AI Systems Are Running Their Own Scientific Experiments Now
The 2024 Nobel Prize in Chemistry went to Google DeepMind's Demis Hassabis and John Jumper for AlphaFold, a system that predicts 3D protein structures. That was a milestone. What came after looks like a race.
Everyone Wants an AI Scientist
In February 2025, Google released an AI co-scientist tool. By October 2025, OpenAI had launched a dedicated AI for science team and announced GPT-Rosalind, described as the first in a planned series of specialized scientific models. Anthropic announced Claude features aimed at biological sciences the same month.
Three major labs, same quarter, same direction.
How These Systems Actually Work
Google's co-scientist uses multiple specialized agents: a supervisor agent, a generation agent, and a ranking agent. The system generates hypotheses and research plans through their interaction, not through retrieval.
Stanford's AI for Science Lab, led by James Zou, took a similar approach. They built a virtual lab of specialist agents that designed antibody fragments capable of binding to SARS-CoV-2. The agents were generating candidates, not looking up known ones.
OpenAI and Ginkgo Closed the Loop
In February 2026, OpenAI connected GPT-5 directly to automated biological laboratories built by Ginkgo Bioworks. The system ran iterative experiments and reduced the cost of synthesizing a particular protein by 40%.
That is the meaningful step. Not AI suggesting experiments. AI running them, seeing results, adjusting, running them again.
The Efficiency Problem Nobody Is Talking About
A Nature study found a wrinkle in the optimism. AI adoption in science may narrow the range of topics the scientific community investigates. The reason: AI systems favor established areas with large-scale training data.
Optimizing for tractability is not the same as optimizing for discovery. The two can look identical until they aren't.
Where This Lands
The pipeline from hypothesis to lab result is compressing fast. A 40% cost reduction on protein synthesis is not a demo. It is an economic argument for replacing more of the process.
Whether AI narrows science while it accelerates it is an open question. The Nature finding suggests those two things might not be separable.
Source: Technologyreview