Google Is Quietly Changing What AI Driven Science Looks Like
Google DeepMind CEO Demis Hassabis stepped onto the Google I/O stage this week and told the crowd they were “standing in the foothills of the singularity.” It was a grand claim, but what made it remarkable was the moment he chose to say it right in the middle of a segment about scientific AI.
The centerpiece of that segment was WeatherNext, Google’s weather prediction software that reportedly issued an advance warning before Hurricane Melissa made a catastrophic landfall in Jamaica last year, potentially saving lives. It is a real, meaningful achievement. But it is also a purpose-built tool solving a specific problem, and that kind of AI driven science may no longer be where Google’s biggest ambitions lie.
For years, the company’s flagship contribution to AI driven science was AlphaFold, the protein-structure prediction system that earned DeepMind scientists, including John Jumper, a Nobel Prize. But according to a Los Angeles Times report, Jumper has now been reassigned to work on AI coding, not scientific tools. Google has faced pressure after its coding products fell behind those from Anthropic and OpenAI, but the move also signals something broader a pivot toward agentic, general-purpose AI systems that could one day conduct research on their own.
That vision is taking shape inside Google’s new Gemini for Science package, announced at I/O. The bundle brings together LLM-based tools including the hypothesis-generating AI Co-Scientist and the algorithm-optimizing AlphaEvolve under one brand. Neither has been made broadly available yet, but Google is now accepting researcher applications for access. Early testers have been enthusiastic, Stanford geneticist Gary Peltz reportedly compared using the AI Co-Scientist to consulting the oracle of Delphi, in a piece published in Nature Medicine.
Google is not the only company pushing AI deeper into research territory. OpenAI this week announced that one of its general-purpose reasoning models had disproved an important mathematics conjecture with no specialised scientific training. If general models can make those kinds of contributions, the pressure on domain-specific tools will only grow.
Pushmeet Kohli, Google Cloud’s chief scientist, captured the shift directly in the journal Daedalus this week, writing that AI is moving from facilitating science to actually doing it. Hassabis echoed a more cautious version of that view in the same publication, saying AI should be thought of as “an amazing tool to help scientists” for the next decade, and perhaps a collaborator beyond that.
Google has been deliberate about its language. The name AI Co-Scientist rather than AI Scientist appears to be an intentional choice, keeping humans visibly at the centre. Still, the company’s direction in AI driven science is becoming clearer: away from building the next AlphaFold, and toward building systems that might one day not need a scientist to guide them at all.
