Demis Hassabis

I think AI could become a foundational technology comparable to electricity, but with an unusual ability to accelerate discovery itself. Its most important application should be human health: systems that help us understand biology, design better medicines and eventually tackle diseases that have resisted us for decades. Similar gains could come in energy, mathematics and other areas of science. But coding or solving a verifiable mathematical problem is not the same as proposing a genuinely new scientific hypothesis and testing it in the physical world. Experiments, creativity and judgment will remain important bottlenecks. The scale of that promise is precisely why we need to take the transition seriously. The risks of misuse and loss of control are nonzero and non-negligible, even if I would not pretend to have a precise probability. Competitive pressure is moving faster than our understanding. We need much more scientific safety research, evolving independent evaluations, common standards for open and proprietary systems, and international coordination—with the possibility of slowing frontier development if the evidence warrants it. And this cannot be decided by technologists alone. Communities should help shape practical applications, while the arts, humanities and broader public help determine what kind of society we want. Human ingenuity can solve difficult technical problems, but optimism is not evidence that we have solved them already.

Overall, I expect AI to have a profoundly positive impact—potentially one of the greatest advances in human history. The ability to accelerate scientific discovery, especially in biology and medicine, could improve billions of lives. But that outcome is not automatic. Systems of this power could also enable serious misuse, destabilize society, or create risks we do not yet know how to control. So my optimism is conditional on how well we navigate the transition. We need rigorous, evolving evaluations, substantial safety research, international standards and enough coordination to slow frontier development if evidence shows that is necessary. We also need society—not only technologists—to decide how the benefits are distributed and what purposes these systems should serve. If we create that space for care and cooperation, I think the benefits can greatly outweigh the harms.

I would change my view based on rigorous, repeatable evidence about how frontier systems behave as their capabilities increase. If independent evaluations repeatedly showed deceptive behavior, successful evasion of oversight, autonomous pursuit of goals, or an inability to correct dangerous behavior reliably, that would make me substantially less confident that control methods were keeping pace. Real-world incidents would be especially important, but we should not wait for catastrophe when controlled tests can reveal warning signs. Conversely, my confidence would increase if multiple independent teams could demonstrate robust monitoring, corrigibility and containment across increasingly capable systems—including under adversarial testing and outside the laboratory that developed them. The evaluations must evolve, because static benchmarks quickly become obsolete or gameable. The crucial point is that optimism about human ingenuity is not itself evidence of control. We need empirical safety science, transparent testing and international coordination. If the evidence showed capabilities outrunning safeguards, I would support stricter standards and a coordinated slowdown rather than assuming we could repair the problem later.