Recent advances in artificial intelligence (AI) are opening up new perspectives in many areas of research. Among these initiatives, the “AI co-scientist” project, developed by Google Research, stands out. It proposes an AI system designed to help researchers generate novel research hypotheses, debate and refine these ideas, and then verify them more effectively than with traditional approaches.
An assistant inspired by scientific reasoning
The Google Research team built the AI co-scientist as a set of specialized agents, each playing a precise role in the scientific process:
- A generation agent who explores the literature and formulates new hypotheses,
- A reflection agent to assess the plausibility, originality and robustness of these hypotheses,
- A ranking agent, responsible for comparing these ideas via a “tournament” system,
- An evolution agent that continuously improves or merges hypotheses,
- And a meta-review agent that identifies limits or recurring points in the debates to optimize the whole process.
The whole process is orchestrated by a “Supervisor”, who delegates the right amount of calculation and time to the various agents. The guiding principle: the more the system “thinks” and confronts its hypotheses, the more it refines its proposals.
Modular AI that interacts with researchers
To validate the AI co-scientist, the team conducted several experiments in the biomedical field. In each case, a human expert remained at the heart of the process to guide the AI and assess the relevance of hypotheses before testing them. The findings of this collaborative work underline the potential of AI capable of bothinnovating and self-correcting when its hypotheses prove flawed.
Although validated in complex biomedical contexts, AI co-scientist is designed to adapt to a variety of scientific disciplines. By modulating its agents (and incorporating new specific tools, such as a bibliographic search engine or dedicated prediction models), it could just as easily prove useful in materials physics, agronomy or engineering.
According to the research team, the AI co-scientist is not intended to replace researchers, but rather to accelerate and enrich their work. It acts as a “partner”, capable of rapidly exploring multiple avenues while enabling humans to focus their skills on finer, more strategic tasks.
Collaboration between humans and AI takes on a new dimension thanks to the AI co-scientist. Its early successes in the biomedical field demonstrate that it is possible to generate and validate scientific hypotheses more rapidly, while benefiting from a framework of continuous iteration and critical review. Ultimately, this type of AI could revolutionize fundamental research and accelerate the development of concrete solutions in fields as varied as health, energy and the environment.
At Humind, a comparable but more pragmatic approach has been adopted, not for fundamental research, but for the concrete application of proven and documented technologies, with the aim of stimulating business innovation. The process is based on a targeted search for information, combined with a multi-agent AI system responsible for analyzing the results, comparing them with the existing product portfolio and proposing different development options. Each option is then thoroughly evaluated, facilitating the prioritization of resources according to various strategic criteria. This approach not only makes it possible to efficiently exploit multiple sources of information and optimize the R&D project portfolio, but also guarantees data confidentiality.
To find out more, take part in our webinar dedicated to innovation support
Learn more about Google co-scientist :
- Article from Google Research: Accelerating scientific breakthroughs with an AI co-scientist
- Technical report: “Towards an AI co-scientist” by the Google Cloud AI Research & Google DeepMind team, 2025
