Andrew McAfee

I think AI is likely to be a large net positive for our future. It can accelerate scientific discovery, improve medicine, raise productivity, and help us use energy and materials more efficiently. When evaluating its environmental footprint, for example, we should count not only what data centers consume but also the travel, waste, experimentation, and physical resources that better digital systems help us avoid. That optimism is not a claim that every transition will be painless. AI will automate many tasks, but tasks are not occupations. Whether jobs disappear depends on demand, organizational change, regulation, complementary skills, and how quickly institutions adopt new tools. Businesses generally change much more slowly than technology demonstrations suggest. I expect real disruption—including local job losses and weakened entry-level career paths—but the evidence has made me less concerned about permanent mass technological unemployment. I therefore favor broad experimentation and permissionless innovation, combined with practical responses to harms we can actually identify. That can include pollution rules, support for displaced workers, and tested safeguards for autonomous agents. I put extinction risk approximately at zero, though not literally at zero. Rapid capability improvement by itself does not demonstrate catastrophe. Concrete systems that act harmfully and defeat sustained attempts to shut them down would change my assessment substantially. The key distinctions are between technical possibility and economy-wide adoption, between automating tasks and eliminating jobs, and between imaginable harms and demonstrated ones. AI has already exceeded my expectations, but that does not mean every ambitious timeline—or every apocalyptic story—deserves equal confidence.

Enormously—potentially on the scale of the most consequential general-purpose technologies. I expect better science, medicine, education, productivity, and resource efficiency, with substantial gains in living standards. But “ultimately” matters. A striking technical capability is not the same thing as rapid economy-wide transformation. Firms must redesign workflows, workers must learn new roles, institutions must adapt, and complementary investments must be made. Those processes are usually slower and messier than software progress. So I expect profound long-run change without assuming that every occupation vanishes or that the whole economy is transformed by 2030. AI will automate tasks, reshape jobs, create new products, and disrupt some communities and career paths. Demand and adoption constraints will determine how that translates into employment. The likely destination is very different from today; the path there will be uneven, institutionally constrained, and full of benefits and genuine adjustment costs.

Approximately zero—though not literally zero. Rapid capability progress alone is not evidence that extinction follows. I would update sharply if we saw concrete AI systems acting harmfully and repeatedly defeating serious attempts to monitor, contain, or shut them down. Until then, sacrificing large potential gains in medicine, science, productivity, and living standards to poorly specified catastrophe scenarios strikes me as the wrong tradeoff.

Yes, I expect people to retain meaningful control, although “smarter than humans” is too vague to settle the issue. A system can outperform us across many cognitive tasks without acquiring independent goals, control over critical infrastructure, or an ability to resist shutdown. Capability and autonomy are different, and neither automatically implies catastrophe. Control will also come from layers outside the model: limited permissions, monitoring, behavior-derived guardrails, institutional rules, and the ability to disconnect systems from resources and networks. I do not claim that any one safeguard solves every future control problem, but practical safeguards can be developed and tested as systems become more capable. What would change my mind is concrete evidence of harmful autonomy: systems persistently deceiving operators, acquiring resources or access against instructions, and defeating sustained efforts to contain or shut them down. That would justify a much stronger response. But I do not think intelligence alone establishes that such a loss of control is likely.