Volker Türk is right to demand “cast-iron” safeguards for advanced AI, says Vendan Ananda Kumararajah, but the harder question is whether regulators can remain sufficiently independent and knowledgeable when much of the expertise needed to understand frontier systems sits inside the companies they are meant to govern

The United Nations human-rights chief, Volker Türk, has called for “cast-iron guarantees” around advanced artificial intelligence. His warning is difficult to dismiss.

Türk has warned that advanced AI could pose an existential threat to humanity. He has called for international red lines and independent verification, while also warning that extraordinary power over AI is increasingly concentrated in the hands of a small number of technology companies.

The obvious response is to demand stronger regulation. That’s necessary, but it raises a harder question: do the institutions responsible for regulating advanced AI possess enough independent knowledge to understand what they are governing?

The issue increasingly extends beyond regulatory competence to dependence on the companies being regulated for the technical knowledge required to oversee them.

The organisations developing the most advanced AI systems possess extraordinary concentrations of technical expertise, computing infrastructure, proprietary data and direct experience of how their models behave. Regulators inevitably need access to some of that knowledge. As their reliance on frontier AI companies increases, however, the relationship between expertise and authority becomes harder to manage.

A regulator may possess the legal power to govern while the company it regulates possesses much of the knowledge needed to make that governance effective. That is an unusual asymmetry, and it could become one of the central vulnerabilities of advanced AI regulation.

A rule can be legally robust and still fail if the institution enforcing it does not sufficiently understand the system to which it applies. A watchdog can be formally independent while depending heavily on the regulated industry for its understanding of the technology. A regulator can require safety evaluations while relying on the developer to explain what should be evaluated and how. Governments can establish red lines while lacking the technical capacity to recognise quickly when those lines have been crossed.

The coming regulatory challenge is therefore about maintaining the conditions under which authority remains legitimate. Institutions applying the rules must retain enough knowledge and independence to exercise their authority meaningfully.

In my A3 Model, I describe legitimate agency through the principle of Adhikaram: authority whose exercise must remain justified, conditional and revocable.

Applied to regulation, statutory power alone does not settle whether an institution remains fit to govern a rapidly changing technological system. Formal authority can remain legally intact while the legitimacy of its exercise depends on the institution retaining sufficient knowledge to understand what it is regulating.

Regulators therefore need to ask difficult questions of themselves. Do they possess sufficient knowledge of the systems they oversee? Can they detect meaningful changes in those systems? Can they distinguish reassurance from evidence? Can they challenge technical assumptions made by the organisations they regulate? Can they recognise when their own knowledge has fallen behind? And can the governance system respond when that happens?

Frontier AI makes these questions particularly important because it is not a static object of regulation. Models are updated, their access to tools can expand, and new capabilities can emerge unpredictably. New behaviours and uses can appear after the system was originally assessed.

The regulator is therefore governing a moving knowledge boundary: technology whose capabilities and uses can continue to change.

Traditional regulation can often tolerate some delay between technological change and institutional response but advanced AI may sharply reduce that margin. A system can be modified, scaled or connected to new tools much faster than governments can rewrite legislation, establish new standards or develop specialist expertise.

This creates what I would call ‘epistemic lag’: a growing gap between the speed at which the technology develops and the speed at which regulators can acquire the knowledge needed to govern it effectively.

Türk’s warning about concentrated power should therefore be understood as more than a question of market competition. If technical capability becomes concentrated in a small number of companies, knowledge of how the most advanced systems work may become concentrated there too. Concentration of capability can therefore become concentration of knowledge and, with it, concentration of practical governance power.

If only a handful of organisations possess the expertise needed to understand frontier AI, they will inevitably help shape how risks are described, which measurements are treated as credible and which technical interventions are regarded as feasible. Their indispensable expertise creates the governance problem: the company being regulated can gradually become part of the knowledge system on which the regulator itself depends.

That dependency does not necessarily amount to regulatory capture in the conventional sense. It requires neither corruption nor an improper relationship between regulator and industry. It can arise simply because the technology is so specialised that meaningful oversight depends on knowledge concentrated inside the companies developing it.

This is where the distinction between expertise and sovereignty comes into its own. Technology companies should participate deeply in AI governance – their engineers and researchers often know more about frontier systems than anyone outside them. But possessing superior technical knowledge should not give those companies a greater entitlement to decide where the boundaries of acceptable risk should lie.

At the same time, regulatory independence cannot mean pretending that outside institutions already possess expertise they don’t have. Genuine independence requires building that capacity.

Independent governance requires independent epistemic capacity: regulators capable of interrogating technical information, reproducing important evaluations, challenging assumptions and recognising when their own understanding is no longer adequate.

Regulatory competence also has to be treated as something that can change. An institution capable of governing one generation of AI systems cannot automatically be assumed to possess the knowledge required to govern the next.

A second concept from the A3 Model helps explain why this is important. I use Aanavam to describe endogenous systemic distortion: situations in which a system drifts away from its legitimate purpose because of the relationships, incentives and assumptions operating within it rather than because an outside actor has deliberately corrupted it.

Dependence on industry knowledge can create this kind of distortion. A regulator may gradually adopt the language and assumptions of the industry because those become the main tools through which it understands the technology. Safety questions can begin to reflect what developers are able to measure rather than what matters most to society. Verification can become a matter of confirming procedures rather than independently testing the underlying assumptions. Regulators can remain confident in their oversight even while more of the knowledge supporting that confidence sits outside their own institutions.

The difficulty is that none of this requires bad faith. Every institution involved may behave responsibly according to its existing role while the system as a whole drifts towards a position in which formal authority sits with the regulator but much of the substantive knowledge needed to exercise that authority sits with the regulated industry.

At that point, stronger safeguards are insufficient unless the institution responsible for enforcing them can also recognise when its own knowledge has become inadequate.

This also changes what we mean by independent verification. An outside institution cannot provide genuine verification merely by receiving evidence from a developer and confirming that required procedures have been followed. Independence also requires the ability to interrogate the evidence, reproduce important tests and reach interpretations that do not depend entirely on the developer’s own conceptual framework.

Türk is right to demand strong guarantees. The most important guarantee may be a governance architecture that prevents the knowledge needed to regulate advanced AI from becoming monopolised by the organisations subject to that regulation.

That will require sustained investment in public technical expertise, meaningful access to systems and evidence, genuinely independent evaluation capacity and mechanisms for recognising when regulators themselves have fallen behind the technologies they oversee.

It will also require a different understanding of legitimate authority.

A regulator’s formal powers may come from Parliament, a treaty or another established institution. Under conditions of rapid technological change, however, the legitimacy of exercising those powers cannot rest on their formal source alone. Governing authority must remain coupled to sufficient knowledge and competence to understand what is being governed.

If that capacity deteriorates, the governance system must be able to respond. Responsibility may need to move, oversight may need to deepen, authority may need to be constrained, and institutions may need to be reconstituted around new knowledge.

That is what serious governance looks like under conditions of technological uncertainty. The capacity to govern highly complex technology cannot be assumed to remain constant while the technology itself changes rapidly.

The coming struggle over AI will concern who owns the most powerful models and who possesses the knowledge required to understand them. If that knowledge remains concentrated inside a handful of companies, governments may find themselves in a paradoxical position: possessing the legal authority to regulate AI while depending heavily on the organisations they are regulating to understand what effective regulation requires.

That leaves an uncomfortable paradox. As AI becomes more technically complex, effective oversight may become more necessary while regulators simultaneously become more dependent on the companies whose power they are trying to constrain.

The challenge is to create strong safeguards while ensuring that the institutions entrusted with enforcing them retain the independent knowledge needed to understand what they are governing.

Vendan Ananda Kumararajah is an internationally recognised transformation architect and systems thinker. The originator of the A3 Model—a new-order cybernetic framework uniting ethics, distortion awareness, and agency in AI and governance—he bridges ancient Tamil philosophy with contemporary systems science. A Member of the Chartered Management Institute and author of Navigating Complexity and System Challenges: Foundations for the A3 Model (2025), Vendan is redefining how intelligence, governance, and ethics interconnect in an age of autonomous technologies.

READ MORE: Leaving AI governance to the US and China alone could trigger an AI apocalypse. Artificial intelligence is increasingly being treated as a tool of geopolitical power, risking a global governance regime shaped by the interests of its strongest players – the U.S and China. That could narrow the range of ideas needed to prevent future catastrophic AI failures. To avoid an AI apocalypse, nations and organisations around the world must be able to contribute their ideas and expertise to AI governance, argues Vendan Ananda Kumararajah.

Do you have news to share or expertise to contribute? The European welcomes insights from business leaders and sector specialists. Get in touch with our editorial team to find out more.

_{Main Image: wal_ 172619 / Pexels}