AI and Authoritarianism

AI and authoritarianism refers to the risk that powerful AI gives authoritarian regimes, or aspiring authoritarians, tools for surveillance, repression, military dominance, and propaganda effective enough that the resulting autocracy becomes essentially impossible to overthrow. The concept is developed primarily in two essays by Anthropic CEO Dario Amodei, Machines of Loving Grace (Amodei, ~October 2024) and The Adolescence of Technology (Amodei, 2026), which set out both the threat and a proposed democratic response.

The forward-looking argument in those essays sits alongside a longer measured record of digital repression, tracked annually by Freedom House, which supplies the empirical base against which claims about AI's marginal contribution can be assessed.

Mechanisms of AI-enabled autocracy

Amodei groups the instruments of AI-enabled authoritarian control into four categories, summarized below with their current state of deployment.

Population-scale surveillance is, in this account, the primary technical instrument of AI-enabled authoritarian control; it is treated in more depth in AI and Surveillance.

Threat actors

Amodei ranks the potential wielders of these tools by severity, in four tiers. First, and most severe, is the Chinese Communist Party, which he identifies as having the strongest combination of AI capability, autocratic government, and surveillance infrastructure, calling it "hands down the clearest path to the AI-enabled totalitarian nightmare." Second are AI-competitive democracies, which possess legitimate defensive tools but risk turning them inward against their own populations. Third are non-democratic countries with large datacenters, which could expropriate that infrastructure. Fourth are AI companies themselves, which control datacenters, frontier models, and influence over millions of users.

Why conventional deterrence may fail

Amodei argues that existing deterrence mechanisms may not hold in an AI-dominated environment. Powerful AI could potentially undermine the nuclear deterrent by detecting nuclear submarines, conducting influence operations against their operators, or launching cyberattacks against detection satellites. The balance of power could prove fragile because AI's recursive self-improvement may create runaway advantages that competitors cannot catch up with. Even a balanced distribution of power would not prevent internal repression: multiple competing powers could each repress their own populations, the scenario in Orwell's Nineteen Eighty-Four.

Proposed defenses

Amodei sets out five policy responses intended to keep AI from enabling durable autocracy:

• Chip export controls — denying the CCP access to advanced chips and chip-making tools, which he calls "the most important single action we can take" (see Export Controls (AI)).
• Arming democracies with AI — but within limits, to "use AI for national defense in all ways except those which would make us more like our autocratic adversaries."
• Hard lines within democracies — treating domestic mass surveillance and mass propaganda as bright red lines, with a possible constitutional amendment for AI-era civil liberties.
• An international taboo — treating certain AI uses, including offensive autonomous weapons, mass surveillance, and mass propaganda, as crimes against humanity.
• Scrutiny of AI companies — governance designed for more than ordinary corporate oversight.

The Entente strategy

In Machines of Loving Grace, Amodei proposes a positive strategy for securing a democratic AI advantage, which he frames as an "Entente" with five elements. First, a coalition of democracies secures the AI supply chain and frontier capability. Second, it uses AI for military superiority (the "stick"). Third, it distributes AI benefits to a wider group of countries in exchange for their support of democratic norms (the "carrot"), an arrangement he likens to "Atoms for Peace." Fourth, it isolates the worst adversaries until they accept the bargain. Fifth, the stated goal is an "eternal 1991," a durable democratic advantage. Amodei writes that this requires close cooperation between private AI companies and democratic governments, and "extraordinarily wise decisions about the balance between carrot and stick" (Source: Machines of Loving Grace).

The measured record

Freedom House's Freedom on the Net, an annual assessment covering 72 countries and about 89% of the world's internet users, scores each country out of 100 across 21 indicators spanning obstacles to access, limits on content, and violations of user rights. Its 2025 edition, An Uncertain Future for the Global Internet, published November 13, 2025 and covering June 2024 to May 2025, recorded a 15th consecutive annual decline in global internet freedom: conditions deteriorated in 27 countries and improved in 17 (Source: freedomhouse.org).

The steepest single-year declines were in Kenya (−6), Venezuela (−4) and Georgia (−4); Bangladesh (+5) improved most. China and Myanmar tied at 9 as the most restrictive environments, against Iceland at 94 and Estonia at 91. Serbia fell from Free to Partly Free and Nicaragua from Partly Free to Not Free. Over the full fifteen years, the governments imposing the most extreme cumulative declines were those of Egypt, Pakistan, Russia, Turkey and Venezuela, each tightening control in response to challenges to its rule.

Two findings bear on the mechanisms Amodei describes. First, of the 21 indicators, the one measuring whether online information sources are manipulated by the government or other powerful actors has declined most consistently over the fifteen-year series — the propaganda channel, rather than the surveillance channel, shows the longest-running measured deterioration. Second, the decline is not confined to autocracies: half of the 18 countries rated Free lost points during the coverage period and only three gained. Germany fell 3 points to 74, on factors including criminal prosecutions over memes about politicians, self-censorship driven partly by threats from far-right actors, and hacking attributed to actors with Russian state ties. The United States fell 3 points to 73, on growing restrictions on civic space including the detention of foreign nationals for nonviolent online expression (Source: freedomhouse.org). That pattern corresponds to Amodei's second tier of threat actors — AI-competitive democracies turning defensive tools inward.

Freedom House frames a surge in government investment in AI as one of three transformations that could reset the trajectory, alongside a boom in satellite internet connectivity and tightening constraints on online anonymity. Report co-author Kian Vesteinsson argued for early intervention: "As AI and other digital technologies make rapid advances, democratic societies should work to embed safeguards for free expression and privacy at the earliest possible stages of development." Freedom House president Annie Boyajian characterized the fifteen-year trend as "persistent but not irreversible" (Source: freedomhouse.org).

Relation to policy and other concepts

Export controls are, in Amodei's account, the primary defensive tool and his strongest single policy recommendation (see Export Controls (AI)). The concept connects to AI Diffusion: Beijing is already exporting AI governance frameworks and surveillance technology. China's domestic AI regulations provide operational examples of authoritarian AI governance — the real-name verification requirements and algorithm filing with the Cyberspace Administration of China (CAC) set out in China — Interim Measures for the Management of Generative AI Services, China — Internet Information Service Algorithmic Recommendation Management Provisions, and China — Provisions on the Administration of Deep Synthesis Internet Information Services are concrete mechanisms. The concept maps to Worlds 1–4 in the Eight Worlds Framework, where China racing combined with a scenario hard to catch up in is treated as most dangerous. It stands in tension with AI Autonomy Risk: building safe AI competes with building AI fast enough to outpace autocracies.

Open questions

• How much of the fifteen-year measured decline in internet freedom is attributable to AI as distinct from the older instruments of censorship, network shutdowns and legal coercion that predate it.
• Whether the four-category account of AI-enabled autocracy, drawn from a single author writing as the chief executive of a frontier lab, holds up against independent analysis; the concept's evidence base here is narrow relative to its reliance across other pages.
• Whether the score declines in Free countries reflect the inward turn Amodei's second tier anticipates, or causes unrelated to AI.

Relationships

• depends-on: Machines of Loving Grace, The Adolescence of Technology — the two essays in which the four-mechanism account and the Entente strategy are set out
• depends-on: AI and Surveillance — the primary technical instrument in this account
• supports: Export Controls (AI) — the account supplies the strategic case Amodei makes for chip controls as the leading defensive tool
• contradicts: AI Autonomy Risk — building safe AI competes for time and resources with building AI fast enough to outpace autocracies
• instance-of: AI and Civil Liberties — the domestic-repression case is the civil-liberties question applied to state actors
• related: AI and Democracy, AI Diffusion, Eight Worlds Framework, AI and Misinformation
• related: China — Interim Measures for the Management of Generative AI Services, China — Internet Information Service Algorithmic Recommendation Management Provisions, China — Provisions on the Administration of Deep Synthesis Internet Information Services — operational examples of authoritarian AI governance

See also

• Machines of Loving Grace (Amodei, ~October 2024)
• The Adolescence of Technology (Amodei, 2026)
• Export Controls (AI)
• AI Diffusion
• Eight Worlds Framework
• AI Autonomy Risk
• AI and Surveillance
• China — Interim Measures for the Management of Generative AI Services, China — Internet Information Service Algorithmic Recommendation Management Provisions, China — Provisions on the Administration of Deep Synthesis Internet Information Services