Google Team: AI Scientists Need Budgets and Rules, Not Just Smarts

AI in science is changing fast. Instead of using one AI program to do one small job, researchers are now building teams of AI 'agents' (programs that can take actions, not just chat) that can carry a research project from start to finish: suggesting an idea, designing an experiment, running it, and reading the results. That is exciting, because a research project that once took a team months might eventually take days.

The paper's main point is that being smart is not enough. Real science is expensive. Testing a new drug or material eats up lab time, rare chemicals, expensive equipment, and huge amounts of computing power. An AI that never gets tired and never asks permission could burn through a lab's entire yearly budget in a week chasing dead ends. So the authors argue AI researchers need something like a marketplace: a system that hands out limited resources, sets priorities, and decides which ideas get tested first.

Then come the rules. If an AI team discovers a promising new drug, who gets the patent? If an AI-designed experiment is dangerous, or the data leaks, who is legally responsible? The paper lays out four things a working system would need: a way for AI and human scientists to agree on research goals, a way to assign credit fairly, a way to track who is accountable, and protections against misuse and data theft.

Finally, the authors look at the bigger picture. If AI starts producing discoveries at high speed, will everyone benefit, or only the companies and countries that own the biggest AI systems? They argue governments should start writing policies now, while the technology is still forming. It's worth being clear about the limits here: this is a concept paper from a group of researchers, with no experiments, no product, and no timeline. It is a map of problems, not a solution.