Google DeepMind Unveils Co-Scientist — an AI for Generating Scientific Hypotheses Based on Self-Play
A multi-agent system built on Gemini autonomously proposed and validated several biomedical hypotheses: it found an IRE1 inhibitor for treating leukemia and rediscovered a mechanism of bacterial resistance gene transfer in just two days of computation.
The article in Nature from May 19, 2026, is not just a demonstration of Gemini's capabilities. It is a key element of a broader architectural battle for control over how scientific knowledge is generated. While the public marvels at new drug candidates for leukemia, a tectonic shift is unfolding in laboratories: multi-agent systems based on language models are transitioning from passive information retrievers to arbiters of scientific truth. Google DeepMind's Co-Scientist is not a tool for scientists; it is a bid to create a new operating system for the entire research process, capable not only of generating hypotheses but also ranking, critiquing, and evolutionarily improving them.
The Core: From Text Generator to Simulator of the Scientific Method
The system is a coalition of seven specialized AI agents divided into three phases: generation, debate, and evolution. The architecture fundamentally differs from conventional chatbots. Agents do not simply output text on demand—they engage in asynchronous interaction where a "generator" proposes ideas, a "reflector" acts as a virtual reviewer, and a "ranker" conducts tournaments of ideas using a principle borrowed from AlphaGo. Instead of playing Go, the system plays scientific debates, optimizing bets not on winning the board but on the novelty and verifiability of the hypothesis.
The key innovation is scaling test-time computation. The system allocates the lion's share of resources not to generation but to hypothesis verification, cross-referencing each claim against scientific literature and databases like ChEMBL and UniProt, and in some cases even invoking AlphaFold for structural validation. Co-Scientist does not just "invent"—it verifies, and this is its fundamental difference from hallucinating predecessors.
Timeline and Context: The Race of Scientific Agents
The announcement of Co-Scientist coincided with the publication of a competing system, Robin, from FutureHouse. Nature deliberately synchronized the release of both papers, signaling to the market the formation of a new product class. Behind this lies months of preparation: early testing of Co-Scientist was conducted in a closed mode with Imperial College London, where the system, under the guidance of Professor José Penadés, rediscovered a mechanism of horizontal gene transfer in bacteria that Penadés' team had experimentally discovered but not yet published. Legend has it that the scientist, upon seeing Co-Scientist's result, asked if the AI had peeked at his drafts—so accurately did the system replicate the logic of human discovery in just two days of computation.
Who Wins and Who Loses
The primary winner is Google Cloud. Rolling out Co-Scientist to the masses via the Google Labs platform is a classic market capture strategy through infrastructure dependency. Researchers who get used to running "hypothesis tournaments" on Google's TPUs are unlikely to switch to another cloud.
Large pharmaceutical companies with existing contracts with DeepMind also win. As demonstrated by Cohere's acquisition of startup Reliant AI to verticalize AI in pharma, the industry is willing to pay huge sums for systems that accelerate R&D.
Postdocs and junior researchers, whose work was 70% literature search and hypothesis generation, lose out. Their function is now automated at zero marginal cost. Companies like OpenAI lose as well, as they have yet to present a similar multi-agent architecture for science, limiting themselves to general reasoning with o1.
What the Media Isn't Saying
The first blind spot is the reproducibility problem of AI-generated hypotheses. If Co-Scientist uses another agent based on the same Gemini model as a "reviewer," the critique becomes a closed loop: a model checks a model by rules written by a model.
The second non-obvious point is the strategic silence around the "rediscovery" of the resistance mechanism. A system trained on a corpus of scientific literature inevitably absorbs statistical patterns, even if a specific article was under embargo. The system's success may be explained not by creativity but by Gemini's ability to pick up weak signals from related publications.
Forecast: The Next 30 Days and 90 Days
In the next 30 days, I expect a wave of registrations on labs.google/science and the first independent attempts to replicate the reported results. Within 90 days, the scientific community will split into those who accept Co-Scientist as a co-author (and submit grants written by AI) and those who demand a complete ban on using such systems in peer review due to the risk of contaminating the intellectual environment with synthetic hypotheses. The market for scientific AI agents will reach $2 billion by the end of the year—and Google intends to capture a significant share.
— Editorial Team
