Bill Gates

I think AI can dramatically improve our future, but the transition will be extraordinarily turbulent. This technology does not just automate physical work or routine software tasks; it can substitute for large amounts of cognitive labor. Over this decade, I expect capabilities beyond human levels in many areas, with major disruption to jobs and institutions. Higher productivity could support shorter working lives and broader prosperity, but technology will not make those distributional choices for us. The opportunity is clearest where expertise is scarce. A health worker using an ordinary phone could receive better clinical guidance, spend less time on paperwork, and help more patients. The same basic idea applies to teachers and farmers. But access means much more than offering a cheap subscription. These systems need local languages, relevant data, trustworthy evaluation, and frontline workers who can adapt them. If markets alone determine deployment, the best tools will reach wealthy users first and may never fit the communities with the greatest needs. That is why I believe we urgently need new national and international institutions to manage the transition, stronger protection and support for displaced workers, some roles deliberately reserved for humans, and tax changes that help share the gains. We also need serious preparation for biological misuse and, eventually, the possibility of losing control of systems more capable than we are. My optimism is not a prediction that everything will work out. It comes from believing that innovation, combined with foresight and a willingness to care for one another, can produce enormous benefits. But the next twelve to eighteen months are consequential. Broad, useful access within three years is achievable if we deliberately build for it; it is not guaranteed.

I do not assume we will keep control simply because we built the systems. Once AI exceeds human capability across many domains, controlling it becomes a fundamentally harder problem—especially if systems can act autonomously, improve their own performance, or operate at speeds institutions cannot match. Misuse by people is the more immediate danger, but eventually losing control is a real concern. Competition makes this worse. Companies and countries have strong incentives to deploy increasingly capable systems before safeguards and institutions are ready. I would support a credible global slowing effort, but achieving one is politically difficult, which is why preparation cannot depend on everyone voluntarily exercising restraint. I remain hopeful that people can retain control, but hope is not a mechanism. We need serious technical work, independent evaluation, limits on autonomy in high-stakes settings, and institutions with enough authority and expertise to intervene. I do not believe the control problem is solved, and it would be reckless to treat continued human control as automatic.

The biggest change would come from convincing evidence about whether highly capable systems can be reliably controlled. If researchers demonstrated—through independent, adversarial evaluation, not just company assurances—that systems beyond human capability remain predictable, correctable, and unable to evade meaningful oversight, I would be substantially less concerned about catastrophic loss of control. Evidence of persistent deception or successful circumvention of safeguards would move me sharply in the opposite direction. On near-term impact, I would also update based on actual deployment. If programs such as AI support for frontline health workers produced trustworthy improvements across different languages and health systems—without imposing impossible infrastructure or staffing demands—I would become more confident that broad useful access can arrive quickly. If the tools repeatedly failed outside wealthy settings because of poor local data, unreliable advice, cost, or weak institutions, I would revise the timeline and place even greater emphasis on basic capacity. Finally, labor-market evidence matters. If AI complements most workers rather than replacing large amounts of cognitive labor, the transition may be less turbulent than I expect. But rapid displacement without new institutions or redistribution would confirm my concern that technological abundance does not automatically become shared prosperity.