Does AI Make Totalitarianism More Likely?

Introduction

Much of the contemporary AI risk discourse focuses on large-scale existential threats to the human species. However, there are more mundane risks that are also worth considering, one of which is the possibility that AI could enable a new wave of totalitarianism.

Background

Throughout history, advances in communication and bureaucratic technology have enabled larger and more powerful states, with increased ability to monitor and control their populations. In particular, the first half of the twentieth century saw the rise of totalitarian regimes that used new technologies to achieve unprecedented levels of control.

For example, the Nazi regime made deliberate use of mass radio to saturate daily life with centralized propaganda. Under Goebbels, the government promoted the inexpensive Volksempfänger radio receiver to reliably deliver state broadcasts to common households. By controlling the primary communication channel, the regime reduced the space in which dissenting narratives could circulate. Beyond radio, the Nazis also used punch-card tabulating systems supplied by IBM (through its German subsidiary) to process census data. This allowed the regime to act on its ideological priorities with greater speed and consistency, rapidly identifying Jews and other targeted groups.

Other totalitarian governments have made use of similar technologies for various other means of suppressing dissent. For example, in East Germany, the Stasi used an immense archive of files, informant reports, intercepted mail, and wiretaps to anticipate and disrupt dissent before it became organized.

On the other hand, some communication technologies have also been associated with increases in liberty. The spread of print in early modern Europe weakened centralized control over information and helped erode religious and political monopolies. Pamphlets and inexpensive books allowed dissenting ideas to circulate beyond elite circles, contributing to movements such as the Reformation and later democratic revolutions. In some ways, the concept of a written constitution as the foundational bedrock of the United States is contingent on widespread literacy and print culture.

The early internet appeared to have similar decentralizing effects. Digital networks lower the cost of publishing, enabling peer-to-peer communication and reduced reliance on state-controlled broadcasters. During the Arab Spring (~2010-2011), activists in Tunisia and Egypt used platforms like Facebook and Twitter to coordinate protests, share information about state repression, and mobilize large numbers of citizens.

This motivates a natural question: will AI enable more centralized modes of organization, like top-down bureaucracies and totalitarianism, or will it empower more decentralized systems, like markets and civil society?

Structural Mechanisms

Despite the concept of fascism making the “trains run on time”, most historical totalitarian governments were economically dysfunctional, especially compared with their democratic counterparts. In some ways, the entire 20th century can be read as a competition between the relatively decentralized liberal market democracies of the West and relatively centralized totalitarian regimes in Europe and Asia, with the former winning decisively in multiple hot and cold wars, economic growth, cultural production, and technological innovation.

What sort of structural mechanisms explain this pattern? Why did totalitarian regimes underperform democracies, and how might AI change those mechanisms?

We can consider various governments as constrained by their cost-benefit curves. For example, costs of planning, consensus, monitoring, coercion, persuasion, coordination, etc., alter which governance mechanisms are most cost-effective for a given regime. The 20th century favored decentralization because centralization was too expensive, but AI may change many of these costs. For example:

Correlates with authoritarianism:

Increased centralized information-processing capacity (Hayek and Kantorovich)

Reduced dependence on broad human labor for wealth generation (Selectorate Theory and the Resource Curse)

Lower monitoring and enforcement costs (Surveillance at Scale)

More reliable coercive force with reduced defection risk (Robot Armies)

Greater narrative control and centralized propaganda capacity (Propaganda)

Regime coordination advantages over opposition coordination advantages (Coordination Asymmetry)

Anti-correlates with authoritarianism:

Enhanced distributed information processing (Policy Modeling and Foresight)

Improved large-coalition aggregation (Consensus Formation)

Monitoring symmetry between state and citizens (Transparency and Auditability)

Diffusion of coercive capacity (Civil-Military Diffusion)

Strengthened informational integrity (Epistemic Defense)

Enhanced decentralized coordination and innovation (Distributed Innovation)

Let’s explore these speculative mechanisms.

Dictatorship

Hayek and Kantorovich

In Seeing Like a State, James C. Scott argues that a central problem of governance is the ability of a state to see, categorize, and measure the land, population, and capital (the “governants”) under its span of control. The world is complex, so centralized planners use abstract, standardized, and simplified models to monitor the population, allocate resources, and make decisions in lieu of situated, practical knowledge (“metis”). In fact, the state’s desire to understand the system it is managing can in turn alter the system itself, favoring governants with easily parseable and measurable characteristics. Scott calls governants that lend themselves to monitoring and control by a central authority “legible.” Scott goes on to argue that pressure towards legibility (whether successful or unsuccessful) can lead to unintended (often disastrous) consequences.

As an example, consider the Soviet Union’s collectivization of agriculture. The state imposed a rigid structure on farming that ignored local conditions, leading to widespread famine1. The legibility of the collective farm system made it easier for the state to extract resources and control the population, but it also made the agricultural system less resilient and more vulnerable to shocks.

Scott’s critique is epistemic. High-modernist schemes largely fail not due to any moral or political issues, but due to failures in the exchange and processing of information. Centralized planners substitute abstract, standardized representations for the dispersed, tacit knowledge embedded in local practice. But all top-down control requires some type of model, and to reject all possible models would be intellectual nihilism. What other option is there? What institutional form can preserve local knowledge while still enabling system-wide coordination?

Along similar lines, Hayek’s famous essay “The Use of Knowledge in Society” argues that the function of economic organization is to aggregate and utilize dispersed knowledge.If a farmer has better insight into the drainage of their field, a shopkeeper knows what items their customers tend to buy, and a factory foreman understands the idiosyncrasies of their particular machinery, then they should each make decisions independently. Instead of top-down control, the decisions are made in a decentralized fashion and markets coordinate their activity via price signals. No one entity needs to understand the whole system.

We can view Scott and Hayek as diagnosing complementary failures of centralized epistemology. Scott emphasizes that administrative legibility suppresses local, adaptive knowledge in favor of simplified representations. Hayek emphasizes that the knowledge required for economic coordination is dispersed, tacit, and constantly evolving, and therefore cannot be centralized in any usable form. Both are ultimately concerned with how large systems originate and process information. The state operates through centralized abstraction; markets operate through distributed adjustment mediated by prices.

The issue is not that a central authority could in principle compute the optimal allocation if only it had more capacity. Rather, the relevant knowledge is generated and updated through decentralized activity itself. Prices function as signals that both transmit and produce information, allowing coordination without requiring any agent to comprehend the entire system.

If markets coordinate via price signals that summarize dispersed information, could a planner simulate those signals? Could optimization theory reconstruct the informational role of prices within a planned system? This ties back to our question about AI and totalitarianism. If AI can originate and process information at a scale and speed that approaches or exceeds human capabilities, it might be able to replace the need for decentralized markets.

This idea has intellectual antecedents. In the 1930s, the Soviet economist Leonid Kantorovich developed the foundations of linear programming while attempting to solve resource allocation problems in a planned economy. He showed that a central planner could in principle use optimization techniques to allocate resources efficiently. However, the computational resources required to solve these problems at the scale of an entire economy were beyond what was available at the time. The Soviet leadership did not adopt Kantorovich’s methods2, and the planned economy continued to struggle with inefficiency and shortages (and ultimately was outcompeted by Western liberal democracies and capitalism).

Nearly 100 years have passed since Kantorovich’s work, and computational resources have increased by many orders of magnitude. The question is whether modern AI could change the relative tradeoffs between centralized and decentralized information processing.

A sufficiently advanced AI system could process real-time sensor data from every factory, farm, and storefront. It could model consumer preferences from behavioral data at a granularity that prices only approximate3. It could run counterfactual simulations of supply chain disruptions, weather events, and demand shocks. Would a sufficiently powerful AI planner even need markets? In theory, one could update its model continuously, faster than any price signal propagates through a market.

If we view the Hayekian knowledge problem not as an argument for markets, per se but instead as a hypothesis for authoritarian regimes have historically underperformed democracies, then just the shift in the ratio of information-processing power between central planners and decentralized markets could narrow the gap in economic performance and make dictatorships more viable.

Selectorate Theory and the Resource Curse

In a previous post, we explored the political economy of authoritarian regimes through the lens of selectorate theory (developed by Bruce Bueno de Mesquita, Alastair Smith, Randolph Siverson, and James Morrow).

To review, every leader survives by satisfying a “winning coalition.” In democracies, the coalition is large (the electorate), so leaders must provide public goods. In autocracies, the coalition is small (a few elites, generals, party insiders), so leaders can maintain power through targeted patronage.

The key variable is the size of the winning coalition \(W\) relative to the selectorate \(S\). When \(W/S\) is large, the leader is pushed toward public goods provision. When \(W/S\) is small, the leader can buy loyalty cheaply. The model predicts that small coalitions produce bad governance because the incentive structure rewards it.

Consider the determinants of coalition size. When wealth requires broad human participation, agriculture, manufacturing, services, the leader needs the population to be productive, which means providing education, infrastructure, healthcare. The winning coalition is effectively large because many people’s cooperation is needed4.

On the other hand, when

[truncated for AI cost control]