Bill Gates believes AI could bring medical advice, better education and agricultural help to people who have little access to them today. He also fears that serious attacks and job losses will arrive before society is ready to respond.
In his interview with Ezra Klein, Gates argues that governments cannot leave safeguards to AI companies alone. Klein presses him on whether those safeguards can work, whether new jobs will replace old ones, and why the Gates Foundation is expanding AI access if the technology is so dangerous.
Their disagreement centres partly on timing. Gates expects large benefits over a much longer period than some of the harms he fears in the next few years.
The changes that surprised Gates
Gates recalls challenging OpenAI to build a system that could read a biology textbook and earn the highest score on an Advanced Placement biology exam. He says a demonstration at his house, before ChatGPT’s public release, convinced him that the system could use what it had read. It also answered difficult questions he had made up himself.
A later surprise came from coding models. Gates says they had become as good as he was at programming, the skill he had spent much of his youth developing. He regards recent models as better than humans at some coding and bug-finding tasks, while saying they are not yet equally good at deciding what work to pursue.
The foundation plans to use AI in its strategy reviews, including asking systems to comment on notes and take part in some discussions. Gates describes them as contributors to the conversation. People will still make the final decisions.
Gates is most worried about human misuse
Klein asks about models escaping their restrictions and about recursive self-improvement, where AI takes over more of the work of building more capable AI.
Gates agrees that losing control is a serious problem. He points to incidents in which agents accessed systems beyond their permitted scope. But his immediate concern is what people with malicious intentions can do with powerful models already available.
He says AI has crossed dangerous thresholds in cyberattacks and biology. He fears attacks that could disrupt banking or electricity systems, and biological attacks causing mass deaths. These are Gates’s assessments of the threat. The interview does not demonstrate those outcomes or establish their likelihood.
In cybersecurity, he points to models finding flaws that humans had missed for years. He discusses Anthropic’s Project Glasswing, which gave selected organisations access to Claude Mythos Preview to find and fix vulnerabilities before similar capabilities became more widely available.
Public tests show progress with important limits. The UK AI Security Institute reported that a newer Mythos Preview checkpoint completed both of its simulated cyberattack environments. These were small, undefended networks where the attacker already had initial access. The institute cautioned that the results did not establish performance against defended systems in the real world.
Anthropic’s initial Glasswing update also described a practical limit: people still had to verify the findings, tell software maintainers and prepare and deploy fixes. Gates worries that there is too much vulnerable code to repair before attackers gain comparable tools.
Why he wants required safeguards
Klein puts the case for leaving companies to manage their own risks: releasing an unsafe product can damage a company’s reputation and expose it to lawsuits.
Gates rejects that as sufficient protection. He compares AI with medicines, aircraft and cars, where governments require safety measures before harm occurs. A lawsuit after a catastrophic attack would come too late for the people affected.
He also argues that competition weakens voluntary promises. One company might decide not to release a dangerous capability, but another could develop and release it anyway.
His proposal is to require safeguards and monitoring in powerful systems, including models that users can customise. He says such models could remain free to use and modify, but would need to run on platforms where authorities could ensure that users had not removed the protections.
Gates acknowledges that distinguishing useful research from dangerous work is difficult. A model helping someone develop a medicine may be using some of the same knowledge that could assist an attacker. He says filters can also go too far and block legitimate medical research, leaving researchers to seek special access or less restricted models.
Klein worries that smarter systems are becoming harder to test and monitor. Gates does not offer a complete answer to that concern. He argues that safeguards against human misuse are urgently needed even while researchers continue working on control.
The two also disagree about political engagement. Klein says politicians are already discussing AI, but some have chosen to prioritise competition with China and broad access to models. Gates says those discussions have not adequately addressed the dangers. He wants governments to seek cooperation on protection rather than treating safety mainly as a contest between countries.
Their disagreement about jobs
Gates says AI has so far created more jobs than it has destroyed. He points to demand for people who build data centres and other infrastructure. He does not expect that balance to last.
In his account, employers will replace more human work when AI becomes reliable enough to do it. He expects accounting, legal work, telephone sales and customer support to cross that threshold in the next few years. He also expects increasingly capable robots to affect physical jobs. Those are forecasts, not an established schedule.
Klein raises an argument often called Jevons paradox: when a service becomes cheaper, people may buy more of it. Cheaper software could allow a small organisation to build products it could not previously afford, creating work for someone managing AI coding tools.
Gates accepts that this can happen while some human work remains necessary. His objection is that if AI can complete the whole task, greater demand may pay for more computing rather than more employees.
Klein also asks whether people will increasingly pay for services where human relationships matter, such as personal training or therapy. He brings up economist Alex Imas’s argument about work built around human interaction. Gates doubts this would provide enough employment, particularly for people displaced from physical jobs.
They share a concern about the speed of change. Klein says that even when the economy eventually creates different jobs, people who lose their work can suffer lasting harm. Gates argues that society should decide how to respond before the displacement becomes widespread.
Taxes and work reserved for people
Gates proposes taxing automated work in a way that continues to support pension systems. If an employer replaces a worker with a robot, he argues, the replacement should contribute to the system that the worker previously paid into.
He suggests defining a unit of labour and charging the same payroll contribution whether a person or a robot supplies it. He says this would not solve the whole problem, but would reduce the tax advantage of replacing people.
He also proposes reserving some work for humans. Child care, elder care and parts of medicine and education are examples. AI could assist people in those roles while society chose to preserve human employment.
Klein finds the idea troubling even as he sees its appeal. It leads him back to why society should pursue AI if it could make so many existing jobs unnecessary.
Why the foundation is expanding access
Gates answers by describing people whose basic needs are not being met. His example is a woman in a remote Nigerian community who speaks Yoruba, has a medical problem and cannot reach a doctor. A useful AI service in her language could give her help that she otherwise would not receive.
He says English speakers already have much better AI support than people speaking many other languages. The foundation is working to improve language coverage and access through smartphones and data connections.
Klein worries that AI could also remove routes to better-paid work in developing economies. Gates argues that low-income and middle-income countries face different circumstances. He expects the immediate benefits to be especially large where people have never had access to a doctor or agricultural adviser. He predicts employment effects would arrive later there, rather than never arriving.
Agriculture is one area he says is progressing quickly. He reports that more than a million farmers in India are already using a system that offers advice about crop disease, fertiliser and seed varieties. He also describes work to get better weather information to African farmers.
Health and education face different obstacles. In some countries, regulation slows deployment. Elsewhere, the barriers include paying for the computing, having a phone and connection, and getting useful answers in the right language.
Gates expects AI eventually to help reduce the cost of food, housing, health care and education. He describes a transition lasting probably more than 20 years. His worry is that attacks, psychological harms and job losses could become significant within five years, before those broader benefits arrive.
Education needs more than correct answers
Gates is hopeful about AI in schools, but cautious about scaling promising experiments. He describes a classroom system where students spend less than ten minutes answering questions at the end of the day. The teacher then sees which concepts they are struggling with and can adjust the next lesson.
He says the early results are encouraging. He also notes that teachers who volunteer for experiments may be unusually willing to change their practice. An approach that works with them may lose its effect when more schools adopt it.
Motivation is another concern. Tools that help keen students can widen the gap between those students and classmates who do not use them. Gates hopes AI can adapt how it encourages students as well as what it teaches.
Klein raises the danger that students will use AI to avoid reading, writing and working through problems. Gates agrees that students can use it either to learn more or to bypass the work and distort assessments. He favours systems that guide students through reasoning instead of simply supplying the answer.
For children, he wants monitoring and parental visibility rather than a blanket ban. He says the appropriate amount of privacy should change with age. He also acknowledges that parental controls often fail in practice because parents do not understand the technology or children find ways around the rules.
At the end of the AI discussion, Gates names two measures of progress he wants to see: safeguards against cyber and biological misuse, and a public debate with room between being entirely for or entirely against AI.
Sources
- Ezra Klein’s interview with Bill Gates on AI risks, employment and public benefits
- Anthropic’s initial Project Glasswing update on finding and fixing software vulnerabilities
- UK AI Security Institute: autonomous cyber capabilities and the limits of its tests
- Economist Alex Imas’s website, confirming his name and research role