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

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

In this live Masters of Scale conversation, Reid Hoffman speaks with Demis Hassabis, co-founder and CEO of Google DeepMind and chief scientist at Alphabet, about the race to AGI, AI’s role in scientific discovery, the risks of recursive self-improvement, and why he believes society is not yet ready to confront conscious AI. Hassabis argues that the best near-term path is to build powerful intelligent tools—especially for medicine and science—while strengthening safety, oversight, and international coordination before AI systems become more autonomous or potentially conscious.

Key Takeaways

• AI should primarily be used to improve human health and scientific understanding.

  • Hassabis sees medicine and biology as the most important early applications of AI.
  • He frames AI as a tool to accelerate discovery, cure disease, and improve the human condition.

• AlphaFold remains a landmark example of practical AI impact.

  • It solved protein folding at unprecedented scale and speed.
  • He says it has already become foundational for biology and drug discovery worldwide.

• Recursive self-improvement (RSI) is powerful but risky.

  • RSI works best in verifiable domains like math, coding, and games.
  • The most dangerous version is “closed-loop” RSI without human review.

• AI safety needs stronger engineering and independent oversight.

  • He calls for hardened sandboxes, better monitoring, and cyber defenses that stay ahead of attacks.
  • He supports third-party auditors/evaluators and lab-to-lab red teaming.

• Society is not ready for conscious AI.

  • Hassabis argues intelligence and consciousness are not the same thing.
  • He believes we should build intelligent tools first, then use those tools to better understand consciousness before considering conscious systems.

• AI sovereignty will likely be pragmatic, not fully self-sufficient.

  • Most countries should not try to build frontier models from scratch.
  • Instead, they should mix frontier models, open-source models, and domain-specific applications with privacy and control.

Topics Discussed

AI for Science and Medicine

Hassabis explains that his long-term goal has always been to use AI to advance science, especially biology and medicine. He points to AlphaFold as proof that AI can solve once-intractable scientific problems and says future work should extend into drug discovery, toxicity prediction, and other biochemical properties.

Recursive Self-Improvement and AGI

The discussion turns to RSI, including AlphaZero-style self-play systems. Hassabis notes that these systems can rapidly improve in domains where outputs can be verified. But he warns that moving toward closed-loop, human-free self-improvement could create an “intelligence explosion” that is hard to control and possibly alien in its reasoning.

Safety, Cybersecurity, and Lab Coordination

He emphasizes that the current AI safety challenge is partly an engineering problem:

• stronger cyber defense,
• more robust sandboxing,
• better monitoring,
• and formalized safety testing.

He also supports industry standards bodies with government oversight and international collaboration, along with third-party auditing and pre-release red teaming.

Consciousness and Moral Standing

Hassabis strongly pushes back on the idea that AI is already conscious because it can say it is. He argues:

• there is no agreed scientific definition of consciousness,
• behavior alone is not enough,
• and AI lacks the biological substrate humans and animals share.

He says intelligence and consciousness can be separated, and society should not rush into debates about AI moral standing before it understands the issue better.

AI Sovereignty and Global Strategy

On AI sovereignty, Hassabis says:

• frontier capability is likely to remain concentrated among a handful of companies,
• open source will remain important but comes with risks,
• and governments/enterprises should focus on using AI effectively rather than trying to build frontier models independently.

Notable Insights

• “The whole point of me working on AI for my whole career was to use it as the ultimate tool to advance science and medicine.”
• “The next phase... is closed RSI, closed loop... and that will obviously be faster, but I think is also a very dangerous way to get to full AGI.”
• “I don’t think we are ready yet as a society” to decide whether AI systems should have moral standing or consciousness.
• He argues that intelligence and consciousness are double dissociable: a system can be very smart without being conscious, and conscious beings may not be especially intelligent.

Practical Implications

• For AI labs: prioritize safety engineering, monitoring, audits, and controlled deployment.
• For businesses: focus on where AI creates real value, especially in science, coding, and operations.
• For governments: collaborate internationally and avoid trying to build every frontier model independently.
• For researchers: invest in neuroscience and consciousness research before assuming AI consciousness is solvable or imminent.

Bottom Line

Hassabis presents a cautious but ambitious view of the AI future: use AI aggressively to accelerate science and medicine, but slow down when it comes to autonomous self-improvement and anything resembling conscious systems. His core message is that humanity should first build useful, governable intelligence—and only then decide whether it is ready to cross the next, much more complicated threshold.