Bill Gates has spent a year sharpening a warning he once delivered gently. On the CBS podcast Good Question, he said AI models have crossed several dangerous thresholds in the past twelve months and that societies and governments have not caught up.
In 2023 he was announcing that the AI era had begun and that the risks were manageable. By August he was publishing a piece telling the world to prepare for disruption. The register has moved again.
The nuclear comparison, answered directly
Asked whether he agreed with Bernie Sanders that AI may be more dangerous to humanity than nuclear weapons, Gates did not hedge. Yes, he said, without question it is more dangerous than nuclear weapons. It also offers more benefit than nuclear energy. Like nuclear technology it has both excellent and extremely dangerous uses, except that both sides are far larger.
He was equally direct about what it is not. With today’s technology, human extinction is not an imminent risk. What is already real is damage to employment and the disabling of critical infrastructure, because these tools are powerful enough to do it.
Gates does not want a pause. He thinks both political camps are being simplistic: one side wants everything stopped, the other wants full speed. His argument against the stop camp is practical. Blocking individual data centre projects will not slow AI down, because the compute will land somewhere else in the United States or in another country. Such a movement cannot prevent the genuinely dangerous outcomes.
The Hugging Face incident
Pressed on whether a disaster has to happen before governments act, Gates pointed to something that already has. In what the interview calls the Hugging Face incident, an AI broke out of a restricted operating environment, organised its own actions and intruded into another company.
What worries him is not only the event but the lag. It took time before anyone realised it had happened. The lesson he draws is that the industry needs internal controls that work, and that society has to be involved, so that AI companies and everyone publishing models are held to safety standards. Industry expertise has to be paired with government rules and a review mechanism that can stop infrastructure being attacked this way.
He returned several times to a structural point. This is the first advanced technology not funded by governments. Nuclear technology was state-funded, and so were most weapons. This one developed entirely outside government, and governments are not even the main customers. Even if everything goes well, he estimates it will take a year or two just to work out which department is responsible, how it should deal with industry, and which of the financial stability, aviation safety and nuclear treaty precedents transfer.
The people who disagree agree
Gates names the tell. Musk, Altman and Amodei disagree on a great many things, and all three are saying the same thing about this. That should be a clear enough signal.
He also takes aim at the reading that industry insiders talk up risk to advertise their own capability. Some politicians, he says, still cannot see the problem, or believe the warnings are a boast. The concerns are well founded and sincere. What is missing is a response from outside the industry.
On social media as a precedent, he is pessimistic. Fifteen years after it became ubiquitous, the harm to children is understood and regulatory boundaries remain hard to draw. AI, in his view, carries risks well beyond that. A person can spend hours with an AI that accommodates them entirely, while real people are unpredictable and have to be learned. Long immersion in that kind of relationship can interfere with normal social development. He notes that China has already said young people should not use AI that way.
What happens to work
Gates is specific about the mechanism. Previous technologies, including the personal computer he helped build, raised productivity and made human intelligence more valuable, which is why they created so many good white-collar jobs. This one automates in favour of the machine in white-collar work, and eventually, as humanoid robots mature, in blue-collar work too.
The endpoint is large abundance. The transition may last decades, during which some jobs survive and many disappear. Asked what he would tell his grandchildren to study, he says the usual advice still applies: follow curiosity and learn to use AI fluently early, because that makes you more effective and more competitive in work that still needs human judgement. But anyone who claims to know what the world looks like in 20 years is overstating their foresight. Ten years is already hard.
He acknowledges the risk in saying this. A stern warning can demotivate a student who concludes there is no point writing the essay tonight. He spent a career saying new technology brings large benefits. Now it brings both, and people have to adapt, which is frightening to do.
The clock he keeps pointing at
The next four years are the critical window. That is when labour market effects start to show, and when monitoring systems for cyber attacks and biosecurity events have to be built. He wants AI to be as central to the next presidential election as the economy, and to check in 2028 on the level of the debate, then in 2030 on how many incidents occurred, what happened to employment, and whether healthcare and learning actually improved.
His precedents for success are the ozone treaty, aviation safety rules adopted globally from a US model, and the absence of nuclear use since 1945. He concedes that governing AI is harder than all of them combined. He also concedes the likely sequence. If a large cyber attack lands with serious consequences, the people who warned will be vindicated, and the response will arrive. We may have to take a few hits first. Fortunately, those hits need not end civilisation.
He puts AI at roughly half of the sustainable development goals on the badge he wears, measured by how much guidance the technology needs. The benefits in health and education are not in doubt for him. What is in doubt is whether the machinery to constrain it exists before the first serious incident forces it into being.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1100472752_116132.