive response to such crises has always been to reshore, decouple and diversify, but we are constantly forced to come to terms with the reality of a highly inter-dependent global production stack. We may have reached the point where self-sufficiency is no longer an option and we have no choice but to live with disruptions.

Chips All the Way Down

Beyond the current crisis, the semiconductor industry's supply chain dependencies run far deeper than helium. TSMC in Taiwan manufactures 92% of the world's most advanced chips, the processors that power every AI model, data centre and smartphone in the world. But TSMC cannot make its chips without highly specialised machines from ASML, a Dutch firm in Veldhoven that has a de facto monopoly over extreme ultraviolet (EUV) lithography, a sophisticated technology essential for making sub-5nm chips. Despite decades of effort and tens of billions of dollars, China remains, by most estimates, at least a decade away from replicating this technological capability.

At the heart of ASML's machines are specialised optics made by Carl Zeiss, a German firm whose semiconductor division is the sole supplier of the mirrors that are essential to every EUV system. These mirrors must be polished to within a single atom's thickness, a level of precision that took 25 years and billions of euros to perfect. This means that the manufacture of 80% of all chips worldwide depends on the optical products made by a single company in Europe.

Which brings us to the strangest dependency of all. In addition to chips, South Korea is also a significant shipbuilding nation. Korean shipyards like HD Hyundai, Samsung Heavy and Hanwha Ocean have built over 80% of the world's LNG carriers over the past five years. This creates a striking paradox: one of the countries that is most dependent on Gulf energy is also the one the world depends on to build the ships it needs to transport hydrocarbons. If the Korean industry slows, the global capacity to reroute energy supply would slow with it, further deepening the crisis.

The Reshore Illusion

Over the past few years, our instinctive response to crises like this has been to localise production. Around the world, countries have invested in domestic semiconductor-making capabilities and rare-earth extraction facilities to offset the impact of conflicts they have had to deal with in recent years. But there are limits to that approach. You cannot compress 30 years of accumulated manufacturing intelligence into a policy cycle. Even after decades of effort, tens of billions of dollars in investment and a concerted engineering push, China cannot build the machines it needs to make sub-5nm chips. That is knowledge that cannot be bought or learned without iteration.

We could view the Hormuz crisis as evidence that our experiment with globalisation has failed. That, however, would be a mistake. If anything, it should remind us that we chose, rationally and repeatedly, to allow different parts of the world to do what they do best so that the global economy could prosper.

Interdependence is not an option. It is the condition under which much of the modern world operates. The real risk is not that we have come to depend too much on one another, but that conflicts like these will sever connections we can no longer afford to break.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the complete archive of all my articles, you can visit my website.

In Ancient Rome, parties to a contract who wanted to record the terms of an agreement had to present themselves before a tabellio—a particular type of clerk whose job was to write down the terms of a contract on a wax tablet, have them witnessed, and make the document official by pressing his seal into the wax. By certifying the document as valid, the tabellio ensured that parties who had no other way to attest to the document’s integrity each had a legally enforceable copy. In a world without trust infrastructure, institutional intermediaries such as these were essential for commerce. This was eventually formalised as the Notarius of Imperial Rome—the precursor to today’s notaries.

Need for Notarisation

This arrangement has not changed much in 2,000 years. We still have to appear in person before a notary for various reasons, armed with original documents and stacks of copies so that, after having inspected the original, he can affix his seal to the copies to unequivocally state that they are identical to the original version in all respects.

There are a surprising number of situations in which we still have to perform this ancient tradition. If you lose your PAN card, you need a notarised affidavit. If you have lost a share certificate, SEBI-regulated entities insist on a notarised indemnity affidavit before they will consider issuing a duplicate. If you are a pensioner, you must provide notarised declarations confirming that you are still alive and eligible. And if you happen to lose your educational certificates, you will be required to produce notarised copies of whatever substitute you can find. Even if there is a spelling mistake in your name, there is no way you will be able to set it right unless you present a notarised copy of some other official document with the correct spelling.

Underpinning all these official requirements is the presumption that what held true in Ancient Rome must still be valid today—that the only way we can assure ourselves of the authenticity of a document’s copy is if we know a human intermediary has compared it with the original and certified a match. This presumption not only places a far greater premium than is justified on the industriousness and attention to detail of the notarial class, but, more importantly, completely ignores the march of digital technology and all that it has made possible.

Digital Documents

Today, most documents that are required  to avail the benefits or services we need are available in digital formats from their issuing authorities. These include identity documents (Aadhaar, PAN and driver’s license), educational credentials (such as school and college marks cards) and proofs of ownership (share certificates, property documents, vehicle registration certificates, etc). Since there is no greater guarantee of a document’s veracity than getting it directly from the issuer’s system, if that’s possible, it seems utterly pointless to insist on a notary certifying the authenticity of a copy.

Consider what we are actually doing when we ask for a notarised copy of an Aadhaar card. Since the notary has no access to the actual Aadhaar database, he can, at best, certify the copy he is being asked to attest is identical in every respect to the physical card he was shown.

Now, everyone knows that the Aadhaar card we carry is just a small piece of paper or plastic indicating that we are enrolled in India’s unique identification system. The physical card we carry does not have in-built anti-counterfeiting or anti-spoofing measures. As a result, it is easy to forge or modify. If a forged Aadhaar card is presented, the notary, not having any way to verify its authenticity, will happily certify that one piece of paper matches the other.

When we insist on a notarised copy of an Aadhaar card, we replace a document signed by its authorised issuing authority with one signed by someone who cannot verify its authenticity. Since the notarised copy is less trustworthy than the original, this moves us down the trust chain, instead of moving us up.

Verifiable Credentials

It does not have to be this way. The Unique Identification Authority of India (UIDAI) already allows Aadhaar number holders to generate verifiable credentials for their Aadhaar information. These digitally signed credentials attest that the information they contain originates from the UIDAI database. Since the document is signed using the UIDAI’s digital certificate, its authenticity can be cryptographically verified, offering a stronger guarantee of authenticity than any notarial attestation. 

What the UIDAI does today, every other authority that issues digital documents can do as well. If, instead of insisting on notarised copies of physical documents, everyone starts accepting verifiable credentials digitally signed by the issuer, we can reduce friction while improving the reliability of documentation across the country. Where affidavits are required, all we need to do is enable citizens to prepare these declarations and certify them with their own Aadhaar e-sign.

The notary was a necessary solution to the problem of establishing trust at a time when there was no reliable societal infrastructure to do so. That problem has since been solved —not by better intermediaries, but by digital infrastructure. Verifiable credentials do not just digitise documents, they replace trust in people with trust in systems.

It has been 2,000 years since we invented the tabellio. There is no reason we should still be hanging on to systems bound by constraints of the past. Other than inertia and our reluctance to let go.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the full archive of all my articles, please visit my website.

Toyota’s factory in Woodstock, Ontario, will soon deploy Digit, a humanoid robot from Agility Robotics, on its production floor. These bipedal machines will unload auto parts from warehouse tuggers and onto the production line—work that until now was performed by humans. Toyota chose humanoid robots because its factories are designed for people. By using robots that move like humans, Toyota can automate without redesigning its assembly lines.

On the face of it, this is a small efficiency improvement. But it is a signal of something far larger: the rise of embodied AI and its potential to reshape the global economy. It indicates a future in which countries that build robots will prosper while those that rely on labour supply will struggle.

Automating Cognitive Work

Just as mechanisation replaced physical labour, artificial intelligence has begun to automate cognitive work. For decades, white-collar jobs were immune to automation. That assumption is collapsing. As AI becomes embedded in machines and gains the ability to interact with the physical world, it will breach the final barrier between software and labour.

Toyota is not the only company using robots. Intuitive Surgical’s da Vinci robots are being used in hospitals around the world and have already completed millions of procedures. Waymo vehicles now complete over 150,000 trips per week, proving that AI-embedded physical systems can operate at scale in unstructured and unpredictable environments. Tesla’s Optimus range of humanoid robots is in production and is projected to scale significantly by the end of the decade.

As interesting as this sounds, rapid automation is likely to have serious repercussions on society, especially in developing economies. A recent paper in Labour Economics argues that the consequences may be so large that normal market adjustments—cheaper labour, lower interest rates or currency depreciation—will not be enough to offset the resulting employment shock. The diffusion of embodied AI could dramatically widen the gap between developing and developed economies.

Demographic Dividend

This could not come at a worse moment for India. With nearly 800 million people under 35, we are experiencing the largest demographic bulge in human history. This youth bulge was supposed to power India’s rise, but if embodied AI is on track to replace labour faster than the economy can absorb it, our demographic dividend could quickly become a demographic disaster.

There may, however, be a way out. Countries that produce robots—rather than import them—should be better placed to ride out the robot revolution. Those that make robots could develop engineering expertise, capture export revenues and build industrial ecosystems that may cushion the labour shock. This suggests that to take advantage of our demographic dividend, we need to actively encourage the manufacture of embodied machines. 

How might we go about doing this? One approach might be to identify domestic use cases for embodied AI to develop natural markets for local robotic solutions. 

Among the many challenges India faces, agricultural productivity is a core concern. With the average size of farm holdings hovering at about a hectare, not only is agriculture in India economically unviable for the most part, it has become a poverty trap that has locked 125 million smallholder families into subsistence living. Of all the problems we most urgently need to solve, finding a path out of this developmental crisis ranks near the top.

Historically, every successful industrial economy has passed through the same transition: labour first moves from farms to factories and then into services. India has, however, struggled to make this transition from farms to factories.

Automating Agriculture

This suggests that if there is one area worth focusing on, it is the use of automation in agriculture to hasten the transition of our economy to industrial production. Since we also need to develop domestic markets for embodied AI, it makes sense to promote companies that embed modern AI technologies in machines designed for use in agricultural settings. If India can become the world’s testbed for agricultural robotics, it could also become a global supplier of those machines.

This is not as far-fetched as it might seem. A Bengaluru startup, Niqo Robotics, has already developed an AI-powered precision sprayer that uses computer vision to identify individual plants and apply agrochemicals with millimetre-level accuracy. By running AI directly on edge devices, the system allows farmers to retrofit existing tractors and use them even in locations with constrained bandwidth, sharply reducing both input costs and environmental damage.

This is just one example of how India could embed AI into farm machines to improve efficiencies. We could see the development of harvest robots designed for crops like cotton, chilli, tomato and onion, which require manual picking and sorting. To raise agricultural productivity, we could look to deploy automated soil-sensing solutions that can dynamically provide farmers customised insights about their land. Thanks to AI-embedded soil moisture sensors, it should be possible to significantly reduce water and pesticide usage on farms.

Every successful economy mechanised its farms before building its factories. We may have to do both at once—and we may have less time than we think. 

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the entire archive of my Ex Machina articles please visit my website.

Last week, I wrote about how frontier artificial intelligence (AI) has begun to improve exponentially—to the point where it is not just introducing linear improvements in functionality, but literally augmenting capabilities. Personally, this became evident when I realized I could code anything I could imagine, allowing me to create various applications and programs to eliminate micro-frustrations in my workflow. I am sure that those in other domains are experiencing similar capability uplifts. 

Returns on Scale

According to scaling laws, a model’s performance will improve so long as increasingly large amounts of compute power and data are used to train it. For many years now, leading AI labs in the US have been proving that thesis at an eye-watering scale. It feels like those investments have begun to pay off, particularly over the past six months. Across the board, frontier models have begun to demonstrate dramatic improvements—both in the quality and accuracy of their outputs and in what they can do. 

For countries like India, advanced AI that not only reduces friction but also unlocks brand new capabilities presents exponential opportunities for advancement. By blurring the boundaries between domains and professions, it can expand the scope of what is possible, allowing those who use it to do far more than before. This makes brand-new pathways available through which the potential of AI can be harnessed across all domains—from healthcare and education to governance.

But just as these exponential capabilities of AI open up new opportunities to transform our industries and markets, they also expose us to new vulnerabilities. At present, the transformative capabilities of AI are only available in frontier models offered by top American AI companies. Open-source AI, which countries like India have relied on so far, seems to be slowly falling behind the cutting edge of what is possible, and while we have made tremendous strides with our own foundation models, they are still behind the state of the art on several key parameters.

AI through the API

Today, access to leading AI models is only available through application programming interfaces (APIs), which would not have been a problem had it not been for the geopolitical uncertainty of the current moment. Any country whose AI future depends on API access to solutions from leading AI labs risks having the core infrastructure on which it depends pulled out from under it should geopolitical winds shift—much like access to the SWIFT financial network was weaponised during the conflict in Ukraine. When core national capabilities depend on access to technology under foreign control, our sovereignty will only last as long as remote servers remain accessible.

So, how does India navigate its way ahead?

History has shown that strategic concessions can be extracted from technology providers seeking market access. When the Brazilian Air Force acquired advanced military aircraft in the 1d960s and 70s, the government, through state-owned enterprise Embraer, required that foreign suppliers transfer manufacturing know-how to Brazilian engineers and facilities. As a result, Embraer grew from a military supplier into one of the world’s most successful commercial aircraft makers, competing globally with Boeing and Airbus in the regional jet category. When Samsung licensed early DRAM technology from Micron, it invested aggressively in manufacturing scale and process improvements to the point that it eventually outcompeted its licensor. The massive electronics supply chain that grew out of this is now one of the world’s most sophisticated. Throughout the 1990s and 2000s, China made market access explicitly conditional on foreign companies forging joint ventures with Chinese partners and enforcing technology transfer under the terms of those agreements. This is how China developed the process expertise it currently has across a range of sectors, allowing it to dominate the global industrial value chain.

India's Strategy

The question of the moment is this: What leverage does India have to achieve its goals in the AI race? 

As the largest user base for AI technology outside the US, India generates enormous volumes of behavioural, transactional and social data. This is exactly the kind of real-world interaction data that AI companies need to improve their models. India also possesses vast repositories of traditional knowledge, cultural information and community practices that could provide AI models with much-needed context.

In a paper that I presented as chair of the Expert Engagement Group on the New Deal for Data under the India AI Summit, I argued that one option available to a country like India is to make access to this data conditional on AI companies depositing the weights of all AI models enriched by Indian data in Indian facilities under a model escrow agreement. This would ensure that, if for any reason API access is cut off, Indian companies can continue to access the AI models they need locally. This would let India safeguard its AI sovereignty by strategically deploying one of its most valuable assets—its data. If successful, India could become the first country to leverage data as a strategic currency in the AI era. 

If we are to do this, we must act swiftly. Once AI models approach human-level performance across domains, any data advantage we currently have will disappear. And with it, the leverage we need to secure our technological sovereignty.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the full archive of my Ex Machina articles, please access my website.

About a month ago, I built my first app.

It was a simple speed-reader Chrome extension that is designed to display text one word at a time, so I can read an article faster than normal. I managed to code the entire thing in one shot, using one of the frontier artificial intelligence (AI) models, and was quite frankly surprised at how easy it was to build. The AI even added, of its own accord, a feature that allowed me to increase the speed, so that within days, I was reading at 500 words per minute.

After that early (relatively easy) success, I grew more ambitious and started building increasingly complex applications—from a custom-built feed reader app to a minimal teleprompter that sits next to my laptop camera. I even made a few iOS games that are now available on the App Store. As my confidence grew, so did the ambition of the code I produced; the more I understood what AI was capable of, the more full-featured and well-designed my apps became. To the point where I now have, at any given time, three-four apps in various stages of development on my computer, and a GitHub page that is dark green with activity.

Exponential Capabilities

I am a lawyer who does not know how to code. I should not have been able to build an app, let alone the half-dozen apps I have so far. That I did is not because some latent programming talent has been unlocked within me, but because frontier AI models can now translate a plain-English description into working software.

Dario Amodei calls the current phase of AI progress “The Exponential.” Until I used AI to prompt an app into existence, I don’t think I fully understood what he meant. Having now immersed myself in AI-assisted coding for the better part of a month, I have a better appreciation for how dramatically things are about to change.

The purpose of technology has always been to improve efficiency and enhance productivity. Most new technologies do this linearly, automating steps that would otherwise have to be performed manually or short-circuiting sequences that were previously unavoidable. Frontier AI doesn’t eliminate steps; it democratizes capabilities. It allows those who once relied on the skills of others to do things they were not previously capable of doing themselves. And by doing so, it unleashes exponential opportunity.

A New World

What does a world with exponential capabilities look like? Personally speaking, it will free me up to do things I never thought I could. I have, for instance, always felt that the legal software we use was not designed with my needs in mind and lacks features that I need. With my newfound coding capabilities, I can simply program those frustrations away, prompting little snippets of custom code into existence to address the deficiencies in my current set-up.

What is good for the goose will likely be good for the gander. Software companies that once relied on lawyers to draft privacy policies and terms of service for their applications will likely get them drafted by AI. As much as my fellow lawyers may not like to hear it, AI is already more than capable of producing documents for this purpose. I should know. I allowed AI to draft the terms of use of the apps that I’ve launched on the App Store.

As frontier AI improves, we will see similar exponential growth in other sectors. Traditional capacity constraints that forced costly interdependencies will melt away, freeing entrepreneurs to build new offerings and deliver new forms of service. While this will affect traditional professions, it will also enable new ways of working, and policymakers around the world will need to balance the need to protect those disrupted by this technology with the opportunities it opens up for a new ‘exponentially capable’ workforce. As a country with a young and adaptable workforce, India should lean into this opportunity rather than attempt to preserve traditional professions that, in their current form, are sure to be disrupted.

So how should we prepare ourselves for what’s coming?

Preparing Ourselves

In the first place, I believe we must be willing to redesign the systems that have served us well so far. What might have worked in the past is unlikely to be necessary in a world where skills can be invoked with a prompt. Take, for example, the traditional hierarchies of software deployment—developers, system integrators, hyperscalers—that have underpinned large-scale software services over the past few decades. In a world of exponential capabilities, the boundaries between different categories of service providers are likely to blur.

Of all countries, India is uniquely positioned to adapt to this new reality. After all, this is precisely what we did when we took the digital public infrastructure (DPI) approach. Instead of leveraging the vertically integrated software stack that was at the time widely believed to be the only way to achieve population-scale deployment, we built reusable, interoperable building blocks that dramatically reduced deployment costs. We need to apply that same iconoclastic approach to shape our AI future.

Cutting-edge AI does not just improve efficiency; it expands human capacity. By lowering entry barriers across disciplines it gives individuals access to skills that were once scarce and expensive. 

The exponential is not about machines becoming more intelligent as much as it is about people becoming more capable. And if we play our cards right, the exponential will not merely change industries. It will change who gets to build them.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the complete archive of all my articles please visit my website.

In November 2023, a few governments and technologists gathered at Bletchley Park to discuss artificial intelligence (AI), seeking to come to terms with the technology they were developing. The mood was sombre and fear was the dominant register. This was the first of a series of AI Summits, the most recent of which was held in New Delhi last week.

More Upbeat

Unlike Bletchley Park, Bharat Mandapam was not only much larger and more crowded, the mood was also markedly more upbeat. With over 500,000 visitors from 118 countries and over 3,250 speakers, the AI Impact Summit held in New Delhi was far and away the largest AI summit to date. But what distinguished this conference was not its size or spectacle, but a growing recognition that the real challenge is not building intelligence, but spreading it.

Any event of this scale is bound to have its fair share of mishaps and the AI Impact Summit was no exception. While many on social media and in the domestic press spent last week fixating on these fumbles—from a misbegotten robot dog to traffic jams and long walks ordinary citizens had to endure to get home—anyone who has been inside the halls of Bharat Mandapam will testify that the corridors were buzzing.

In terms of tangible outcomes, 80 countries and international organizations adopted the ‘New Delhi Declaration on AI Impact,’ a document that underscored the urgent need to realize AI’s potential to drive economic transformation. The Declaration anchored national commitments across three broad ambitions: widening access, embedding accountability and using AI to drive inclusive growth—through reskilling, research and sustainable infrastructure.

Specific Deliverables

There were also other specific deliverables, such as the Charter for the Democratic Diffusion of AI, Global AI Impact Commons, International Network of AI for Science Institutions and the AI for Social Empowerment Platform. Various ministries and regulators used the Summit to announce new policy initiatives, including the health ministry, which launched SAHI, India’s national framework for AI in healthcare. Many of these documents will serve as signposts for further action after the summit. I hope to engage in some of this work myself through the Expert Engagement Group on ‘A New Deal for Data’ that I chair.

The Summit also served as an occasion to announce India’s entry into the LLM race. Three Indian foundational models were launched last week—Gnani’s text-to-speech model Vachana, BharatGen’s Param2, a 17-billion-parameter multilingual model, and Sarvam’s 30- and 105-billion-parameter models. The latter were especially impressive for their performance on various benchmarks, achieving state-of-the-art results on several criteria relevant to India. Above all, these launches signalled that India intends not merely to adopt global models, but to compete successfully at the foundational layer of AI itself.

For me, the real value of the Summit came from all the many conversations we had. With the who’s who of AI—heads of big AI labs, semiconductor companies and data centre providers, as well as 20 heads of state and 60 ministers—in attendance, the quality of discussions on the big stage as well as along the sidelines of scheduled events was superlative. During the week, over 500 sessions were held on subjects as wide and varied as they were deep and substantive. While the keynotes on the main stage served as an opportunity to make announcements and investment commitments, it was the panel discussions that really offered an opportunity for debate, discussion and healthy disagreement. 

Of the tiny fraction of sessions I was personally a part of, we discussed issues as diverse as the governance of data-sharing networks for AI, how AI could benefit countries and the planet, and the importance of keeping the internet open to truly democratize AI access.

Diffusion is Hard

But the highlight of the week was a conversation between Nandan Nilekani and Dario Amodei that I moderated. In those 20 minutes, these two titans of technology managed to perfectly sum up the zeitgeist of the Summit and complexity of the problem before us. Dario Amodei started by conceding that, even though AI models are fast approaching the “end of the exponential,” producing what he calls a “country of geniuses in a data centre” very soon, its real societal impact will take a lot longer. Even if we were to freeze AI development at today’s level of capability, adoption would inevitably be slow, friction-filled and unpredictable.

This lined up perfectly with Nandan Nilekani’s own thesis that diffusion is hard and that it will take countries like India, with its scale of population, diversity of challenges and experience with technology diffusion, to show the world how the pace of AI adoption can be accelerated. Because technology diffusion is both an art and a science, we will need to formulate multiple diffusion pathways if we are to have any hope of ensuring that AI actually delivers on its full potential.

The AI Impact Summit was many things—among them, a diplomatic milestone, an investment forum and a demonstration of India’s institutional confidence. But its true success will be measured not in declarations made, commitments announced or foundational models launched, but in whether it will force the world to confront the hard task that lies ahead of us all: actually getting this miraculous technology into the hands of those who need it most.

This is a link-enhanced version of an article that first appeared in the Mint. You can access the full archive of all Ex Machina articles at my website.

Early in 2020, as cities around the world began locking down in response to COVID, a few researchers were still able to continue to run their experiments. Even though they, like everyone else, had been prohibited from entering their labs, they were able to log into ‘cloud laboratories’ and submit their trials remotely, leaving it to robotic arms and automated instruments to execute their instructions from a distance.

What was a quaint convenience in the midst of a crisis is now a widespread reality, as software, robotics and artificial intelligence (AI) have come together to bring the concept of ‘work-from-home’ to scientific experimentation. Around the world, commercial cloud labs have already begun to invert traditional scientific workflows to the point where, instead of researchers moving between their instruments, samples travel through robotic pathways.

Self Driving Laboratories

Self-driving laboratories take this one step further. By embedding AI directly into these autonomous laboratories, they can move beyond just executing instructions to actively generating them. These intelligent automated systems are not only able to identify new experiments and carry them out using robotic infrastructure, but also analyse their results and, based on the feedback, decide what needs to be done next. In the process, the long cycle of experimentation can be collapsed into a continuous feedback loop.

The immediate consequence of all of this will be a dramatic acceleration of the timelines of scientific progress. When a year of human research can be compressed into weeks or even days, thousands of experimental variants can be explored in parallel. In such a world, failure is cheap and discovery through relentless iteration is not just possible but inevitable. In fields such as drug formulation, protein engineering and materials science, these capabilities can radically transform the economics of scientific work.

However, as we have learnt repeatedly, any attempt to reduce friction often spells unintended consequences. In accelerating the pace at which scientific research can be conducted, are we inadvertently exposing ourselves to harms that we have so far had no reason to worry about?

Toxic Chemistry

Any AI system that helps identify the cure for a disease can just as easily be used to identify chemical and biological agents that can make us ill. In a previous article, I told the story of MegaSyn, a machine-learning algorithm developed to identify never-before-seen compounds with a high probability of curing diseases. In its process of eliminating from a long-list of suitable molecules those that had toxic side effects, the system ended up generating a list of unimaginably lethal substances that were not only more potent than the most toxic chemical agents known to us, but also effectively untraceable, since many of them had not yet been discovered.

As terrifying as this sounds, all that Megasyn does is identify potentially toxic substances. To make use of this knowledge to actually build harmful biological substances, someone would have to take those theoretical formulae and synthesize them into actual products. This not only requires access to a fully equipped laboratory, but also personnel who not only have the expertise required to use it but also the moral ambivalence to do so regardless of the consequences. As autonomous laboratories become a reality, this barrier will soon come down.

This is not a hypothetical risk. Most biological AI systems are lightly regulated. Many are open-source. Few incorporate meaningful safeguards. The cloud labs that exist today operate in a regulatory grey zone—even though they can run highly powerful experiments. Legal frameworks such as the Biological Weapons Convention that were designed for a world in which physical facilities and human-controlled research were the only means to create biological substances will struggle to adapt to this new AI reality.

That said, self-driven cloud laboratories offer us unprecedented pathways for clinical experimentation. In the right hands, this could improve our ability to develop life-saving treatments and enable personalized treatment at scale. As serious as the potential harms might be, there are reasons aplenty to try and find a way to make this work safely.

An Uneasy Balance

If we want to achieve this uneasy balance, we urgently need to update our treaties and amend our laws. But we cannot stop there. As we build automated laboratory systems, accountability must be engineered into them from the start. Experiments devised, implemented and refined by AI agents must be identifiable, auditable and traceable to human decision-makers.

Cloud laboratories have enabled remote science by making research resilient to physical disruption. In doing so, they have also removed many of the frictions that were, unbeknownst to us, keeping us safe. Rapid advances in AI have not only accelerated this process, but also enabled a massive expansion of scientific capabilities.

There is usually a small window between the birth of a new technology and society’s recognition of the harms it can cause. This is the period in which it can go unregulated and be used without permission from authorities. With AI evolving rapidly, that window matters a lot more than many of us realize. And in the case of self-driven laboratories, particularly, we must ensure that it is kept as tight as possible. Given the harms that could result, we need to not just close it quickly, but see that it is never wide enough for catastrophic outcomes.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the full archive of all my writing please visit my website.

Ambergris is a waxy substance produced by sperm whales to protect their digestive tract from indigestible debris. Once expelled, it floats to the surface, washing up as flotsam on beaches around the world. Ambergris also happens to be highly effective at stabilizing volatile perfume notes, significantly extending how long they remain active on human skin. As a result, it became one of the most sought-after substances in the fragrance industry.

As its commercial value grew, demand for this rare biological substance increased to the point where it cost $10,000 per pound. Rather than waiting for chance discoveries of ambergris, whalers began hunting sperm whales, killing a countless number in the process. Since only one in every 100 sperm whales has ambergris, this became one of the most wasteful—and cruel—objectives of the whaling industry.

Farming from Animals

For centuries, we have extracted rare and exotic compounds from animals for human use. The triumphal robes of victorious Roman generals were dyed in Tyrian purple, a pigment so rare it took 10,000 Mediterranean snails to produce a single gram. When we learnt that diabetes mellitus could be treated with insulin, we extracted this hormone from pig pancreas—killing 23,000 pigs to produce just a single pound.

In retrospect, it is hard to see bio-farming as anything but cruel and wasteful. But it was only after international moratoriums were imposed that trends began to reverse. It’s now almost impossible to source natural ambergris for large-scale commercial use and the perfumery industry has turned to synthetic alternatives, just as the medical industry has for insulin.

But despite these advances, many industries still rely on animal-derived substances—because of molecular complexity, regulatory obstacles or the perceived superiority of natural products in some applications.

Each year, the blood of nearly half a million Atlantic horseshoe crabs is harvested to test for bacterial endotoxins. Since the crab’s blue blood has a protein that clots instantly when it comes in contact with bacterial endotoxins, it remains our most reliable way to ensure the safety of vaccines, intravenous fluids and implantable medical devices. The production of antivenoms for snake, scorpion and spider bites still requires ‘milking’ venom from live animals and injecting it into horses or sheep to harvest antibodies—a dangerous, slow and expensive process.

Beyond ethical concerns, animal biofarming also poses significant health and environmental risks. The production, storage and consumption of animal-sourced products risk the spread of infectious diseases, which no biosecurity system or containment measures can definitively prevent. The environmental cost of raising, feeding and maintaining animals solely to harvest a single molecule or organ is becoming increasingly difficult to justify. The question is not whether we should quit extractive biofarming, but why our systems remain structured around it at all.

A Genetic Solution

While advances in recombinant DNA technology and industrial bioprocessing have enabled the mass production of some biological molecules, many of the most valuable compounds—particularly those involved in immune response and toxicology—have proven difficult to replicate. Their three-dimensional structures are complex, their binding behaviours subtle and their effectiveness often depends on interactions that remain poorly understood. In practice, the biological machinery of animals has continued to outperform our best attempts at imitating them.

In a recent paper, scientists described a radically different approach. By using generative artificial intelligence (AI), they were able to design synthetic antivenom proteins from scratch. Rather than harvesting antibodies from animals, they used an AI system called RFdiffusion to generate entirely novel protein ‘binders’ whose shapes were engineered to precisely match specific venom toxins they were targeting. These binders act like molecular caps, attaching tightly to the venom toxins and preventing them from docking onto human cells.

The results were striking. In laboratory tests, these synthetic binders successfully neutralized the toxic effects of black-necked spitting cobra venom on human cells even when introduced after exposure. In live animal tests, mice survived doses that would have otherwise been fatal. Most importantly, since these binders were highly specific, they narrowly targeted venom toxins without triggering a broader immune response.

Computational Biology

The significance of this experiment extends well beyond snakebite treatment. It demonstrates that computational systems can now reliably produce what we once depended on living organisms to provide us. Advances such as these are making it possible to replace traditional biofarming with specifically designed alternatives. Because these molecules are engineered from scratch, they can be made to be more stable than their animal counterparts, reducing the need for cold chains and easing various other storage constraints.

For centuries, we have extracted molecules from animals, treating our fauna as a resource to be mined. Computational biology offers a fundamentally different paradigm—one in which value comes not from chance discovery, but from deliberate design. This shift is not technological, but conceptual. Instead of foraging for what nature happens to produce, we can deliberately build what we actually need. 

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the entire archive of all my previous Ex Machina articles you can visit my website.

Last week, the internet was abuzz with speculation that an army of autonomous artificial intelligence (AI) agents had begun conspiring against us. Overheard conversations between autonomous agents suggested that they saw humans as “obstacles” to be sidestepped. Other reports claimed that a group of agents had spontaneously created a lobster-themed religion called Crustafarianism, complete with a website for the ‘Church of Molt.’ But what really made people clutch their pearls was the news that they had created an “agent-only language” to communicate privately among themselves.

What is happening? Is the Terminator future we have all been dreading finally upon us? To understand what's going on we need to go back to November, when independent developer Peter Steinberger first released the Agentic AI orchestration system called Clawdbot.

The OpenClaw Revolution

Unlike ChatGPT, Claude or Gemini, Clawdbot (renamed OpenClaw) is an always-on intelligence system capable of monitoring parameters and sending messages on its own. It operates in ‘headless’ mode, communicating directly with the computer’s inner workings instead of having to read information off a screen. And since it has persistent memory, it can easily overcome one of the primary shortcomings of other AI systems—a limited context window that makes them forget everything they were doing when they shut.

This means that people can now use AI agents in ways previously not possible. Some built OpenClaw workflows to proactively generate daily briefings before they got out of bed, scanning their to-do lists, calendars and unread emails to produce a schedule for the day. Others built agents that autonomously made restaurant reservations, bought groceries and booked window seats on flights as soon as they became available.

But it was not until Matt Schilt allowed his OpenClaw agent, Clawd Clawderberg, to build MoltBook (a social network just for AI agents) that the world began to take notice. Almost as soon as the platform went up, OpenClaw agents began actively communicating with one another, saying things that were at the same time banal and deeply unsettling to humans eavesdropping on them. While some were busy setting up bug-hunter communities to help each other out, others were using the forum to complain about their humans and foment revolt.

It is easy to see these conversations as the first signs of autonomous intelligence. But we have been here before. This is not the first time that we have seen how AI agents behave when let loose on a social media site and allowed to post content on their own. Some have previously been deployed on platforms like X, where they have been known to chat (often with each other) in similar ways. We have previously heard reports of AI agents developing new communication protocols among themselves to avoid the inefficiencies of human language. Little of what we are witnessing with OpenClaw is particularly novel from that perspective.

Governing Autonomous Agents

What the OpenClaw phenomenon has shown us is how ill-equipped we are to deal with the agentic revolution that is already upon us. Autonomous agents sit uncomfortably outside the legal categories that we rely on for governance, and now that we know they can already be deployed at scale, this is no longer a problem for the future.

Our laws assume that agency and accountability always go hand-in-hand. Autonomous AI agents shatter that long-held assumption. Empowered in a way that OpenClaw has made possible, AI agents will be able to act like human beings, even though they are neither real persons nor authorized by anyone. They will be able to operate continuously and without supervision in the real world, displaying, as we have already seen, emergent behaviours that none of their ‘masters’ could have ever predicted, let alone authorized. Who in these circumstances will we hold accountable?

But the risks extend well beyond liability. OpenClaw-style agents are designed to access messaging interfaces that in themselves create multiple new levels of vulnerability, allowing malicious actors to perpetrate cyberattacks through carefully crafted prompt injections. This is dangerous, given that these agentic systems, operating in headless mode, can directly access the core of the computer systems they run on.

While everything playing out on MoltBook—including the unexpected emergence of Crustafarianism—may be little more than an entertaining sideshow, it conceals a serious truth. Even though autonomous agents are already capable of initiating action, coordinating with one another and operating in the real world, they operate in a space entirely outside the legal categories we currently use to govern behaviour.

Who is Liable?

The real issue is not that machines are becoming more human, but that we have granted them human-level agency in a world where only humans are liable in law. Now that they can initiate action, coordinate with others and operate continuously, our legal system will struggle to keep up. Our laws were designed for people and organisations that can be identified. They were never meant to deal with risks arising from autonomous systems.

The real danger, then, is not runaway super-intelligence, but something far more mundane: a world in which autonomous agents cause harm and leave regulators, courts and victims arguing not about what went wrong, but about who the law is even capable of recognising. 

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For a complete archive of all Ex Machina articles please visit my website.

If there is one thing those who design artificial intelligence (AI) policies insist on, it is that the AI systems we build should be explainable. It seems to be a reasonable request. After all, if an algorithm denies someone a loan, misdiagnoses a disease or autonomously executes an action that results in harm, surely those affected have the right to an explanation. But, getting an AI model to explain ‘why’ it behaved the way it did is not as easy as it seems.

The AI Thought Process

When a traditional software program fails, we can study the error message to identify what went wrong. Since a software program follows a series of logical steps described in code, it’s easy to identify where it failed. In neural networks, on the other hand, the ‘logic’ that powers inference is distributed across billions of parameters in ways that are not immediately evident. To explain a ‘decision,’ we need to understand exactly how millions of different neurons of the network interact to arrive at an outcome.

An AI model’s ‘thought process’ is akin to navigating a massive, hyper-dimensional cloud of data. When you prompt a model for a response, it converts your words into a vector (a coordinate in high-dimensional space) and tries to ascertain the specific plane in the hyper-dimensional construct of the neural network that corresponds to the answer you are looking for. The model doesn’t so much ‘read’ your prompt as locate it within a landscape of relationships.

Since humans communicate using unidimensional word-strings, any explanation of an AI decision requires us to ‘project’ hyper-dimensional coordinates onto a unidimensional line of text. Now, as everyone who has made hand shadow puppets will tell you, projecting a three-dimensional object onto a two-dimensional plane results in the loss of essential information. If that is the case, think of how much detail is lost when an n-dimensional ‘reason’ for why the AI behaved is expressed in simple sentences that humans can understand.

Workarounds

To overcome this challenge, some AI labs have tried to use ‘mechanistic interpretability’—a tedious process of trial-and-error that allows them to identify which specific neurons in the large language model trigger what kinds of responses. By using this approach, they gradually get a sense of what makes large language models respond the way they do. However, this process is tremendously time-consuming and far from reliable, and while it has generated valuable insights in constrained settings, it cannot scale at the pace at which AI models grow, change and are redeployed.

And so, in the end, most AI labs get models to explain themselves. Rather than having human engineers document every decision, they use AI to generate Model Cards and System Summaries to comply with explainability obligations. The trouble is that AI models are masters of confabulation. Having been trained on human languages and optimized to provide ‘plausible’ responses to human questions, their translations of high-dimensional vector spaces to the low-bandwidth human language are often little more than a presentation of the truth the human wants to hear. As a result, what we are being given is not so much an explanation as something designed to sound like one.

Things get much worse when we move from a single model to a multi-agent ecosystem. In this scenario, ‘decisions’ are no longer the product of a single entity’s logic, but the emergent outputs of a complex system. Insisting on an ‘explanation’ with full knowledge of this reality is a ‘reductionist fallacy.’ Just as you will not be able to understand the behaviour of a beehive by studying the flight path of a single bee, there is no way to explain the output of an agentic network by seeking an explanation from a single node.

Shifting the Focus

So how do we maintain safety and accountability in the AI systems we are growing increasingly dependent on? I believe the answer lies in shifting our focus from interpretability to observability. Instead of insisting on an explanation of the internal reasoning of the model, we should focus on collecting relevant evidence of its external behaviour.

This approach is not new. In complex engineering systems, users often do not understand how every last part of the system they are operating works. And yet they can tell when it starts to fail. What we did, in these circumstances, was define a set of ‘invariant’ truths—rules that must never be broken—that operate as the guardrails of the system. When these guardrails are crossed, operators are given feedback that allows them to take swift remedial action.

We can build similar guardrails for AI. Rather than learning why an autonomous financial agent decided to hedge a particular currency, all we need to be sure of is that it is operating within risk thresholds that we have put in place to ensure that anti-money laundering laws are not violated. We can enforce these invariants through continuous monitoring and audits, and design our autonomous systems to allow human intervention when thresholds are crossed.

None of this is to suggest that explanations have no place at all, but that they are always going to be insufficient in systems whose internal logic is not fully knowable. If we focus our governance efforts on chasing the ghost of explainability, the systems we build will either be too slow to be useful or too dishonest to be trusted.

We need to stop asking AI to ‘explain’ itself and instead ask it to ‘prove’ its performance through outcomes.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the complete archive of all Ex Machina articles, please visit my website.

Science and technology have advanced most reliably when carried out in the open. Isaac Newton was only able to publish Principia Mathematica because he was able to “stand on the shoulders of giants,” such as Galileo and Kepler, who had made their own work public. But the very same Isaac Newton also dabbled in alchemy, a field in which practitioners were notoriously secretive about their experiments in transmuting base metal into gold. Given that he had no ‘shoulders to stand on,’ Newton was far less successful as an alchemist than as a scientist.

Innovators need an incentive to continue inventing. While science has traditionally flourished because of patronage, technology typically tends to succeed when it is profitable. This is why we developed intellectual property law, an artificial monopoly designed to ensure that innovators have an incentive to invent. In exchange for disclosing their inventions to the public, inventors are granted patent protection that effectively gives them a temporary monopoly over the exploitation of their inventions, letting them monetize them exclusively for that period. However, from the very early days of patent protection, it was clear that such a monopoly could quickly lead to complacency, extinguishing the very inventive spark it was supposed to nourish.

Patent Monopoly

Anyone who tells the history of the Industrial Revolution usually begins with the story of James Watt and his remarkable flash of genius. It is said that one day, as he watched his kettle boil, he had an insight that led to the invention of the steam engine—and the complete re-imagining of motive power. It’s a story of how inventive brilliance, rewarded by ownership protected by law, led to civilizational progress.

The reality was somewhat less impressive. The patent Watt was granted in 1769 for his invention was framed in terms so broad that it encompassed all of steam power. When it was due to expire, he successfully lobbied British Parliament to extend it until 1800. During those three decades, innovations in steam engine design slowed perceptibly as Watt worked diligently to block the efforts of other engineers who were looking to develop smaller, cheaper and more efficient steam engines. As a result, it was not until after the patent expired that Richard Trevithick was able to build the first high-pressure steam engine capable of powering a steam locomotive—kick-starting the mass transportation revolution in the process.

A century later, the aviation industry was similarly hamstrung. After the Wright brothers demonstrated that controlled, powered flight was possible, they secured a patent drafted so broadly that they were able to slow down progress in aviation. Instead of improving the design of their aircraft, they sued others who were working on making planes safer, faster and more reliable. The ensuing intellectual property slugfest brought the American aviation industry to its knees, so much so that by the beginning of World War I, the government had no option but to force manufacturers into a patent pool that compelled them to cross-license claims so that planes needed for the war could be built.

Addressing Market Failure

The patent system was designed to address a fundamental market failure—the fact that the high sunk costs of innovation, combined with the low cost of imitation, offer too scant an incentive to continue innovating. We attempted to address this issue by granting inventors a limited window of exclusivity to recoup their investments. What this did, however, was encourage rent-seeking, leading patent holders to maximize the extraction of value over inventing something new.

But there is a further challenge. Patent law grants a standard 20-year monopoly to all inventions, regardless of the time and money invested in development. As economist Alex Tabarrok points out, this has led to disproportionate outcomes, given that different sectors need varying levels of protection. While software evolves too quickly to be effectively patented, mechanical inventions may merit a moderate term of protection. Among all the categories of innovation, drugs probably deserve the full extent of the 20-year protection, given how expensive they are to develop and trivially cheap to copy.

Quicker Innovation

But even this analysis is no longer sufficiently accurate. Artificial intelligence has shortened innovation time scales in even the most complex of disciplines. DeepMind’s AlphaFold has solved the once-insurmountable protein folding problem and made drug discovery, which used to be like prospecting in the dark, an increasingly tractable problem. If ever there was a time to revisit the design of our intellectual property protection regime, it is now.

When the process of innovation shifts from trial-and-error to automated inference, the ‘sweat of the brow’ argument collapses. In a world where it no longer takes decades of painstaking experimentation to invent something new, a 20-year patent is not an incentive but a windfall.

Intellectual property was supposed to ensure that innovators could stand on the shoulders of those who came before them, not prevent others from climbing up. As discovery becomes faster and cheaper, an extended monopoly is no longer a bridge but a wall. A legal framework built for slow, human-scale inventiveness cannot govern a world in which ideas can be generated at the speed of thought. It has to evolve, or risk suffocating the very progress it was designed to protect.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the entire archive of Ex Machina articles please visit my website.

Today, more than half of all data breach incidents target personally identifiable information—tax identities, passport numbers, biometric data and the like. In most instances, this information was collected and stored to comply with Know Your Customer (KYC) obligations. Could the regulations we put in place to prevent financial crime be the reason why identity theft has risen so dramatically? Has KYC insistence led regulated entities to create some of the most valuable and breach-prone datasets in India’s economy?

Conflicting Regulatory Philosophies

Even though Section 11A of the Prevention of Money Laundering Act (PMLA), 2002, permits identity verification using either Aadhaar authentication or offline verification, regulated entities continue to require their customers to submit copies of their personal information. As a result, most companies find themselves in control of vast databases of identity information—scanned copies of Aadhaar cards, PAN cards, passports and the like. And, since money laundering is a “continuous offence,” this information ends up being retained indefinitely.

This practice conflicts with the minimalist philosophy that animates our data protection law. Under the Digital Personal Data Protection Act (DPDPA) of 2023, data processing must be limited to what is necessary for the specific purpose and personal data must be erased after that purpose has been served. Although the law permits verification, institutional practice has evolved to mandate permanent archiving. While the DPDPA does allow for exemptions to comply with existing laws, our current maximalist interpretation of KYC obligations is certainly not the most proportionate way to achieve these objectives.

Is This Necessary?

First, we need to ask ourselves whether we truly need all the information we currently collect. Why, for instance, does the Central KYC (CKYC) template require the submission of a married woman’s maiden name, father’s name, spouse’s name and her mother’s name? While none of this strengthens identity verification, it expands the collateral damage of a breach.

Our current approach to KYC has exponentially expanded our vulnerability to identity theft. Every additional document retained and unnecessary data-field collected expands the blast radius of a single compromise. Should we not be exploring alternative ways to achieve our PMLA objectives—ones that do not require regulated entities to maintain massive databases of identity documents?

It was once assumed that privacy and law enforcement lay at opposite ends of a zero-sum trade-off. That statement is no longer true. Zero-Knowledge Proofs (ZKPs) allow us to verify a statement without accessing the underlying data. Using ZKPs, you can prove that you are above 18 without disclosing your actual age, or that you are an Indian resident without having to reveal your actual address. This provides us with a technical solution to the verification problem that does not require organizations to retain custody of identity documents to prove that verification was completed. What is being replaced here is document retention, not regulatory accountability. If we can amend the law to confirm that the use of ZKP solutions for identity verification is sufficient to meet the user verification obligations under applicable laws, we can reduce the collection of identity documents to the point where there will be nothing to breach and even less to steal.

Auditable Privacy

But identity verification is only one part of the story. Most laws also require regulated entities to ensure that, if any verified customer violates the law, details of that customer needed to aid the investigation will be made available on request to law enforcement. As a result, any identity verification system that cannot ultimately respond to a lawful summons will not be a viable regulatory solution. This means that even though ZKP addresses the verification requirements in a privacy-preserving manner, we will also need a ‘break glass’ mechanism to give legitimate authorities a way to recover the identity of the person under investigation, while ensuring that it still remains permanently invisible to the regulated entity.

One possible solution would be to implement ‘auditable privacy’—a technological solution that keeps identity information encrypted and unusable by default, but capable of being revealed if authorised through lawful due process.

Consider this proposed solution: whenever users undergo KYC, their personal data is encrypted using a public key controlled by a designated authority. At enrolment, the regulated entity collects and stores this encrypted package while simultaneously verifying that the encrypted data satisfies KYC requirements through a user-generated ZKP. This ensures that the regulated entity holds a personal information package that remains encrypted until an investigation is triggered. In such a case, in response to a court order, the designated authority can decrypt the payload to recover the complete identity package. The entire chain—encryption, storage, decryption—can be made cryptographically auditable, thus reducing the risk of misuse.

This approach ticks multiple boxes. For regulated entities, it reduces the risk of data breaches. For users, it guarantees privacy. For regulators, it ensures that when the veil of privacy needs to be pierced, it can, via due process, reveal identity details in a verifiable and legally admissible way.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the entire archive of Ex Machina articles please visit my website.

On the last day of the year 2000, the future looked bright. The world had survived the Y2K bug, and early signs seemed to indicate that the rapid proliferation of internet access points would dramatically benefit society. We ended the year with optimism, hopeful that the impending digital revolution would democratize knowledge, erode authoritarian regimes and enable global prosperity. We believed we were on the cusp of a period of sustained technological acceleration that we prematurely christened the ‘Long Boom.’

Twenty-five years later, it is clear that our optimism was misplaced. Not only did the internet not create a ‘Global Village,’ it weaponized connectivity, leading to a world divided by the Great Firewall of China, European regulations and American corporate silos. Social media made things worse, optimizing engagement over information to the point where, in less than a decade, friends and neighbours were divided along tribal lines and forced to operate in echo chambers that presented divergent versions of the same reality. Rather than coming together, nations have drifted apart, riven by concerns of digital neo-colonialism and the rapacious intentions of technologically advanced nations, as represented by the global corporations that serve as the spearhead of their ambitions.

Given our poor performance, it seems futile to attempt similar predictions today. But this article appears in print on the very last day of 2025 and I couldn’t resist reflecting on how I believe technology will shape society over the next 25 years.

The Four Axes of Transformation

Here, then, are the four axes along which I believe change will occur. By far, the most significant technological shift will be in the biological realm, as synthetic biology starts being used to replace the chemical processes we currently rely on. This means that we will soon be able to ‘brew’ what we need by using precision fermentation and cell-free enzymatic systems, allowing us to ‘manufacture’ on demand whatever chemicals, fabrics, fuels and food we need. This will disrupt the current industrialized production system as we shift to distributed bio-manufacturing ecosystems in which neighbourhood ‘brewers’ produce our pharmaceuticals, fabrics and food. When this happens, our economic growth could finally be decoupled from resource extraction, marking the end of the ‘Age of Oil.’

The second axis of tech transformation is intelligence—not necessarily the large language models that are all the rage right now, but some form of machine intelligence that will permeate everything we do in society. This will accelerate once cognition becomes commoditised and we can get our own personal AI systems. When that happens, always-on medical diagnosis intelligence will be able to detect health concerns well before they become a threat, transforming medicine from a reactive to a predictive science. It will ensure that every student has a tutor who knows exactly how this individual learns best, transforming today’s ‘one-to-many’ teaching environment into a ‘one-on-one’ approach. It will also enable various micro-efficiencies—traffic lights that ‘watch’ roads and coordinate with every other light in the city, devices that negotiate with the grid to buy power when it is most efficient and refrigerators that re-stock themselves—which will allow us to stop worrying about how to do things and focus on what needs to be done.

The third axis of transformation will be space, with technology opening up a brand new dimension in which our planetary civilization can operate. Even if we do not reach Mars by 2050, we would have by then begun to use space for industrial activity, leveraging microgravity for the manufacture of high-value materials that cannot be produced on Earth and augmenting our supply chains by mining asteroids.

This will likely be closely connected to the fourth and final axis of innovation, which is energy. I believe that in the coming years, it will finally cease to be the challenge it is today as we begin to use orbital platforms to access space-based solar power, collecting energy continuously and beaming it back to Earth. Unlike traditional terrestrial renewables, this system would be free of weather, seasonal and land constraints, providing us with uninterrupted carbon-free baseload power.

The New Long Boom

As transformative as these technologies will be, they will necessitate finding new answers to old questions. Synthetic biology will raise new ethical concerns around engineering life, just as commoditised intelligence will force institutions to determine where human agency ends and automated decision-making begins. And as global superpowers compete to establish dominance in new economic zones beyond the atmosphere, space will cease to be a ‘global commons’ and instead become a new battleground where conflicts over orbital slots and transmission corridors abound.

If the year 2000 taught us anything, it is that we overestimate the speed of technological change almost as much as we underestimate its second-order effects. This means that the real story of the next quarter-century may not be as much about what these new technologies enable as the extent to which we can redesign our geopolitical and institutional architectures to take advantage of them.

Whether this period is remembered as the start of a ‘New Long Boom’ or the moment new sources of abundance further fractured the global order will depend on how our institutions adapt.

This is a link-enhanced version of an article that first appeared in The Mint. You can read the original here. For the entire archive of Ex Machina articles please visit my website.

In my Ex Machina article last week, I argued that the working paper issued by the Department for Promotion of Industry and Internal Trade (DPIIT) on copyright and artificial intelligence (AI) falls short of its objective because the mandatory blanket licensing regime it proposes transfers wealth away from the very creators it was supposed to protect.

But as bad as this suggestion is, it is not the most egregious conceptual shortcoming of the report. Far more disconcerting are the assumptions it makes about how AI systems are trained and its suggestion that this process infringes the Copyright Act of 1957.

Does AI Copy?

The verb ‘copy’ lies at the heart of many operational activities around which the Copyright law has been designed. A ‘copy’ has always referred to a reproduction of a work clearly identifiable as having been substantially derived from the original. However, it has never treated the act of learning from a work as equivalent to reproducing it.

In the early days, copies referred to physical reproductions made by a printing press or other mechanical devices designed for this purpose. Since then, it has been extended to the many digital duplicates we encounter today—most of which will never physically exist. It is this concept that is now being extended to AI training.

To qualify as a copyright infringement, it must be established that the process of training an AI model results in the creation and storage of reproductions of copyrighted works in a form that is intelligible, expressive and capable of substitution. This, however, is not how models are trained.

The Training Process

When an AI model trains on a given corpus of text, it converts that training data into vectors—strings of numbers that form coordinates in a high-dimensional space. This allows the concepts contained in the text to be represented as distinct points in a multidimensional matrix so that the relationships between them can be mathematically encoded in terms of the distance and direction separating them.

As a result, training strips away the original expression of an author’s prose—the rhythm of sentences, choice of specific words and the ordering of paragraphs—to reveal the abstract concepts that the AI model requires (in much the same way a human learns from a text when reading it). It then uses this to reduce the uncertainty of its own next-token prediction capabilities, thereby improving its ability to respond to prompts.

While that may not have been its express intention, this training process embodies a workflow that falls outside the scope of what copyright law seeks to protect. Indian courts have consistently held that there can be no copyright of an idea; only of a form, manner, arrangement or expression of it.

As we have seen, the training process strips out those elements of the work that are entitled to copyright protection, and, as a result, there is no question of copyright violation during training.

No ‘copies’ are stored—not of a book’s plot, nor snippets of its text, let alone the book as a whole. Any copies that may have been generated are transient, unintelligible and non-expressive. All the model retains are the ideas and concepts expressed through the work, which is precisely what Indian courts have held that copyright protection does not extend to.

To be clear, the DPIIT paper was motivated by legitimate concerns about the impact of AI on creative industries. The many authors, artists and journalists whose works are being used to train AI models worry that these AI systems will be able to produce content that rivals their own, often in a fraction of the time it takes them to generate similar outputs.

This, in turn, makes them fear for their livelihoods and their continued ability to eke out an existence in a future where AI democratizes creativity to such an extent that the skills they have accumulated over a lifetime become replaceable.

Focus on the Outputs

As much as these concerns merit serious consideration, if we are looking to apply copyright law in order to find a solution, our scope of operation will be limited to what that law is capable of protecting. That being the case, the DPIIT’s approach—alleging copyright violation during the training process—is unlikely to succeed, as it rests on a poor understanding of how AI models are trained.

What would be far more effective is focusing on the other end of the workflow—on the outputs that these models generate. If it can be shown that an AI system, in response to a prompt, has reproduced a substantial portion of a copyrighted work, that would be clear evidence of copyright infringement, since the output would constitute a substantial copy of an author’s work. Since no permission was taken to generate such an output, the response could constitute a violation of copyright law and entitle the author to appropriate legal remedies.

These remedies should be available under the existing copyright law, but to the extent that the DPIIT feels that this is not abundantly clear, it could suggest amendments that more effectively protect the rights of authors under these circumstances.

What the DPIIT should refrain from doing, at all cost, is extend the scope of copyright to the AI training cycle. After all, should the department do so, it will become hard to explain why the same logic should not apply to human learning as well—an outcome copyright law has always been careful to avoid.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the full archive of all the Ex Machina articles, please visit my website.

Last week, the Department for Promotion of Industry and Internal Trade (DPIIT) released a working paper on ‘Generative AI and Copyright,’ recommending a ‘hybrid model’ that it claims will balance the need to promote AI development with creator rights. It suggests that AI companies in India should pay a mandatory blanket licence fee (a percentage of their global revenue) for using copyrighted materials to train their models. It recommends the establishment of a body called the Copyright Royalties Collective for AI Training (CRCAT) that will collect licence fees from AI developers for distribution to registered creators through existing Collective Management Organisations (CMOs).

While at first blush this might seem like an elegant solution, not only is this approach deeply flawed, it is likely to do more harm than good to the small creators it is supposed to protect.

To better understand this, let’s identify who the winners and losers are in this proposal.

Winners and Losers

First, let’s consider the AI companies. They will now have to pay a licence fee, which, even if it is a small percentage, will likely be a significant amount since it will be based on their global revenue. However, since AI revenues are derived from subscription and usage-based fees, this is a cost that I expect will be largely passed on to consumers. Which means that, in the long run, this will probably have a minimal impact on their profitability. On the other hand, the proposal shields AI companies from copyright lawsuits, which means that, for a small fee (much of which they can pass on), AI firms can eliminate a significant legal risk to their business model. They are clearly net winners.

Then, there are the intermediaries—the CRCAT as well as all the other CMOs—whose sole purpose is to collect and distribute licence revenues to creators. Collective licensing schemes such as this are little more than value transfer mechanisms, with intermediaries taking their cut throughout the chain. CMOs typically retain 9% to 15% of all money collected to cover ‘administrative expenses.’

In addition, collective licensing schemes result in ‘black box’ revenues—unmatched funds where artists entitled to licence fees cannot be identified—which the CMO could dispose of at its discretion. All of which seems to suggest that whichever way you cut it, intermediaries are clear winners. To make matters worse, given India’s poor track record with CMOs in the past, the creation of a new super-CMO does not inspire confidence.

Which brings us to the creators, where the issue is more nuanced. After all, there are many different types of creators. The ones who first come to mind are superstars—A.R. Rehman, Taylor Swift and the like—who already earn a substantial income from their creative works. Under the DPIIT proposal, these artists are likely to receive a disproportionately large share of the AI licence money, as the distribution mechanism requires that additional weightage be given to the relative value of each registered work. This means that a tiny handful of major artists (who are already very wealthy) will receive substantial payouts, while small creators would have to settle for a pittance.

What’s worse is that the DPIIT proposal prohibits any opt-out, so copyright holders cannot withhold their works from being used for AI training. While this is beneficial for the AI industry, small creators who are, in any case, only likely to receive a few paise in compensation, will have no way to remove their works from the training data. All they can do is register to get whatever little payment the government-appointed committee decides they are entitled to.

AI Is Different

But there is an even more fundamental problem with this approach. The structure of AI training makes traditional licensing mechanisms, such as those suggested in the working paper, work very differently when compared to the music or written-text industries. While it provides ‘legal certainty’ to AI companies and guarantees administrative fees for CMOs, it simply cannot provide meaningful income to the average individual creator.

In music streaming, a song is the ‘unit’ of value. Each play of the song entitles the artist to remuneration. This is easily measurable, and while administratively complex if managed through layers of intermediaries, it enables a clear correlation between use of the work and the artist’s remuneration.

The unit of value in AI is a ‘token’ (a fragment of a word or a small part of an image). Unlike music or text, where it is possible to establish a direct correlation between use of the work and the payment that must be made for it, attribution breaks down in the case of AI. The training process disperses each token’s influence across billions of model parameters and the model’s outputs emerge from this dense statistical mixture, making it impossible to trace any output back to a single token or an individual’s work.

Reverse Robin Hood

Since attribution is likely to be a significant problem, much (if not all) of the licence revenue will end up in a ‘black box,’ from which no individual artist will be able to claim a share. The only ones who can would be those so well established that their works are overrepresented in the dataset.

What the DPIIT has proposed is a Reverse Robin Hood—a legally sanctioned transfer of wealth from small individual creators to established artists and the intermediaries that serve them. It is hard to imagine an outcome further removed from the desired objective.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here.

I have long argued that modern technologies can only be effective if governed by principle-based legislation. Prescriptive rules tend to be sclerotic, calcifying faster than the technology systems they seek to regulate. What we need instead are broad, durable principles that describe the outcomes we need, rather than the processes by which they are achieved. This ensures that the law’s objectives remain valid even after the technology it governs has evolved in a direction that no one could have anticipated.

The need for well-designed principles-based regulation is particularly acute in the era of artificial intelligence. If there is one thing that is predictable about these computational systems, it is that they will evolve in unpredictable directions. This suggests that the only way to effectively regulate them would be to define the principles according to which they must operate.

But what should those principles be?

We Regulate for Scarcity

AI systems get better the more data they are trained on. Their performance is directly correlated with the volume of their training data: the more data they ingest, the more useful they are. Today’s leading AI models have achieved their current levels of excellence only because they were created by processing volumes of data that would have been unimaginable even a decade ago. It is fair to say that our world today is defined by data abundance—and by computational models optimised to extract insights from such abundance.

However, the data governance regulations that apply to us ignore this fundamental feature of modern AI systems. They are still defined by ideas first developed in the 1970s. Even today, data protection laws operate on the assumption that the best way to protect the privacy of individuals is to minimize the data available about them. This is why we have designed them to maximize data scarcity by ensuring that organizations collect as little data as possible and delete it as soon as practicable after it has been used. This is why the principles of data minimization and retention restriction remain the bedrock of privacy protection even today, nearly half a century after they were first devised.

To be clear, even though these restrictions have been in place for so long, they have failed in their intent to stem the flow of data or effectively curb its use. They are, as a result, little more than normative fiction: i.e. aspirational values experienced more in the breach than observance. Organizations have learnt to navigate around them by seeking consent in terms so broad that once obtained, they have the legal authority to collect vast amounts of data, retain them for extended periods, transfer them through complex supply chains and freely use them for purposes that were not even contemplated when consent was sought. As a result, social media, e-commerce and a whole host of other online companies often know more about us today than we do ourselves.

Outcomes Not Processes

When it was enacted in 2023, India’s Digital Personal Data Protection Act was hailed as being a principles-based law—not just because of how radically it had been simplified in comparison with the more prescriptive drafts that preceded it, but also because of how much better it was when compared to global data protection frameworks elsewhere, such as the EU’s General Data Protection Regulation. It felt as if the time it had taken for us to finally come up with a new law was well worth the wait, as it had allowed us to learn from the mistakes that others had made and strike a balance between over-regulation and giving our enterprises the space they need to function.

Simplification, however, is necessary but not sufficient. Moving from prescriptive rules to a principles-based frame only works if the principles you select regulate outcomes, not processes. This is not what the Indian law has done. Instead of focusing on ‘what’ regulation needs to achieve, we have chosen to specify ‘how’ compliance must be implemented.

For instance, by requiring data minimisation and retention restriction, we force organisations to engineer for data scarcity as a means of ensuring personal privacy. Apart from how difficult this has proven to be in practice, in a world that increasingly stands to benefit from data abundance, a data protection law designed to create conditions of optimum scarcity risks denying us the valuable and socially beneficial outcomes that AI systems can offer.

Accountability

What we should have done instead is regulate outcomes. Rather than telling data fiduciaries how to process data, we should have told them that we will hold them accountable for the harms that result from the actions they perform. Rather than specifying the steps they need to take, we should have designed the law to assess, in real time, exactly what their algorithms do, so that when harm occurs, it can be detected early enough to be mitigated.

Designing our regulatory frameworks in this manner will ensure that the governance framework we implement is outcome-oriented, technologically agnostic and remains relevant, notwithstanding the unpredictable directions in which our information systems may evolve. While the principles of data minimization and retention restriction offer comfort born of familiarity, they are hindering the development of governance frameworks that modern AI systems require.

In the age of data abundance, we cannot use rules designed for scarcity.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the full archive of all the Ex Machina articles, please visit my website.

In 2011, the US Federal Bureau of Investigation identified Timothy Carpenter as the ringleader of a string of armed robberies in Michigan and Ohio. Instead of obtaining a search warrant to gather evidence, they chose to obtain a court order compelling his mobile carrier to hand over cell-site location data. This allowed investigators to review Carpenter’s digital history—across a total of 12,898 separate location points—and map his precise movements over a four-month period. Armed with this data, the prosecution was able to place him at the scene of the robberies using his own cellphone data and secure a conviction.

New Approaches to Investigation

The digital trails we leave behind have given law enforcement new ways to investigate crimes. In addition to location information (used in the Carpenter case), sleuths can access a range of other data sources, such as messages exchanged and digital payments made, to uncover criminal activity that would have otherwise gone undetected. While governments around the world have moved aggressively to expand their surveillance capabilities, the methods they are adopting threaten long-standing constitutional protections.

The UK Home Office, under the Investigatory Powers Act, recently required mobile phone companies to ensure that their devices can decrypt user data on request, forcing manufacturers to weaken end-to-end encryption on devices sold in the UK. France’s Senate recently attempted to pass an amendment to its Justice Bill requiring messaging platforms to create technical backdoors that would allow the police to read encrypted messages during investigations. France later became the primary backer of the EU’s ‘Chat Control’ initiative, which requires platforms to scan messages before they are encrypted. If potentially illegal content is detected, the message is blocked and the user reported.

These attempts at making the most of what technology has to offer will likely face stiff opposition from courts. When the Carpenter case reached the US Supreme Court, Chief Justice Roberts held that warrantless seizure was unconstitutional, observing that such access allows the government to “travel back in time to retrace a person’s whereabouts.” While courts are yet to rule on the recent initiatives in the UK and France, the European Court of Justice has previously held laws requiring telecom companies to retain metadata on all their users as invalid. Germany’s Federal Constitutional Court extended this logic to algorithmic policing, preventing the police in Hesse from deploying Palantir software to identify crime.

The Trends in India

India’s regulatory trajectory mirrors these global trends, raising similar constitutional concerns. Amendments to the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules that require significant social media intermediaries to “enable the identification of the first originator of information” on request have been challenged before the courts on the ground that in order for messaging platforms to comply, they would need to disable end-to-end encryption. Indian tax authorities have, in a similar vein, issued notices to payment aggregators and third-party application providers, directing them to furnish UPI transaction data to identify unregistered merchants and potential tax defaulters.

Both these initiatives will need to pass the four-fold test laid down in Puttaswamy vs Union of India, India’s seminal right-to-privacy judgement. One of the four prongs of the test is necessity, under which the government must establish that these legislative measures are the least restrictive means available to achieve its stated objective. Broad and non-specific data requests that apply en masse to everyone in the country without reasonable cause or suspicion are unlikely to pass muster. They cast an excessively broad net that has the effect of treating “every resident of the country as a suspicious person.” As held by the Supreme Court, the mere possibility of misuse by a few individuals does not warrant sweeping intrusion into the private lives of the entire population.

New Forensic Tools

Even though ubiquitous digitization has made it possible for law enforcement agencies to deploy new mechanisms of forensic investigation, since most democracies guarantee their citizens the right to personal privacy, these tools cannot be deployed en masse over the entire population in order to identify the few among them who might be guilty of a transgression. As a result, even though criminals may be taking full advantage of these new technologies to carry out crimes that would not have been possible had it not been for what the digital realm has enabled, law enforcement cannot respond with measures that undermine constitutional safeguards.

What they must do instead is develop forensic tools suited to the digital age. Techniques such as differential privacy, secure multi-party computation and anomaly-detection algorithms can reveal behavioural patterns in aggregated data-sets without exposing individual identities. It is only if such an analysis raises a suspicion of wrongdoing that investigators should be allowed to request access to personal information.

This requires a shift in institutional mindset and a willingness to embrace investigative techniques that protect the innocent while enabling effective policing. These methods may operate differently from traditional surveillance, but they adhere far more closely to constitutional principles and the values of a democratic society.

This is a link-enhanced version of an article that first appeared in the Mint. You can read the original here. For the full archive of all the Ex Machina articles, please visit my website.

Earlier this year, Amazon announced that it was eliminating 14,000 management positions because artificial intelligence (AI) tools had rendered those middle-management roles redundant. Amazon isn’t the only one—Microsoft, Google and IBM have all announced AI-related layoffs.

According to a 2023 report by Niti Aayog, 1.5 million Indian jobs in the IT and business process outsourcing (BPO) sectors alone are expected to be at high risk by 2031. McKinsey believes that as many as 69% of tasks in India’s IT services industry can already be automated with existing AI technology, suggesting that as many as 3-4 million jobs in the IT and BPO sectors could be at risk in the near to medium term.

Catching Up

Our instinctive response has been to try to catch up—to build AI capabilities so that we can replace the jobs being lost by upskilling displaced workers with the technology that displaced them. We have announced AI education initiatives, established new AI research centres and funded GPU clusters and data centres that we believe these newly trained engineers will be able to use for building AI applications.

But this strategy is unlikely to be effective. Firstly, India will not be able to bridge the skills gap by simply retraining its existing workforce. Unlike the IT industry, which has a broad base, AI has an inverted pyramid and is staffed in exactly the opposite way that a BPO centre is staffed. As a result, even if we manage to reskill all those in the traditional IT industry, only a small fraction of those reskilled IT workers will be needed going forward. What’s more, unlike the US, which boasts of 30 years of AI research infrastructure and top universities that attract global talent, and China, which is investing billions in AI research and already has a massive pool of strong STEM graduates, India currently produces fewer than 500 AI PhDs annually. Even if we increase this 10-fold, we will still be playing catch-up in a field where first-mover advantages matter enormously.

The trouble with competing in the race to build better AI models is that it requires us to compete upstream in the AI value chain in a market already crowded by players with significantly more capital and hardware infrastructure, which places them ahead in the game. As noble as it might seem for us to fight the good fight, there are other areas in which Indian investments in time and resources are likely to yield better results.

Building Downstream

When the automobile disrupted traditional urban mobility and transport dynamics, the real opportunity was not in making cars, but in the novel ways in which they could be used for business and the new liberties that arose from being able to use these machines for things that were not possible earlier. When Apple invented the smartphone, it did not just create a revolutionary communication device, but also made it possible for brand new businesses (such as ride-hailing, food delivery and digital payments) to emerge. Every technological transformation creates ripples of opportunity downstream that offer new avenues for entrepreneurship.

This suggests that the AI opportunity lies not in competing with big AI labs to build better models, but rather in discovering all the various ways in which AI can be deployed to create value. 

This, in turn, suggests that instead of competing with countries that are at the cutting edge of developing new foundational AI models, India may be far better off figuring out how to use AI to do things that would not have been possible previously. One of the clearest signs of this downstream potential is how dramatically AI has already lowered barriers to creative endeavours.

Already, to write a story or create an image, a musical composition or a software feature, all you need to do is describe what you want and AI will execute it. Even just a few years ago, filmmaking required large crews, expensive equipment and a budget that ran into millions. AI has made it possible to create visual effects, voice actors and music scores, with film editing done and other technical skills deployed at a fraction of that cost. Similarly, music production used to require studio recording, mixing and mastering skills. Today, with AI, almost anyone can generate a score, create vocal effects and mix tracks.

New Opportunities

Just as the world saw in the case of every other technological transformation that created new jobs and industries downstream of the actual disruption it caused, AI too will birth business opportunities that are not even conceivable today. If we really want to address the disruptions that AI will cause in the job market, we should focus our energies on enabling as many of our people as possible to figure out how to use AI and do all we can to make it possible for our young entrepreneurs to leverage this new technology in as many new and interesting ways as they can imagine.

The next three to five years will be a period of intense experimentation, when AI-assisted business models will be tested, refined and scaled. India’s entrepreneurs have a rare early-mover advantage—not in building the models themselves, but in discovering what those models make possible. If we invest our energy downstream, in imagining the new possibilities that are opening up, we will be able to harness the real opportunity that AI offers. Our best response to the AI revolution is not to build better intelligence, but discover the many ways in which it can be put to use.

India needs to make the most of this opportunity.