Demis Hassabis has spent years arguing that the technology behind most AI chatbots cannot reach human-level intelligence on its own. In 2026 the Nobel Prize-winning scientist stepped down as head of DeepMind, Google’s AI lab, which he co-founded. Day-to-day control of Gemini, the company’s main family of AI models, passed to his successor, Koray Kavukcuoglu, the lab’s chief technology officer. Hassabis became chair of the lab and chief scientist of parent company Alphabet, but neither title puts him in charge of the business that ships Google’s AI products.
The technology he doubts works by predicting text and stringing it together, in systems called language models. He sees them as strong at matching patterns across huge amounts of text but weak at understanding cause and effect in the physical world. His preferred path is AI trained to simulate and predict how that world behaves, which he calls world models. Building them, he has said, has always been the lab’s plan for reaching human-level intelligence.
Language models have proven hard to bet against. Uneven and imperfect, they have still succeeded across a wide range of fields, and rivals OpenAI and Anthropic built their businesses on pushing the approach further. They found paying customers and everyday users at a scale no other AI approach has matched. Google kept shipping its own language models, including Gemini, but measured against what rivals invested, that work had been a secondary bet. By 2026, Gemini had faced repeated delays and fallen behind the most advanced models from OpenAI and Anthropic.
The reshuffle amounts to an acknowledgment that the approach rivals favored has outpaced the one Hassabis championed, though the company did not describe it that way. DeepMind made its name on long-shot scientific research, so handing over the product reins changes what the lab is chiefly known for.
For Hassabis, the new titles make official a change already under way in how he spends his time. He speaks often in public about AI regulation, safety and science, in the role of an AI statesman. That role lets him keep making the case for world models without control over Gemini’s roadmap.
Google still has more money, data and computing power than nearly any rival. Whether those resources can bring Gemini level with OpenAI’s and Anthropic’s best models is a question for a business Hassabis no longer runs.
