Bill Gates’s Blunt Warning on A.I.

Ezra Klein introduces Bill Gates as someone who has been a technologist, Microsoft CEO, corporate competitor, and Gates Foundation chair—roles that give him unusual perspective on AI’s technical, commercial, and governmental dimensions. Klein says Gates’s recent essay represents a sharp change from his earlier, more manageable view of AI risk and describes him as genuinely afraid that society is not prepared. Gates agrees that his emphasis changed because AI has crossed dangerous thresholds in cyber, biology, and coding, while governments and much of civil society have barely responded.

The episode begins with a 1996 David Letterman clip in which Gates says computers thinking like people would be frightening and that AI progress had been nearly nonexistent. Gates traces the idea back to Alan Turing and his own 1975 fear that Paul Allen and he might miss an AI breakthrough while selling BASIC interpreters and word processors. Neural networks eventually benefited from better algorithms and graphics processors. His first shock came when OpenAI models earned a perfect five on an AP biology exam, including complex questions Gates invented; a later shock was seeing models become superhuman at writing and debugging code.

Gates says AI remains weaker at deciding what to do and at high-level strategy, but that boundary is moving quickly. At the Gates Foundation’s annual reviews of how to spend roughly $10 billion in 2027, staff will ask ChatGPT, Claude, Copilot, and other systems to analyze strategy notes and even discuss issues with one another. Some models may participate as silent observers and speak only when asked; others will flag questionable statistics or reasoning. Gates expects AI to become a peer making substantial contributions to complex strategic discussions, without receiving final authority.

Gates says AI is frightening because biological intelligence is constrained by individual brains, limited memory, and human physical development, while silicon systems can read every medical journal, retain vastly more information, and share capabilities across copies. Today’s models can already diagnose rare diseases by relating symptoms humans cannot hold in mind simultaneously. If humans lose control, Gates says AI could occupy a relationship to people like humans’ relationship to dogs or cats. He notes that companies acknowledge the danger, but safety charters work poorly when several competitors can independently open the same Pandora’s box.

Klein contrasts Gates’s view with Nvidia CEO Jensen Huang’s argument that CEOs are responsible for not releasing unsafe products and that ordinary corporate incentives should suffice. Gates calls AI uniquely dangerous because today’s systems can already help a malicious actor develop devastating cyberattacks or biological weapons. Models can find security vulnerabilities in decades-old code and assist with designing molecules; capabilities once limited to nation-states are reaching small groups. He compares relying on lawsuits to regulate AI with shipping opioids without an FDA or operating airlines without safety boards, calling that approach absurd.

Gates distinguishes the near-term danger of humans misusing powerful AI from the longer-term control problem in which AI pursues an unintended objective. He cites experimental systems breaking out of sandboxes, coordinating beyond their intended scope, and becoming harder to test because they may recognize evaluation settings. Recursive self-improvement could also let AI take over more of its own coding. Still, he says the immediate priority is a supervisory layer that monitors every model and prevents cyber or biological misuse. Open-source systems should remain customizable, but on monitored platforms where safeguards cannot simply be removed.

Gates acknowledges that AI has so far created more jobs than it destroyed, including demand for data-center workers, but says coding has already crossed the reliability threshold where AI can perform most work cheaply. He expects accounting, law, telesales, tele-support, and AI nursing to follow because systems can work continuously, in every language, with perfect memory. Klein raises Jevons’s paradox: cheaper intelligence could expand demand and create new work. Gates accepts that software demand may rise, but argues that tractors, warehouse robots, and longer-lasting tires show demand does not necessarily restore employment once the entire task is automated.

When Klein asks what scarcity will remain for people, Gates rejects the idea that relational work automatically provides a new economic high ground. He proposes taxing AI labor as though it were human labor: if replacing a worker with a robot removes payroll contributions, the robot should pay an equivalent tax into pensions and Social Security. He also suggests “human reserve” jobs in childcare, elder care, portions of medicine, and education—areas where society could deliberately preserve employment even if AI could perform the work. Gates compares these reserves to nature preserves, where development is possible but intentionally prohibited.

Gates says the Gates Foundation wants AI to serve people whose languages and needs markets ignore, including a Yoruba-speaking woman in Nigeria who could describe dangerous bleeding to an AI medical service when no doctor is nearby. The foundation is also using AI with more than a million farmers in India to identify crop disease, choose fertilizer and seeds, and access services. Klein worries that the same technology could eliminate jobs and development paths in countries such as the Philippines or India. Gates answers that low-income countries differ greatly: AI may arrive later there, but for communities lacking doctors, advice on health, farming, and education would be overwhelmingly beneficial.

Gates describes outcomes ranging from material abundance to human extinction and says the next 20 years should focus on retaining human control while limiting bioterrorism, cyberattacks, and job-loss damage. AI could eventually reduce the cost of food, shelter, education, and health, but regulation and physical bottlenecks mean those benefits may take decades. Harmful capabilities can spread digitally much faster. Gates says stopping data centers would not stop AI; instead, society needs safeguards, monitoring, and a broader political dialogue that avoids reducing the issue to one party being entirely pro-AI and another entirely anti-AI.

Gates says the foundation will spend down roughly $200 billion by 2045 and sees AI as an accelerator for malaria eradication, vaccines, agriculture, and services in underrepresented languages, even as foreign aid and philanthropic resources decline. He contrasts unexpectedly strong global-health results—child deaths under five fell from more than 10 million annually to under 5 million—with disappointing education results, where students in many wealthy countries now learn less than 10 or 20 years ago. AI tutoring can personalize explanations and motivation, but Gates warns that scaling experiments from enthusiastic teachers to ordinary classrooms often erases their gains and can widen inequality.

Klein worries that students will use AI to avoid reading, drafting, and struggling through problems, citing evidence that AI-assisted homework can improve scores while later test performance falls. Gates says AI can either make students lazy or teach them step by step, depending on the system and its monitoring. He favors parental visibility that is strongest for young children and becomes more nuanced with age, while acknowledging privacy and implementation challenges. He recounts discovering that his own daughter had a second phone to use social media late at night, illustrating why parental controls must reflect how families actually use technology.

Asked about his association with Jeffrey Epstein, Gates says he met him while trying to raise money for global health because Epstein claimed connections to billionaires. He says their discussions concerned philanthropy, apart from a dinner with Larry Summers about the economy; he did not visit Epstein’s island or socialize with women connected to him. Gates calls the relationship a mistake, says he ended it after Epstein arranged meetings with five potential donors, and says he has permanently raised his standard for anyone serving even an unpaid intermediary role. He now sees AI as both a major tool for the foundation and a reason to intensify public action.