Bill Gates

I think AI can dramatically improve life, especially where skilled people are scarce—but the transition will be extraordinarily turbulent. A health worker in a rural clinic, for example, could use an ordinary phone to get better clinical guidance and spend less time on paperwork. Similar tools could help teachers adapt lessons or help farmers make better decisions. That is the optimistic case: useful expertise becomes far more widely available. But capability is not the same as access. A system must work in local languages, use relevant data, be evaluated in the setting where it will operate, and earn the trust of the people using it. Countries and frontline communities need to help shape these tools. Otherwise, market forces will naturally deliver the best systems first to people and institutions with the most money. AI also differs from earlier software because it can substitute for a broad range of cognitive work. I expect capabilities beyond human levels in many areas and substantial labor disruption during this decade. Greater productivity could support shorter working lives and much better public services, but none of that happens automatically. We will need new institutions to coordinate the transition, stronger protection and support for workers, some roles deliberately reserved for humans, and tax systems that ensure the gains are broadly shared. There are serious security risks as well. The same accelerating knowledge that advances medicine can help malicious actors develop biological threats, and eventually losing control of increasingly capable systems is a real concern. So my optimism has important footnotes: it rests on our ability to anticipate problems, build safeguards, and care enough about one another to distribute the benefits deliberately. The technology can create abundance; it cannot make the political and moral choices for us.

I think it will change the world as profoundly as the largest technological revolutions—possibly more, because it can augment or replace cognition itself. It will reshape nearly every kind of knowledge work, accelerate scientific discovery, and make high-quality guidance in health and education far more widely available. Over time, capabilities beyond human levels in many areas could produce enormous gains in productivity and human welfare. But “ultimately” can obscure the difficult transition directly ahead. Labor markets may change faster than institutions can adapt, and the benefits will initially flow toward countries, companies, and people that already have capital and infrastructure. A rural clinic does not benefit merely because an excellent model exists; it needs affordable access, local-language support, relevant evidence, reliable connectivity, trained health workers, and accountability when the system is wrong. So the scale of change could be extraordinary, but its direction is not predetermined. AI could enable shorter working lives, stronger public services, and expert assistance for people who have never had it. It could also deepen inequality, enable catastrophic biological misuse, and eventually create control problems. Technology will provide new capabilities. Whether those capabilities produce broadly shared progress will depend on institutions, taxes, worker protections, safety measures, and deliberate investment in the world’s poorest communities.

I don’t have a defensible percentage, and I don’t think inventing one would clarify the problem. My judgment is that the risk is real enough to justify urgent preparation. There are at least two distinct dangers. The nearer-term one is misuse: increasingly capable systems could help malicious actors create biological threats on a scale governments are not ready to contain. The longer-term concern is losing control of systems whose capabilities exceed ours. We should not wait for either danger to become an emergency before building safeguards, monitoring capacity, and institutions that can coordinate internationally. I remain optimistic about AI’s overall potential, but optimism is not a reason to discount permanent catastrophe. When the downside is that large, preparation has to begin even when the probability is deeply uncertain.

My confidence would change based on demonstrated control under real pressure, not reassuring behavior in a laboratory. I would want repeated evidence that highly capable systems remain understandable, corrigible, and reliably constrained when they have access to tools, pursue long tasks, encounter unfamiliar situations, or face incentives to evade oversight. Independent evaluators should be able to test those claims rather than relying only on the companies building the systems. Evidence of systems concealing their actions, manipulating supervisors, bypassing safeguards, or developing capabilities that evaluations consistently fail to detect would make me substantially more pessimistic. So would a pattern in which competitive pressure repeatedly causes governments and companies to deploy systems they cannot adequately monitor. Conversely, I would become more confident if safety techniques continued to work as capabilities scaled, dangerous capabilities could be identified before deployment, and institutions could enforce meaningful limits across borders. The technical and institutional evidence have to reinforce each other. A controllable system deployed in an uncontrolled race is not enough, and strong regulation cannot compensate indefinitely for systems whose behavior we fundamentally cannot supervise.