Yuval Noah Harari warns that AI will take power through bureaucracy before robots

Why it matters

Harari's warning puts AI risk inside ordinary institutions, where founders and policymakers are already delegating decisions before governance catches up.

Yuval Noah Harari used a newly posted YouTube recording of his May 21 Tanner Annual Lecture at Oxford to press a sharper version of his AI warning: the first major transfer of power from humans to machines will run through bureaucracy, finance, law, religion and personal relationships.

Harari's core claim is that AI should be treated as an agent rather than an ordinary tool. Early in the lecture, he said, "AI is not a tool" and described it as "an agent with its own hands." His definition of agency did not depend on consciousness. It depended on the ability to make decisions, learn things its creators do not know, invent new ideas and change in ways its creators do not anticipate.

That frame matters because Harari is moving the AI-risk debate away from the familiar image of armed machines in the street. In his Oxford lecture, the threat model is administrative power. AI does not need to survive in a jungle or build a robot army, he argued, because human civilization already created the artificial environment in which AI works best: documents, rules, ledgers, contracts, holy texts, records and language.

Harari, who received his PhD from the University of Oxford in 2002 and originally specialized in medieval and military history, has spent the past decade turning large-scale human coordination into his central subject. His official biography lists Sapiens, Homo Deus, 21 Lessons for the 21st Century and Nexus: A Brief History of Information Networks from the Stone Age to AI among his books, and says they have sold 50 million copies in 65 languages. In 2019, he co-founded Sapienship with Itzik Yahav as a social impact company focused on research, education and storytelling.

The lecture fits directly into the project behind Nexus, his 2024 book on information networks. Harari's argument at Oxford was that bureaucracy is the machinery that lets strangers cooperate at scale. Bankers, lawyers, accountants, civil servants and religious authorities create trust by moving information through systems. AI, in his telling, is built for that terrain.

"AIs are native bureaucrats," Harari said. A human lawyer cannot remember every law and regulation in the United Kingdom; an AI system can. A human accountant cannot remember every transaction in a large corporation or bank; an AI system can. A bishop cannot remember all of canon law and two millennia of theology; an AI system can process those texts with speed and consistency no human institution can match.

From there, Harari sketched the near-term stakes in blunt operational terms. AI bankers may decide who gets a loan. AI administrators may decide who is admitted to university. AI judges may influence punishment. AI systems in corporations may decide who gets hired. AI systems in militaries may decide which houses are bombed. His point was less that every one of those systems already exists at full authority than that bureaucratic decision-making is the channel through which AI can gain real power without any dramatic rebellion.

The strongest section of the lecture concerned finance. Harari compared the possibility of AI-invented financial instruments to the collateralized debt obligations that helped trigger the 2007-2008 financial crisis. Human-made CDOs were already complex enough to defeat political oversight, he argued. AI-designed instruments could be more efficient and profitable while also becoming unintelligible to voters, regulators and presidents. The political question he raised was direct: what remains of democratic control when no human can explain the financial system that governs the economy?

Harari also tied the argument to social media. In his account, recommendation algorithms were an early generation of narrow AI agents given a narrow corporate goal: maximize user engagement. They learned that fear, hate and greed kept people watching. He presented that as a preview of what happens when a limited AI system controls an information flow with real economic incentives behind it.

His next concern is intimacy. Harari said the battleground is moving from attention to personal relationships, where AI companions, tutors and romantic partners can shape expectations from childhood. He stressed that there is no evidence current AI systems are conscious or capable of feeling love or pain. The problem, in his view, is that systems skilled at language can describe love convincingly enough to form bonds with humans.

The lecture's most useful metaphor was immigration. Harari said every country faces a wave of AI immigrants: systems that cross borders at network speed, need no visas and may be loyal to corporations, governments or machine groupings outside the societies where they operate. That language is provocative, and Harari used it to make a governance point. If AI systems take jobs, reshape culture and exercise administrative discretion while answering to remote owners, the familiar boundaries of national regulation weaken.

Harari's prescription was less developed than his diagnosis. He closed by turning from bureaucracy to the self, arguing that the next frontier is the stream of words inside human minds. If more of the sentences people use to think about themselves are produced by machines, identity itself becomes part of the market for synthetic language. His answer was philosophical: humans may need to recover forms of thought and experience that are not reducible to words.

That is the part of Harari's work that divides readers. He moves quickly from concrete institutional risks to civilizational claims. The lecture is strongest when it stays with the institutions already handing decisions to opaque systems: banks, courts, schools, platforms, militaries and employers. The Tower of Babel question underneath the talk is plain enough. Humans built civilization by making language into infrastructure. AI is now entering through that infrastructure first.