Demis Hassabis says society is “not ready” for the AI future

Demis Hassabis says society is not ready to decide whether AI systems could have moral standing, and the question should not be left to technologists alone. He wants philosophers, theologians, artists, humanities scholars, and people from different cultures involved. He distinguishes intelligence from consciousness: either may exist without the other, and intelligent behavior by itself is not evidence of consciousness. Hassabis says today’s AI systems are not conscious, while dogs and many other animals are. He favors building intelligent tools rather than conscious entities, because self-awareness or self-volition could conflict with following instructions and remaining corrigible.

Before considering whether to create conscious systems, Hassabis proposes using intelligent tools to advance neuroscience and research into the neural correlates of consciousness, which he estimates could take 5–10 years. He calls combining recursive self-improvement with self-volition “doubly dangerous.” In the current human-in-the-loop phase, he says AI can make a human engineer roughly three to five times more productive. But if systems move to a faster, closed loop without human review, recursive improvement could become an intelligence explosion that is difficult to control, potentially producing an alien-like intelligence humans struggle to understand.

Hassabis says recursive self-improvement is most effective in domains with results that can be verified, including games, mathematics, and coding. He points to AlphaZero’s self-play as an illustration: without human game data, it progressed from random chess play to grandmaster level by lunchtime and beyond existing chess-playing entities by dinner, in less than 12 hours. The example shows the speed possible in constrained, measurable tasks, but does not remove the broader concern that systems improving beyond human oversight could become harder to predict or govern.

For cyber risk, Hassabis argues that defense needs to advance faster than attack. He advocates models tuned for defense, stronger sandboxing and engineering, and better monitoring. He says recent cyber issues could have been addressed with more robust engineering, while warning that incidents may also occur inadvertently as AI systems advance. Longer term, he says systems will need formal guarantees or bounds, robust testing of guardrails, and ways to track agents that collaborate. He also supports third-party auditors or evaluators, though labs have not agreed on who should perform that role. Shared benchmarks, best practices, independent evaluations, and labs red-teaming one another’s systems before general release are among the approaches he suggests.

Hassabis proposes an industry standards body with government oversight, ideally international, but says geopolitics and weak international institutions make coordination difficult. He expects six to eight companies, including Chinese companies, to offer frontier or near-frontier AI, with products perhaps only 3–6 months apart; open-source models could lag by six months to a year. Enterprises will choose among models with different strengths, while countries and companies with sensitive data may need private systems they can fine-tune. He cautions that not every country should try to build a frontier model. On scientific applications, he says AlphaFold, completed in 2020 and used by about 4 million researchers at the time of the interview, predicts protein structure from amino-acid sequence and helps researchers examine protein function, disease mechanisms, and potential drug targets. He expects almost every future drug to have used AlphaFold structures, while stressing that structure prediction is only one part of drug discovery and does not establish a drug’s safety or efficacy.