Measuring a Concentration of Power

Introduction

In the last few years, ‘an AI-enabled concentration of power’ has become a common expression in the AI-governance lexicon, and it has been put to use in quite a few situations.

The misuse of advanced capabilities by malicious actors could lead to one (Brundage et al., 2018). So could centralised control over compute, which in some scenarios would concentrate power in whoever oversees the centralisation (Sastry et al., 2024). AI might stabilise authoritarian regimes that would otherwise have fallen (Hezarany, 2024). Automation might hand an unacceptable share of the economy to a handful of AI companies (Dennis et al., 2025), or to the systems themselves (Kulveit et al., 2024). AI could destabilise free societies in which power is spread across many actors (Bullock et al., 2025), and democracies specifically (Katzke et al., 2024).

I have had a mild obsession with this phrase for some time now. Although it sounds precise, like a technical expression, it’s pretty vague on closer inspection and, at times, self-contradictory in ways that matter for how we use it (and what we do about it).

The clearest scope condition I know comes from CLTR (2026): a situation where AI lets one person or a small group gain enough power -economic, political, military, or epistemic power – that most people are disempowered, creating a system that becomes firmly entrenched and self-reinforcing so it cannot be effectively challenged or changed.

The key phrase that I discuss in this blog post and a recent paper is the “enough power” aspect. This quantitative way of looking at the problem of a “concentration of power” is comparable to that of a “concentration of wealth” in that one actor, or a small number of actors, has too much of something, while others have too little of that same thing. In my view, we should reconsider this framing.

In the following post, I’ll clarify what I mean by “this framing”, why I think it mischaracterises the risk, the problems with common proxies, an alternative to those proxies, and some implications.

What is an AI-enabled Concentration of Power?

The scenarios mentioned have a number of features in common: they all appear to be unfair, risky, difficult to reverse, and worth preventing.

But the expression “concentration of power” also suggests what they have in common-that is, they all lie at one end of (what one imagines as) a “power distribution scale.”

Towards the lower end of the imaginary scale, you find ideal democracies and various utopian anarcho-space-communisms. Here, this thing called power is distributed among all people equally. At the higher end, you have authoritarian regimes, techno-feudal autocracies, and the like. Almost no one possesses power, and those who do can use it to have a major impact on the lives of others.

Framed this way, the risk of an “AI-enabled concentration of power” looks a lot like a distribution problem. That is an appealing thought, because distribution problems are tractable. We simply distribute this thing called power to everyone. Everyone rises to the permanent upperclass. Done. 

I think this framing oversimplifies the problem. 

Any statement concerning a distribution assumes that the thing being distributed can be measured in an agent-specific way – and an individual agent’s power is difficult to measure. When it’s easy, it’s often not worth doing. 

What do I mean by this? We can measure how bad wealth inequality is, in a meaningful way, because a persons  wealth can be measured (for example, in dollars). We can use a Gini coefficient or the proportion of wealth held by the top 1% to measure wealth concentration and look at these figures over a number of decades. For instance, in 2022 the top 1% held about one third of all US household wealth, which was an increase from around a quarter in 1989. Nobody disputes that such a claim is meaningful (although people disagree on what to do about it).

Power does not have a unit like a dollar. You can’t directly measure a person’s ability to influence others or reach your goals despite their resistance. 

Sometimes, people try anyway and come up with a proxy: something measurable, like a dollar, that can stand in for power and be counted instead. They often choose one of two options. The first counts the resources that seem to make someone “powerful.” The second looks at who is using whatever resources they have to win political, military or economic contests (See also: Morriss’ vehicle and exercise fallacy (2002)).

The next two sections take each proxy in turn, and argue that both proxies measure something which can be helpful to know about it. Still, neither of them is suitable as proxies for the thing that makes concentration of power scenarios objectionable – which is the difficulty the last section is about. 

Counting Resources: Who owns what?

It could be argued that we already know the factors which make an actor powerful, such as oil, money, energy, weapons, territory, computing power and data. Let’s take these resources as a proxy. 

If you measure these resources and look at the way they are distributed in a society, you have a substitute measure of power based on where those resources come from. At the end of the scale where concentration is low, the resources are evenly distributed among different individuals; at the end where concentration is high, they are – quite literally – concentrated in the hands of a small number of privileged people.

This is a common heuristic, and a standard approach in quantitative international relations.

In US strategic planning, the DIME framework divides national power into four components-diplomatic, informational, military, and economic-considering an evaluation of a state’s power to be a consequence of a state’s “capacity” across these four areas. More granular quantified versions are also available. For example, there is the Composite Index of National Capability, which forms the basis of a great deal of quantitative international relations research. It combines six indicators of a state’s material capacity-population, urban population, iron and steel production, energy consumption, military spending, and military personnel-into a single figure showing the state’s share of global capability, thereby allowing any state to be positioned on a scale ranging from zero to one. 

There are two problems with seeing concentrations of power as unequal access to “valuable” resources:

First, we divide the original problem of measuring power into multiple smaller measurement problems. Rather than comparing the overall power of two actors, we compare their military power, economic power, soft power, cyber power, computing power, and other forms of power. Each of these areas in turn has to be broken down. Military power can be divided into land, sea, air, nuclear, and so on; economic power into GDP, trade influence, currency reserves, control of key suppliers, innovation, and so forth. This process of breakdown goes on without reaching a clear conclusion. Moreover, there is no single “dollar” and no exchange rate by which the various components can be compared. Although we can compare the oil reserves of two countries, we cannot determine how much power a barrel of oil has compared to an influential news station, a court decision, a seat on the UN Security Council, or a gun.

Indexes such as CINC work by assigning weights-for example, by balancing steel production against the number of military personnel-since weights are necessary in order to arrive at a figure. The index was developed in 1963 to help explain the wars that took place from 1816 onwards, using the same indicators that determined those wars. In 1943, steel output was almost a direct measure of how many ships and tanks a country could field, and the United States’ tonnage advantage over both Germany and Japan translated fairly directly into military supplies. Is this still the case? Does it matter, in a world with nuclear weapons?

Lincoln Allison describes power as “partially quantifiable” in order to express this characteristic; within a restricted area and a given context you are often able to rank the actors, but when looking across different areas such ranking has no real foundation.

The second and more serious point against this measure for power is that when you are concerned about the concentration of the resources necessary for one person to subdue another, you don’t normally deal with the problem by distributing those resources to everyone else. The CINC figure shows the amount of resources a country needs in order to win wars- should that be the one to be made more equally distributed?

Would the world be freer if everyone had nuclear weapons? Nine states have them today. A population can be individually capable of lethal violence and unfree at the same time and an “open” world in which every actor has frontier bio-design or cyber-offence capability is not obviously freer than one in which only two actors have access to that capability.

Counting Exercises of Power: Who keeps winning?

Someone else might say that resources are the wrong thing to focus on. Instead of counting the stockpiles that allow a person to win a political contest, a war, or an economic competition, we should just look at who wins those political contests, wars, and economic competitions: a measure derived from the exercise of power.

This proxy creates another scale from low to high concentrations of power:

On one end, a powerful person or a small number of people exerts control over others by continuously demonstrating their ability to imprison dissidents, supress protests, and alter the results of elections. On the other end, everyone occasionally wins and loses political contests, wars, or elections- otherwise, people get along, regard the existing order as legitimate, or at least seem to accept it without much questioning.

In this view, we “measure” a concentration of power by looking at who wins political contests when push comes to shove. If the same people keep winning, whether by coercive force or at the ballot box, we assume they hold power. A concentration of power, then, is a state of affairs in which a small minority wins all or most political contests and benefits accordingly, while everyone else makes do.

Again: There are certain reasonable uses for this measure. It records the behaviour of political actors and can therefore be employed in things such as democracy indices. In my opinion, “how many consecutive terms has the incumbent won” is a better indicator of the type of regime than “how many tanks does the incumbent own”.

It has two related problems, though.

The first problem is that visible exercises of coercive power tend to indicate that an authority is contested, and thus contestable.  People can treat an illegitimate, oppressive regime as legitimate because they see no prospect of changing it. A government that jails dissidents is spending effort to hold a position that is not secure. A government that does not need to jail anyone may face no opposition worth the trouble.

The way people estimate how the rest of them think has a great deal to do with whether or not political contests take place. If each individual believes that all the others back the regime, then no one will be the first to take action and as a result the illusion of a consensus becomes self-fulfilling. (see: Power Lies Trembling, Private Truths, Public Lies).

AI probably makes this worse in certain respects: first of all, predictive policing and widespread surveillance enable the state to intervene before dissent becomes apparent; secondly, information environments mediated by AI can create the illusion of a consensus by means of synthetic commentary, personalised feeds, and chatbots which answer political questions in the way that the regime desires, so that the very person whose legitimacy is in question ends up setting the private judgments that cause illusion of collective consent.

The other issue is the distinction between winning and having the ability to win the things one wants. Arguably, the most concerning scenarios are ones where coercion has become unnecessary. In these worlds, there are a great many groups which might win political contests, but none of them really matter. The topics that matter are not considered “political” or worth discussing. (See also: Crenson, 1971)

A political or military victory can only be taken as evidence of power under two conditions: the contest was worthwhile to enter and the result was one that the person involved wanted. “Winning” is not sufficient evidence that either of these things is true. Otherwise, you end up with (what I’d call) the “drunk-guy-at-a-bar-theory” of power. The drunk guy at the bar who keeps winning fights. He prevails in every contest he enters. He also can’t choose which fights happen, can’t avoid them, and is no better off for having won. Whatever we mean by power, he doesn’t have much of it, and a measure that counts contests won would rank him highly.

The same objection applies to CINC. Treating war outcomes as the measure of power assumes that states are – mostly – aiming to win wars. However, a state which achieves its objectives without having to go to war has no victories to claim, and a state that continues to fight and win might still end up being drawn into conflicts which it would be better served to avoid. 

Alternatives and other Implications

I don’t think it’s a useful approach to measure power in an absolute way if one is trying to tell the difference between a concentration of power and an acceptable situation. In the cases referred to, power is not literally concentrated, and searching for a substitute in the sources of that power or in its most obvious effects does not help much.

The distinction between a concentration-of-power situation and one that is acceptable is not so much a question of degree as it is one of kind; in my view the appropriate question to ask about a particular arrangement is whether its collective decision making processes are, in some fundamental way, legitimate. There are a few ways to talk and think about legitimacy. As a foundation, I think it’s helpful to distinguish between firstly, arrangements that (broadly) produce outcomes that can be justified to the people who live under it, and ones that don’t. And secondly, between arrangements that acquire and maintain power in a way that can be justified to those who live under it, and ones that don’t.

Legitimacy criteria can, of course, be quantified to some degree. But it’s much more difficult, and more contentious than quantifying who owns what, or who wins elections. Take, for example, the V-Dem democracy index: you can find the various indices used in calculating a regime’s democracy score here. V-Dem gives five democracy scores: electoral, liberal, participatory, deliberative and egalitarian. There is little consensus about what democracy entails, and each index reflects a different view of legitimacy: The egalitarian index includes health equality, educational equality and the equal distribution of resources. By contrast, the deliberative index looks at whether public justifications refer to the common good and whether counter arguments are respected – aspects that are reasonable democratic values and also exactly the kind of issues that people debate politically. Even the process of aggregation involves a (somewhat controversial) normative decision : the overall electoral democracy index combines its five subcomponents both by addition and by multiplication, since it is a significant political question whether a country can make up for falsified elections with a free press or whether that failure should cause the entire score to be lowered.

I’d also like to discuss three different implications I think are important to consider: 

Firstly, AI does not have to be the super-powerful multi-tool instrument through which a user attains power. Most of the literature treats an AI-enabled concentration of power as a scenario where the use of AI systems or agents lets its developers or users dominate everyone else. But illegitimate regimes can occur through many other routes. AI systems could generate enough economic, informational, or military disorder that a degree of state intervention which we normally should not agree to seems to become necessary. If automation displaces a large share of the workforce within a few years, if the information environment is such that nothing can be verified, or if a cyber-attack brings critical infrastructure down, each of these situations would lead to a demand for emergency powers (And emergency powers have a well-documented habit of outliving the emergency.)

The AI governance equivalent of “counting resources” is also often applied to compute. It is a reasonable thing to track: whoever controls the chips and the data centres, it’s often said, gets to decide what is built, who gets to build it and on what terms. This is often followed by figures such as “three cloud providers hold about two-thirds of the global cloud market, and a single company designs most of the chips used to train frontier models”. However, owning chips is not the same as controlling them, let alone controlling which AI systems are built and on what terms. It goes without saying that the same quantity of compute buys wildly different kinds of leverage depending on external factors e.g. algorithmic efficiency gains, export controls, energy supply, the terms of a cloud contract, and whether the models it trains are helpful for something important.) Moreover: Compute is one route to leverage over deployment, development and user decisions among several, and not obviously the shortest. A regulator can determine what may be deployed, a distribution platform can determine what reaches users, a handful of firms supply training data, and a few suppliers upstream of the chip designer can constrain everyone downstream. An actor with no compute at all can hold considerable leverage over how AI is developed and used, and an actor with a great deal of compute can hold surprisingly little.

Secondly: The severity and duration of a concentration of power scenario are separate questions. The usual definition of a concentration of power tends to combine the most severe forms of domination with the most deeply rooted ones (and vice versa). This makes sense if we consider concentrations of power as a distribution problem. But political scientists researching regime survival generally agree that it is the institutional structure (e.g. party organisation, elite bargaining, and succession rules) not the degree of repression that makes a regime last.. And for good reason!  It’s an extraordinary claim that the most repressive regimes are more stable than others. And anecdotally, it’s not exactly airtight. 

The Third Reich lasted only twelve years, Stalin’s rule lasted about a quarter of a century. At the same time, some of the most long-lasting authoritarian regimes of the twentieth century – for example, Mexico’s PRI, which ruled for seventy-one years ( See also: Langston, 2017)  or Singapore’s PAP, which has been in power since 1959 – have been and remain relatively lenient. The reasons which account for the durability of a state of domination need not be the same as those which make it damaging during the time it lasts. The two parameters should therefore be examined separately. 

I think it is important to take this especially seriously in the context of AI Governance. The durability of a regime is not a negative feature in itself.  Worries about “Lock-ins” are worth taking seriously, but the elements which contribute to an arrangement’s stability can be neutral, and even positive, regardless of whether the arrangement is one that is worth maintaining or not. Institutions that survive the tenure of individual leaders, clear rules governing succession, elites who find it easier to operate within the system than to leave it, and an order that is seen as legitimate by the population that lives under it are all good things to aim for. 

Purely “preventing entrenchment” cannot be the objective. An intervention that made every arrangement easy to overturn would remove the protections against the abuses of power along with the threats. What we want entrenched is not more or less power but arrangements of a particular kind – ones whose holders can be required to justify themselves and can be removed. AI could create new patterns of entrenchment or radically change the existing ones. For instance, sophisticated AI might be used to automate surveillance and enforcement in such a way that makes it riskier or less obvious for elites to defect or for leadership changes to take place, thus decreasing the chances of institutional change.

Lastly, and perhaps this goes without saying: Addressing this risk is very difficult. It shouldn’t surprise us that we can’t solve the fundamental challenges AI systems pose to our social contract by arming everyone to the teeth with AI agents and their own data centre. Resources are useless without the ability to control them and make informed decisions about what to do with them. Making sure people can access this ability is much more complicated than open sourcing things. Export controls are not good or bad because they “reduce a concentration of power” – and much more usefully discussed as a measure that might “preserve someone’s standing to contest what the compute is used for”. In some ways, preventing an AI-enabled concentration of power means making sure that AI does not produce arrangements of unaccountable and uncontestable rule. I don’t think we’ve figured out how to do that yet, but now seems like a good time to try making progress.

Thurnherr, Lara, Operationalising AI-Enabled Concentrations of Power (20th July 2026). Available at SSRN: https://ssrn.com/abstract=7201418 or http://dx.doi.org/10.2139/ssrn.7201418