Source commit: 36b14ec
Imagine this scenario:
It's 2 a.m. in a biology lab. A postdoc named Lin stares at a dense list of papers on her screen, eyes sore. Her topic is liver fibrosis—a disease in which the liver slowly hardens and may eventually fail. She needs a new therapeutic target, but in the past three months she has read over two hundred papers and run countless experiments with no breakthrough.
She sighs, opens an interface, and types: "Based on the currently known mechanisms of liver fibrosis, propose several new target hypotheses and assess the feasibility of each."
Ten minutes later, the screen shows five detailed hypotheses, each with references, suggested experiment designs, and risk assessments. One of them is something she had never considered, but its logical chain is so perfect it makes her heart race.
This isn't science fiction. This is what Co-Scientist—a multi-agent research assistant announced by Google DeepMind on June 3, 2026—is doing.
---
From AlphaFold to Co-Scientist: DeepMind's Scientific Ambition
Many people know DeepMind only for AlphaGo, the AI that beat the world Go champion. But its greatest achievement may be AlphaFold.
Proteins are the LEGO bricks of life. Every enzyme, antibody, and hormone in your body is a protein, and their function depends on their three-dimensional shape. For decades, determining a protein's structure required months or years of experiments costing millions of dollars.
AlphaFold changed all that. It predicts protein structures with near-experimental accuracy in minutes. When released in 2021, it shocked the biology community. By 2024, it had predicted over 200 million protein structures—covering nearly every known species.
AlphaFold solved the "structure problem." But science is more than structure. Research is a full pipeline: propose a hypothesis → design experiments → analyze data → revise the hypothesis → experiment again → publish. It's a long, expensive, failure-filled process.
Co-Scientist targets the first and hardest step of that pipeline: proposing hypotheses.
---
What Does "Multi-Agent" Mean? Why Isn't One Agent Enough?
To understand Co-Scientist, you must first understand "multi-agent."
Previous AI assistants usually had one "brain" answering your questions: you ask, it thinks, it answers—like a consultant.
But scientific hypothesis generation isn't something a single consultant can handle. It requires multiple "roles" collaborating:
- A "literature expert": reads papers, summarizes known findings, and finds gaps in knowledge
- A "creative generator": boldly proposes new hypotheses based on known facts
- A "critic": nitpicks, evaluating each hypothesis's loopholes, risks, and logical flaws
- An "experiment designer": translates hypotheses into actionable experimental plans
- An "integrator": consolidates everyone's input and gives the final recommendation
- Experimental failure (the vast majority of hypotheses are wrong)
- Insufficient funding (one experiment can cost tens to hundreds of thousands of dollars)
- Lack of time (a good project may take a decade)
- Technical limitations (some experiments can't be done with current technology)
- Disciplinary barriers (cross-domain collaboration is extremely difficult)
That's exactly what Co-Scientist does. It's not one agent working, but a group of agents "holding a lab meeting." They play different roles, debating, correcting, and complementing each other.
It's much like a real research team. A good lab isn't a dictatorship by the PI; students challenge and inspire each other. Co-Scientist simulates that process with AI.
Its foundation is Gemini—Google's large language model. But Gemini here isn't an "answer machine"; it's one of the "participants." It may play both the creative generator and the critic, letting them "debate with themselves."
---
It's Already Collaborating on Real Research
DeepMind says Co-Scientist has participated in three real research projects:
Liver fibrosis. One of the trickiest liver diseases. After injury, the liver over-repairs itself, producing scar tissue (fibrosis) that can progress to cirrhosis. Co-Scientist helped the research team propose new therapeutic target hypotheses.
ALS (amyotrophic lateral sclerosis). The disease that Stephen Hawking had. It gradually kills motor neurons, leading to full paralysis. Its cause remains unknown and there is no cure. Co-Scientist participated in ALS-related hypothesis generation.
Aging research. Aging is an extremely complex biological process involving genes, metabolism, inflammation, cellular senescence, and more. Co-Scientist also provided hypothesis support in this field.
These three projects share a common feature: they are all hard-to-treat diseases with "unclear mechanisms and limited treatments." Precisely in such "we don't know where to go" fields, AI's hypothesis-generation ability is most valuable.
If a problem already has a clear answer, AI isn't needed. But if a field doesn't even know what questions to ask, AI's "imagination" may be the opportunity for human researchers to break through.
---
Why Now?
You might ask: AI proposing scientific hypotheses sounds great, but is it really feasible? Why didn't anyone do it before?
The answer: the technology wasn't ready.
Proposing good scientific hypotheses requires not memorization but "creative reasoning." It demands that the AI do several things:
First, understand a large body of cross-domain knowledge. Liver fibrosis involves cell biology, molecular biology, immunology, and even metabolomics. The AI must integrate these fields and find cross-domain connections.
Second, identify knowledge gaps. Knowing "what is still unknown" is harder than knowing "what is known." This requires a panoramic understanding of the existing literature and the ability to spot contradictions, omissions, and inconsistencies.
Third, generate testable hypotheses. A hypothesis isn't better for being bolder—it must be "falsifiable": you can design experiments to verify whether it's right or wrong. An "unfalsifiable" hypothesis has no scientific value.
Fourth, assess risk and feasibility. A hypothesis may be logically sound but too costly to test, ethically infeasible, or dependent on technology that doesn't yet exist. A good research assistant helps researchers avoid these pitfalls.
Only in 2024-2025 did large language models begin to develop the rudiments of these abilities. The 2026 Co-Scientist is the first time they've been systematically integrated into a "multi-agent research team."
---
Not Replacing Scientists, but Amplifying Them
When many people hear "AI doing science," their first reaction is: oh no, scientists will lose their jobs.
That worry is unnecessary. At least for the next decade, tools like Co-Scientist are positioned as "assistants to the assistant," not "replacements for scientists."
Why? Because the ultimate bottlenecks in research have never been "can't think of a hypothesis." They are:
In fact, it's more like a "super literature assistant" + "brainstorming partner." It lets one scientist read a thousand papers in a day, propose fifty hypotheses, and assess each one's risk—work that previously required a team months.
Throughout scientific history, many breakthroughs came from "cross-domain connections." The discovery of DNA's double helix happened because physicist Francis Crick walked into a biology lab. Penicillin was discovered because Fleming noticed an accidental petri dish contamination. AI's value is that it can discover these "cross-domain connections" faster and more systematically than humans.
---
A Deeper Meaning
The release of Co-Scientist marks the maturation of a trend: AI is shifting from "tool" to "collaborator."
Early AI was a calculator—you input numbers, it outputs results. Later AI was a search engine—you input a question, it outputs relevant documents. Today's AI is a chat assistant—you input a need, it outputs suggestions.
Co-Scientist represents the next stage: AI as a "partner." It doesn't just answer your questions; it thinks through problems with you. It has multiple "personas" (literature expert, creative generator, critic) that discuss internally and then present you the results. You're no longer the "questioner"; you're the "team leader," and AI is your "research team."
The significance of this shift goes far beyond a research tool. It hints at the future working relationship between AI and humans: not "humans commanding machines," but "humans and machines collaborating."
---
Epilogue
Back to the 2 a.m. lab.
Following Co-Scientist's suggestion, Lin designs a new experimental plan. Three months later, she has preliminary data. A year later, her paper is published in a top journal. Five years later, a new drug based on her discovered target enters clinical trials.
The starting point of it all: one of five hypotheses proposed by an AI in ten minutes.
This story sounds like a fairy tale, but DeepMind is turning it into reality. Co-Scientist isn't perfect—perhaps 90% of its hypotheses are wrong. But science has always been a process of "finding one success among countless failures."
AI's value isn't that it's always right. It's that when human researchers "don't know where to go," it can offer directions—even if only one of them turns out to be right.
In science, sometimes one correct direction is worth more than a hundred wrong answers.
DeepMind's Co-Scientist is learning to be the one who "points the way."
Or rather, it's learning to be a qualified research assistant—one that has read every paper, never tires, and is always willing to propose one more hypothesis.
Perhaps that is AI's gentlest and most powerful form in science: not replacing human wisdom, but amplifying human curiosity.