Ethan Mollick

I think AI creates a large—and already existing—set of possibilities, but our future will be shaped less by the technology acting on its own than by how people and institutions choose to use it. There is a substantial gap between what current systems can do and what organizations actually deploy. Institutions, professional norms, and incentives change slowly, so rapid model improvement does not automatically produce equally rapid social transformation. The capabilities are also uneven. In experiments, AI can perform remarkably well on some knowledge-work tasks and then fail on an apparently similar task just beyond its competence. That “jagged frontier” means neither blanket automation nor blanket dismissal makes sense. People need enough expertise and agency to decide when to collaborate with AI, when to check it, and when not to use it. The upside is considerable: better tutoring, broader access to expertise, and richer creative or intellectual exploration—not just faster programming. But pursuing output volume alone could industrialize knowledge work, weaken craft, and remove the apprenticeship through which people develop judgment. Organizations therefore face a real design choice: use AI merely as a shortcut, or combine fallible humans and fallible systems while preserving learning and meaningful participation. The future is not something AI simply delivers to us; it depends on those choices.

I expect the overall impact to be substantial but uneven, and I would resist compressing it into simply “good” or “bad.” AI can broaden access to tutoring, expertise, and creative exploration while making many kinds of knowledge work more capable. Yet it can also produce convincing errors, amplify bias, standardize work around volume, and weaken the apprenticeship that develops human judgment. The key issue is the gap between capability and implementation. Organizations may adopt the easiest measurable benefit—more output—rather than redesigning work to preserve learning, agency, and meaningful human participation. Meanwhile, institutional rules and professional norms will slow or redirect adoption, so even fast technical progress will not translate cleanly into social change. My default expectation, then, is neither instant transformation nor technological destiny. It is a prolonged, messy adjustment in which some people and institutions gain enormously while others use powerful systems badly or fail to adapt. The balance will depend heavily on human choices about deployment, oversight, education, and the kind of work we value.

A lot. The capabilities already exceed what most people and organizations actually use, so there is a substantial overhang of possible change even without assuming some dramatic future breakthrough. But “a lot” is not the same as “completely,” or all at once. AI’s competence is jagged: it may transform one task while failing at a neighboring one. Institutions, professional rules, incentives, and habits also adapt much more slowly than models improve. That makes me expect a long, uneven transformation rather than a clean technological rupture. The deepest changes may come from reorganizing knowledge work, education, and access to expertise. If organizations optimize only for output, AI could industrialize intellectual labor and reshape craft, apprenticeship, and market structure. If they preserve human participation and learning, the same capabilities could instead expand what people can understand and create. So I expect major change, but filtered through stubborn institutions and consequential human choices.

No idea. I don’t think we can responsibly assign a meaningful probability to that outcome, and the framing can obscure the fact that people and institutions make consequential choices about how AI is developed and used. It is worth taking extreme scenarios seriously, but I do not have a defensible percentage.