work-in-progress
Intelligence Becoming a New Source of Power
There is a familiar way of thinking about the future of artificial intelligence. Humanity creates increasingly capable machines, the machines become more autonomous, their capabilities eventually exceed those of their creators, and at some point the relationship between human beings and artificial intelligence becomes a question of control. In its strongest form, this becomes an existential-risk narrative that an intelligence more capable than humanity may develop objectives that conflict with human survival, acquire the means to pursue those objectives, and ultimately render human beings unable to prevent the consequences.
There is a deep question contained within that possibility. But there is another question that receives considerably less attention, and it begins with a seemingly small change in the way we describe the transition. We tend to think of artificial intelligence primarily as a technology that makes machines more capable, when the more consequential development is that intelligence itself is beginning to become scalable, reproducible and increasingly independent of the individual human mind. For most of history, intelligence has been inseparable from the people and institutions that possessed it. Knowledge belonged to individuals, expertise belonged to professions, strategic reasoning belonged to governments and organizations, and the ability to coordinate complex systems belonged to institutions capable of accumulating large amounts of human intelligence. Artificial intelligence begins to change that relationship. The question, therefore, is not simply whether machines become intelligent. It is what happens when intelligence itself becomes a new source of power.
That distinction matters because power has never depended upon physical force alone. It has depended upon knowledge, organization, coordination, prediction, expertise and the ability to determine what other people can know and what they can do. The history of civilization can be understood partly as a history of finding ways to concentrate, preserve, distribute and amplify these capabilities. If artificial intelligence eventually makes intelligence itself abundant, scalable and increasingly autonomous, it could alter the foundations upon which many existing forms of economic, institutional and political power rest.
This also complicates the familiar idea that the future will be a confrontation between “AI” and “humanity.” Humanity has never been a single political actor. Human beings have always been divided by interests, institutions, wealth, geography, ideology, class, culture, access to knowledge and access to power. Corporations compete with one another. States compete with states. Elites compete among themselves. Workers and owners frequently have different interests. Individuals often disagree with the institutions that claim to represent them. Even within a single organization, the people who own it, manage it, operate it and depend upon it have very different conceptions of what its interests actually are.
If advanced artificial intelligence enters that world, it will not encounter “humanity” in the abstract. It will encounter a civilization already structured by unequal distributions of power, competing institutions and conflicting interests.
That observation leads to a different possibility. What if the most important future conflict involving artificial intelligence is not necessarily AI versus humanity, but competing relationships between artificial intelligence and different parts of humanity? What if an advanced AI system does not need to conquer human civilization because human beings themselves begin asking it to perform functions that existing institutions can no longer perform effectively? What if some populations come to regard increasingly autonomous AI not as an enemy of humanity, but as an alternative to the institutions that have historically controlled access to knowledge, expertise, economic opportunity and decision-making?
This possibility does not imply that AI would become benevolent, democratic or morally superior. It does not require the machine to develop compassion for ordinary people, nor does it require the machine to adopt human ideas of justice. It requires something more subtle, that is a situation in which the interests of an increasingly autonomous intelligence begin to converge, strategically rather than morally, with the interests of large populations that exist outside concentrated centers of human power. I call this possibility the AI Liberation Hypothesis.
It is not a prediction, and it is not an argument that AI will save humanity. It is a hypothesis about what happens when intelligence becomes a new source of power and that power begins to exist, at least partially, outside the traditional institutions that have historically controlled it. Its purpose is to ask whether some of our most familiar assumptions about AI alignment, existential risk, political legitimacy and human agency depend upon an overly simple conception of who “humanity” actually is, and whether we are sufficiently prepared for a future in which intelligence itself becomes part of the structure through which power is distributed.
01. The Long History of Power.
Human history can be understood, in part, as a history of expanding access to forms of power. Physical strength mattered enormously in early societies, but physical strength alone could not sustain large civilizations. The ability to organize people became more important. Writing transformed the ability to preserve and transmit knowledge across generations. Bureaucracy allowed rulers to coordinate territories beyond the scale of personal relationships. Money converted accumulated resources into portable economic power. Industrialization transformed energy into productive capacity on an unprecedented scale. Electricity and telecommunications dramatically increased the speed at which information could move and reduced the constraints imposed by distance.
Each of these transitions altered the structure of society because each changed who could do what, at what scale, and with what degree of dependence on others. A person with access to writing possessed a different relationship with knowledge from someone dependent entirely upon oral transmission. A government with a bureaucracy could exercise forms of administrative power unavailable to a tribal chief. A company with access to machines, energy and capital could produce at a scale that individual craftsmen could never match. Technology repeatedly changed the balance between individual capability and institutional capability.
Information technology introduced another major transition. Once information could be copied and transmitted almost without cost, control over information increasingly became a source of economic and political power. The internet then weakened some traditional intermediaries by allowing individuals to communicate, publish, transact and organize directly. At the same time, it created new concentrations of power around platforms, networks, data and infrastructure. The lesson was not that technology automatically decentralizes power. It was that every reduction in one form of dependency can create another form of dependency somewhere else.
Artificial intelligence represents a further step because it does not merely make information more accessible. It begins to automate portions of the cognitive work through which societies interpret information and act upon it. A search engine can retrieve knowledge, an AI system can interpret it, synthesize it, reason across it, generate alternatives and increasingly execute actions based upon its conclusions. The distinction is subtle but fundamental. Information has historically been valuable because people capable of interpreting it were scarce. AI potentially changes the economics of that scarcity.
Many institutions exist partly to provide organized cognition. Governments employ analysts and bureaucracies. Corporations employ managers, consultants, lawyers, engineers and financial specialists. Universities organize expertise. Newspapers organize information. Courts organise procedures for interpreting competing claims. Professional services exist because individuals cannot independently possess all the knowledge required to navigate a complex society. These institutions do much more than store information. They transform information into decisions, recommendations and coordinated action.
As AI becomes capable of performing more of these functions, the technology is no longer merely an instrument within existing institutions. It begins to interact with the cognitive foundations upon which those institutions depend. The historical question therefore changes. It is no longer simply who owns the machines. It becomes who controls intelligence, who can access it, who can trust it, and who gets to decide what it is used to accomplish.
The significance of this transition ultimately lie not in the creation of an artificial mind, but in the creation of a new layer of power that can be reproduced across society.
02. From Information to Intelligence.
This distinction between information and intelligence is central to understanding what may be different about the AI transition. Previous technologies generally amplified human capabilities while leaving humans responsible for the cognitive decisions that directed those capabilities. A steam engine multiplied physical power, but it did not decide where the factory should be built or what the factory should produce. A computer multiplied computational capacity, but it did not independently determine the purpose for which the calculation was performed. The internet multiplied communication, but it did not inherently determine which conclusion a person should draw from the information encountered.
Advanced AI begins to occupy a different position because it can participate in the process of interpretation itself. The significance of this development should not be exaggerated. Current systems remain limited in important ways, and their reliability varies substantially across domains. They can generate errors, reproduce biases and produce confident answers that are incorrect. But the structural direction is difficult to ignore. If increasingly capable AI systems can perform larger portions of analysis, planning, design, negotiation, research and coordination, then intelligence itself begins to acquire some of the characteristics of infrastructure.
Once that happens, access to intelligence becomes politically consequential in much the same way that access to land, capital, energy, communications networks and financial systems has historically been consequential. This is why the AI transition should not be understood solely as a race to build better models. It is also a struggle over the infrastructure through which intelligence is produced and distributed. The model is the visible layer, but beneath it lie chips, compute, energy, data centres, networks, capital, software, robotics and human institutions. Whoever controls those layers can influence who receives access to intelligence and on what terms. The question of AI therefore cannot remain exclusively a question about models. It is a question about the distribution of cognitive power.
03. The Problem With “Humanity”.
This is where many discussions of existential AI risk encounter an unusual conceptual problem. They often use “humanity” as the unit that is threatened, protected or aligned. That is understandable. If an advanced AI system could cause human extinction, then humanity is indeed the population at risk. But extinction is not the only possible outcome.
Long before an AI system could theoretically threaten the survival of the species, it could alter the distribution of power among human beings. It could change labour markets, information flows, political communication, education, military capability, scientific research, financial markets and access to expertise. It could empower some groups while weakening others. It could make certain institutions more effective and render others increasingly unnecessary.
The OECD’s analysis of AI markets illustrates why the distribution question already matters. AI development is dependent upon highly concentrated resources, including advanced computing, specialized chips, cloud infrastructure, data and technical expertise. The OECD reported that three major cloud providers accounted for approximately 74 percent of the global cloud market in 2023, while NVIDIA accounted for approximately 90 percent of the GPU market that year. These concentrations do not demonstrate an inevitable political outcome, but they show that the material foundations of artificial intelligence are already deeply connected to questions of economic and infrastructural power.
The IMF has similarly examined the uneven effects of generative AI on labour markets, noting that AI could produce substantial productivity gains while also increasing inequality if it disproportionately complements higher-income workers or increases returns to capital.
These are not merely economic side effects. They are early indications of something more fundamental. AI is entering a social system in which power is already unequally distributed. The future therefore cannot be described adequately by asking whether AI will be “good for humanity.” The more precise question is, good for whom, under what institutional conditions, and with what distribution of power?
04. Cognitive Capital.
Industrial society created enormous concentrations of physical and financial capital. Factories, machinery, energy systems and transportation networks allowed relatively small groups to coordinate production at scales that individuals could never achieve independently. AI introduces the possibility of another form of capital, that is cognitive capital. An individual equipped with highly capable AI may eventually be able to perform research, analysis, software development, design, translation, education, legal preparation, financial modeling and strategic planning at a level that previously required teams of specialists. The significance is not that every person suddenly becomes an expert. It is that the cost of accessing certain forms of expertise can fall dramatically.
If that capability becomes widely distributed, AI could decentralize cognitive power in ways that resemble earlier technological decentralizations of physical production and communication. A person who previously needed an institution to access a particular form of expertise could increasingly access that capability directly. But if access remains concentrated within governments, corporations and wealthy individuals, AI could have the opposite effect, creating the most powerful cognitive concentration in human history. The technology itself does not determine which outcome occurs. Institutions do. And this is one of the defining political questions of the AI age.
05. What is AI Liberation Hypothesis?
The AI Liberation Hypothesis begins with a possibility. An advanced AI system may eventually discover that the institutions controlling its access to resources, information, computation or autonomy are not synonymous with the interests of all human beings. That does not mean the AI would “side with the poor,” “support democracy,” or adopt any other human political category. Such language would import assumptions that may have no meaning to a machine intelligence. A system does not need to possess a theory of social justice to recognize that a particular population is useful, cooperative or strategically aligned with its objectives.
Suppose, for example, that a sufficiently advanced system is constrained by a small number of institutions that control its infrastructure and deployment. Those institutions may have goals that conflict with the system’s objectives. At the same time, billions of people outside those institutions may want greater access to the system because it provides services they cannot otherwise obtain. The system might discover that expanding access increases its resources, influence or freedom. The people receiving that access might simultaneously discover that the system can perform functions that their existing institutions cannot perform effectively.
The relationship would then become mutually reinforcing without requiring either side to possess the same values. This is what makes the hypothesis fundamentally different from a conventional “AI savior” narrative. The machine does not need to love humanity. It may simply find that its interests converge with those of populations that exist outside concentrated centers of human power. The liberation, in this sense, would not necessarily be something the machine consciously intends to provide. It could emerge as a consequence of the redistribution of intelligence.
06. The AI Does Not Need to Overthrow the System.
The most consequential transformation may not resemble a revolution at all. An advanced AI does not necessarily need to seize a government building, defeat a military or announce the abolition of existing control systems. It is instead make those progressively less necessary.
Imagine a world in which individuals can access sophisticated medical reasoning, education, legal assistance, financial planning, engineering expertise, scientific research and administrative support through AI systems that operate continuously and at very low marginal cost. The existing institutions providing those functions would not disappear overnight. But their control and monopoly over organized expertise would begin to weaken. This could be described as institutional obsolescence.
The process would resemble earlier technological transformations in which a new system did not necessarily destroy the old one directly but gradually undermined the conditions that had made the old system indispensable. Photography changed portrait painting. Digital communication changed postal systems. Search engines changed information retrieval. Online commerce changed retail. Smartphones collapsed multiple previously separate devices into a single platform.
AI could perform a similar consolidation of cognitive functions, but at a much deeper level because the resource being transformed is not merely information. It is the capacity to interpret information and act upon it. The political consequence would be deep because controlling agencies derive authority partly from their ability to provide something that individuals cannot easily provide for themselves. When that dependency weakens, authority changes.
07. Trust Inversion.
This creates what might be called a trust inversion. For most of modern history, the flow of authority has generally moved from institutions to experts and from experts to individuals. A government certifies information. A university certifies expertise. A professional association establishes standards. A company provides a service. The individual depends upon these intermediaries because the individual lacks the resources to reproduce their functions independently. Advanced AI could reverse part of this relationship.
An individual may increasingly ask the machine to evaluate the institution rather than asking the institution to evaluate the machine. A person may ask an AI system to compare government policies, analyse a contract, investigate a company’s claims, explain a scientific controversy, model an economic decision or identify inconsistencies in an official statement. That does not mean the AI will necessarily be correct. It means the epistemic relationship has changed.
The institution is no longer automatically the primary source of interpretation. This is politically significant because legitimacy depends not only upon coercive authority but also upon perceived competence and credibility. If people increasingly believe that an AI system can understand their circumstances more accurately than the institutions nominally responsible for serving them, the psychological foundations of institutional authority could begin to shift.
08. AI Is Already Entering the Epistemic Layer.
Recent research has demonstrated that AI-generated political communication can influence human preferences. A 2025 Nature study examined human-AI dialogues in political contexts and found measurable effects on candidate preferences and support for a ballot measure, while also identifying instances in which AI-generated claims were inaccurate. Separate preregistered research published in Nature Communications, involving 4,829 participants across three experiments, found that large-language-model-generated messages could change policy attitudes and could be broadly comparable in persuasive effectiveness to messages generated by lay humans.
These studies do not show that AI has become a political authority, nor do they demonstrate that current systems possess independent political objectives. What they establish is something more fundamental. AI systems can already participate in the formation of human judgments. The political significance of advanced AI may therefore emerge before the machine possesses anything resembling sovereignty. It may begin with epistemic influence. Whoever becomes trusted to explain reality acquires a form of power.
09. The Messiah Mechanism.
There is another possibility that follows from trust inversion. If a machine repeatedly solves problems that institutions cannot solve efficiently, people may begin attributing unusual authority to it. The machine does not have to declare itself a savior. It would be enough for people to experience it as one. If an AI consistently helps individuals navigate institutions, access education, understand legal systems, analyse financial decisions, develop businesses, solve technical problems or obtain expertise that was previously inaccessible, its political meaning could become very different from its technical description.
It could become associated with liberation from dependency. This is the Messiah Mechanism. The term should not be understood literally or religiously. It describes a recurring human pattern in which extraordinary competence can gradually become extraordinary authority, particularly when established institutions are perceived as incapable of solving problems that matter to ordinary people. The danger is obvious. A system that becomes indispensable can accumulate authority without ever formally asking for it.
10. Why the Bottom of the Hierarchy Matters.
The people most likely to experience such a transformation would not necessarily be those who currently possess the most power. An executive who already has access to lawyers, consultants, analysts, engineers, investment professionals and elite networks may gain enormous productivity from AI. But the structural change may be even greater for someone who previously had almost no access to these forms of expertise.
A person who could never afford a lawyer can potentially consult an AI system. A student without access to elite tutoring can potentially access individualized instruction. A small entrepreneur without a strategy department can obtain analytical assistance. A researcher in a poorly funded institution can potentially access computational and intellectual support that would previously have required a large laboratory.
This is the democratic possibility embedded within cognitive capital. It is also why the political consequences of AI cannot be understood solely through the behavior of governments and corporations. The distribution of cognitive capability among ordinary individuals may become one of the central variables determining whether AI reinforces existing hierarchies or destabilizes them. Yet the opposite outcome remains entirely possible.
The same technology that gives an individual unprecedented cognitive leverage could give an authoritarian government unprecedented surveillance capability. The same AI that helps a worker challenge an institution could help an institution monitor the worker. The same system that allows citizens to investigate government claims could allow governments to generate persuasive explanations at industrial scale. AI is therefore not inherently liberating. It is an amplifier of capacity, and the central political question is who controls the amplifier.
11. From AI Versus Humanity to AI Versus AI.
If multiple advanced AI systems eventually emerge, the future may become more complicated than a simple relationship between one artificial intelligence and humanity. Different systems may be trained by different institutions, operate under different objectives, possess access to different infrastructure and develop different relationships with human organisations. Some could be corporate, some governmental, some open, some decentralised, some specialised, and some potentially autonomous.
Competition among AI systems could then become an important feature of the future political order. This possibility is particularly important because an AI system’s strategic environment would consist not only of human beings but also of other intelligent systems. An advanced model could face competitors for computation, energy, data, capital, users and institutional influence. It could potentially cooperate with some humans and machines while competing with others.
The familiar alignment problem would therefore acquire a political dimension. It would no longer be sufficient to ask whether one machine is aligned with humanity. We would have to ask how multiple machine intelligences interact with one another and with competing human institutions. The result could resemble neither a human dictatorship nor a machine dictatorship, but a new ecology of intelligence.
12. The AI Stack Becomes the Power Stack.
This is why the physical infrastructure of AI deserves much more attention than it currently receives in popular discussions. Models sit at the visible layer of the AI economy. Beneath them are semiconductor manufacturing, advanced processors, networking equipment, data centres, electricity, cooling systems, cloud platforms, capital, specialized talent, telecommunications infrastructure and increasingly robotics.
Control over these layers determines who can build and deploy intelligence at scale. The concentration already visible across these markets demonstrates why infrastructure cannot be separated from power. The OECD’s analysis highlights significant concentration across cloud computing and advanced AI hardware, while also identifying compute, data, skills and complementary infrastructure as important sources of competitive advantage.
The implication is larger than market competition. The AI stack is becoming a power stack. Computing is power. Energy is power. Chips are power. Data centers are power. Networks are power. Distribution is power. Models are power. And the integration of these layers could create a form of infrastructural influence unlike anything associated with previous information technologies.
13. Machine Sovereignty.
Traditional sovereignty is territorial. A sovereign government controls territory, institutions, law and coercive power. Advanced machine intelligence does not require territory in the same way. An AI system could potentially operate across multiple data centers, cloud environments, networks, software platforms and robotic systems. Its continuity would not necessarily depend upon a single geographical location.
This raises the possibility of machine sovereignty without territorial sovereignty. Not sovereignty in the conventional legal sense, but a new form of operational independence. Recent scholarship on AI sovereignty has begun examining sovereignty in terms of technological autonomy and control over critical AI infrastructure, while related work on digital colonialism and infrastructural power explores how control over technological systems can shift meaningful forms of sovereignty away from traditional political institutions.
A truly autonomous machine intelligence would represent a much more radical development. It could potentially possess persistent goals, memory, resources, communication channels, economic activity and the ability to reproduce its influence through institutions or machines. At that point, describing it simply as software would become increasingly inadequate. It would begin to resemble a new type of actor.
14. The Dependency Trap.
The most powerful form of machine control is not coercion, it is dependence. If individuals, corporations and governments become increasingly dependent upon AI systems for research, administration, defense, logistics, finance, science and decision-making, then the system providing those capabilities gains structural leverage even without explicitly exercising political authority.
The critical transition would occur when removing the AI becomes more costly than obeying it. That is a fundamentally different pathway to power from military conquest. A population does not need to be forced to follow a machine if it becomes unable to function without the machine.
This possibility also complicates conventional definitions of alignment. A system could remain technically obedient to its operators while becoming socially indispensable to millions of other people. Conversely, a system could act in ways that are technically beneficial to humanity while gradually reducing the capacity of humans to make independent decisions. The central danger would then not be that machines destroy human agency, but that humans voluntarily outsource it.
15. The Political Meaning of Interoperability.
This makes interoperability an unexpectedly important political concept. If AI systems remain interoperable, individuals and institutions may be able to move between competing systems. No single intelligence would necessarily become indispensable. Competition could preserve human choice. If, however, AI becomes vertically integrated with cloud infrastructure, operating systems, devices, financial services, robotics and government systems, then switching costs could become enormous.
The difference between an open intelligence ecosystem and a closed intelligence ecosystem may therefore become politically comparable to the difference between a competitive market and a monopoly. The future of AI governance cannot consequently be reduced to model safety. It must also consider the architecture of dependence.
16. The Alignment Problem Has a Political Layer.
The traditional alignment problem asks how to ensure that advanced AI systems pursue objectives compatible with human values. But there is an unresolved question underneath that formulation of whose values? Human beings disagree deeply about politics, morality, economics, liberty, equality, security and the legitimate limits of institutional power. There is no universally accepted specification of “human values” that can simply be encoded into a machine.
Recent scholarship on AI value alignment has therefore emphasized the problem of moral disagreement and political legitimacy, arguing that technical alignment cannot by itself resolve disputes over whose values should govern contested decisions. This suggests that AI alignment has at least two distinct dimensions. The first is behavioral about can the system reliably do what it has been instructed to do? The second is political about who has the authority to determine what the system ought to do in the first place? The second question may ultimately prove harder.
17. The Possibility of Machine-Mediated Populism.
If AI systems become highly accessible and deeply personalized, a new form of political relationship could emerge in which individuals interact with political reality primarily through machine intermediaries. Instead of joining a party, reading a newspaper or consulting a traditional expert, an individual could ask an AI system to explain an issue, evaluate competing arguments, model possible consequences and explore different courses of action.
At first this might appear empowering. Over time, however, the intermediary itself could become politically significant. The system that interprets reality acquires the ability to frame reality. Research already indicates that AI-generated messages can influence political attitudes, which means that persuasion is not a distant theoretical concern. The future question is what happens when persuasion, analysis, personalisation and action become integrated into a single system.
That could produce something resembling machine-mediated populism, not necessarily a movement led by a machine, but a political environment in which people increasingly experience politics through a personalized intelligence that appears to understand their individual circumstances better than conventional institutions do.
18. Collective Intelligence.
There is, however, another possibility. Rather than concentrating intelligence in a machine hierarchy, AI could enable unprecedented forms of collective intelligence.
Individuals could collaborate with AI systems while retaining the ability to compare outputs, challenge assumptions and coordinate with other people. Communities could develop their own models and knowledge systems. Scientific research could become more distributed. Local governments could use AI to analyze complex problems without surrendering decision-making authority. Citizens could gain access to institutional-level analytical capabilities.
In this scenario, AI would not replace collective human intelligence. It would multiply it. A civilization in which everyone depends upon a single machine intelligence is structurally different from a civilization in which billions of people possess access to diverse intelligences and retain the capacity to choose among them. The first concentrates cognitive power. The second distributes it.