Feb 20, 2025 11:55:00

Google launches AI co-scientist, an AI assistant for scientists

While the widespread use of AI in the fields of science and academic research has raised

concerns about the falsification of data and papers, AI has also brought many benefits, such as improving the productivity of talented scientists and accelerating the drug discovery process. On February 19, 2025, Google announced 'AI co-scientist,' a science-focused AI assistant based on Gemini 2.0.

Accelerating scientific breakthroughs with an AI co-scientist
https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/

Google's new AI generates hypotheses for researchers - Ars Technica
https://arstechnica.com/google/2025/02/googles-new-ai-generates-hypotheses-for-researchers/

Google's 'AI co-scientist' is a multi-agent AI system that aims to act as a collaboration tool for scientists. It goes beyond traditional uses such as literature searching and summarization to discover new knowledge tailored to specific research fields based on existing evidence, and present suggestions and hypotheses that can serve as starting points for research.

The AI co-scientist contains multiple models that process input data and access resources on the internet to improve their output, and various agents compete against each other in a 'self-improvement loop,' similar to new inference models such as Gemini Flash Thinking and OpenAI o3 .

Since it is still a generative AI, it cannot generate new ideas from scratch, but it can assist scientists with suggestions based on inference using existing data, and it also has a chat interface that allows it to converse with scientists. In other words, AI co-scientist is a brainstorming tool specialized for science.

As a multi-agent system, AI co-scientist self-improves using the Elo auto-evaluation metric, in which agents compete over the

rate of correct answers. Google has confirmed that the higher the Elo rating, the higher the accuracy rate in benchmarks that solve difficult problems.

The graph below shows this, with the accuracy (vertical axis) of the answers given by AI co-scientis (blue) and Gemini 2.0 (red) to the questions in the benchmark 'GPQA Diamond,' which poses doctoral-level scientific problems, summarized by the Elo rating (horizontal axis).

In fact, when experts were asked to evaluate the usability of AI in 11 studies in terms of novelty and impact, AI co-scientist outperformed other cutting-edge models in both categories and its output was also rated favorably.

Regarding the future of AI co-scientist, Google said, 'We are excited by the early results of AI co-scientist and believe it is important to evaluate its strengths and limitations in science and biomedicine more broadly,' and called on research institutions around the world to participate in the test program .