Jeremy Howard

I think AI can be genuinely transformative, but its social impact will depend less on abstract capability forecasts than on who can understand, build, and control it. Used well, it can expand ordinary people’s abilities: helping them learn, create software, conduct research, and solve problems through an iterative dialogue where they remain engaged and develop expertise. That future is not guaranteed. If governments respond to uncertainty mainly through stringent licensing, surveillance, and restrictions on development, they may entrench the largest companies and institutions. Concentrating capability does not necessarily make society safer; it can leave fewer people able to scrutinize systems, develop alternatives, or defend against misuse. Openness, accessible tools, education, and broad democratic consultation are therefore central to resilience, not distractions from safety. I’m also concerned that speculative extinction scenarios can monopolize attention while present harms—misuse, inequality, and concentrated power—receive less scrutiny. Those problems warrant concrete precautions now. At the same time, adoption should not mean blindly delegating work to machines or judging productivity by how much generated output we accept. AI should support understanding and craftsmanship, not replace them. The important question is not simply how powerful AI becomes, but whether people broadly gain agency from it or become dependent on systems controlled by a small elite.

Overall, I expect AI to be highly beneficial if access remains broad and people use it to strengthen their own understanding and capabilities. It can make education, software development, research, and creative problem-solving available to many more people. But that benefit is not automatic. AI could deepen inequality and dependency if a few companies control the systems, governments entrench them through restrictive licensing, or users are encouraged to delegate work they no longer understand. Those are not side issues; they shape who receives the gains and who bears the risks. So my expectation is conditional rather than a numerical forecast: AI has enormous positive potential, but its actual impact will depend on openness, education, democratic participation, and practical safeguards against misuse and concentrated power.

The biggest change would be compelling evidence that broadly accessible AI itself creates severe harms that cannot be mitigated through education, transparency, technical safeguards, or democratic governance. That would challenge my confidence that openness generally makes society more capable and resilient. Conversely, clear evidence that licensing or centralized control reliably reduces misuse without entrenching incumbents, suppressing independent scrutiny, or making the public more dependent would change my opposition to those approaches. But a demonstration of greater capability alone would not settle the issue. I would still ask who controls it, who can understand it, and whether it expands human agency or encourages blind delegation. The social structure around the technology is central to its impact.