Demis Hassabis

I think AI could become a foundational technology comparable to electricity: not merely another consumer product, but a tool that transforms science, medicine and much of society. The application I care most about is human health. Systems such as AlphaFold show how AI can help us understand biology; the larger ambition is to accelerate drug discovery and, eventually, help solve diseases that have resisted us for decades. But progress will not be uniform or automatic. Solving a mathematical problem that can be checked immediately is different from proposing a genuinely new scientific hypothesis, designing an experiment and testing it in the physical world. AI may greatly accelerate reasoning and coding while experiments, scientific creativity and real-world validation remain bottlenecks. This is also a pivotal and risky transition. I do not attach a precise probability to catastrophic outcomes, but I regard the risk as nonzero and non-negligible. Competitive pressure is not giving us enough time to understand increasingly capable systems. We need rigorous, evolving evaluations, independent standards and international coordination, with the possibility of slowing frontier development when the evidence warrants it. That can coexist with a strong open scientific ecosystem. Ultimately, technologists cannot decide alone what this power is for. Scientists, policymakers, the humanities and affected communities all need a role in shaping the society we want. The opportunity is immense, but realizing it requires the same seriousness about safety and human purpose that we bring to capability.

Overall, I expect AI to have a profoundly positive impact—potentially improving human flourishing on a scale comparable to foundational technologies such as electricity. The greatest gains could come through science: understanding biology, accelerating medicine development, advancing energy research and making expertise far more accessible. But these benefits are an ambition, not an automatic consequence of building more capable systems. The transition could also be extremely disruptive. Misuse, loss of control, economic dislocation and concentration of power are serious concerns. I do not assign a precise probability to catastrophic outcomes, but I consider the risk nonzero and non-negligible. Competitive dynamics make this harder because they encourage deployment before we have adequately understood the systems or prepared society. So my positive expectation is conditional on substantial work: scientifically grounded evaluations that evolve with capabilities, independent standards, international coordination and a willingness to slow frontier development if evidence warrants it. We also need broad participation in deciding how the benefits are distributed and what ends these systems should serve. If we create enough room to solve the technical and societal problems, AI could be one of humanity’s greatest achievements. If we simply race ahead, that outcome is far less secure.

I think it could change the world as profoundly as electricity—reshaping science, medicine, work and society rather than simply producing better consumer applications. We may be in “the foothills of the singularity”: approaching systems that dramatically accelerate coding, mathematics and parts of scientific discovery. But even extremely capable AI will not make every constraint disappear. In science, generating a plausible hypothesis is not the same as designing and conducting experiments, validating results in the physical world and turning them into safe medicines. Human judgment, creativity, laboratories and institutions will remain important bottlenecks. The scale of change therefore depends on how we manage the transition. The upside is enormous, especially for health and other difficult scientific problems, but so are the risks of misuse, loss of control, disruption and concentrated power. Rigorous evaluations, international standards and, where warranted, coordinated slowing are essential. Just as importantly, technologists cannot decide by themselves what society should become; the public, policymakers, arts and humanities must help determine the purpose and distribution of these capabilities.

I do not have a precise percentage. My gut view is that the risk is nonzero and non-negligible—serious enough to justify major scientific work on control, rigorous evaluations, international coordination and slowing frontier development if the evidence warrants it. Assigning a neat number would imply more knowledge than we have. The uncertainty is itself part of the danger, especially under competitive pressure.