Nature. ¡°Accelerating scientific discovery with Co-Scientist.¡± Published May 19, 2026.
Nature. ¡°Teams of AI agents boost speed of research.¡± Published May 19, 2026.
AI Has Started Asking Scientists Questions
- The rise of AI that goes beyond writing papers and proposes hypotheses to test
AI no longer remains merely an assistant that summarizes academic papers. The Co-Scientist study published in Nature by researchers affiliated with Google DeepMind and Google Research shows that AI has entered a stage where it can generate scientific hypotheses, critique them against one another, and revise them. The future of science does not lie in AI finding answers on behalf of humans, but in helping humans reach better questions faster.
[Key Message]
* AI has moved beyond being a tool that finds answers and has begun proposing questions that scientists can test. The important shift is that AI is no longer limited to summarizing or writing papers; it is starting to participate in hypothesis generation, the starting point of research.
* The core of Co-Scientist is not that one AI produces an answer, but that multiple AIs generate, challenge, and revise ideas together. This structure mirrors the scientific thinking process of a real lab meeting, where hypotheses are proposed, weaknesses are identified, and ideas are refined.
* In drug development and biomedical research, AI co-scientists can greatly accelerate early-stage exploration. They can quickly connect complex clues related to drug repurposing, therapeutic targets, and antibiotic resistance, then suggest candidates that may be tested experimentally.
* AI does not eliminate the work of scientists; it makes human judgment more important. As AI produces more hypotheses, humans must become more rigorous in deciding what to trust, what to doubt, and what to verify.
* The faster discovery becomes, the heavier the responsibility for verification becomes. An AI-generated hypothesis is not a conclusion but a starting point, and experimental testing, reproducibility, and ethical judgment remain the responsibility of humans and the scientific community.
***
AI Once Looked for Answers. Now It Creates Questions
The hardest part of science is not getting the right answer. The truly difficult part is figuring out what to ask. Why do cancer cells respond only to certain drugs? Why do some bacteria survive even in the face of antibiotics? Among drugs that have already failed, are there candidates that could be used again for entirely different diseases? Science always begins with questions like these. Experiments are the process of testing those questions in reality, and papers are the records that organize the results.
Until now, AI has mainly been used as a tool for reading and organizing those records. It summarized vast numbers of papers, found literature that researchers had missed, classified data, created graphs, and wrote code. These forms of AI were already impressive enough, but they were still closer to assistants standing behind the scientist. They were helpers that quickly scanned knowledge that had already been produced, not colleagues that helped create the starting point of new research.
This time, however, the direction is a little different. The Co-Scientist study published in Nature shows that AI can go beyond simply organizing materials and directly propose hypotheses that scientists may be able to test. The researchers designed Co-Scientist as a Gemini-based multi-agent system. Put simply, rather than having one AI produce an answer on its own, the system consists of multiple AIs, each taking on a role like a researcher: generating ideas, challenging one another, and leaving behind stronger hypotheses. The Nature paper explains that this system generated hypotheses and proposed testable candidates in biomedical tasks such as drug repurposing, discovering new therapeutic targets, and explaining mechanisms of antibiotic resistance.
The importance of this shift is clear. A paper is closer to the end of research, but a hypothesis is closer to its beginning. The statement that AI can write a paper for us is far less significant than the statement that AI has begun to say to scientists, ¡°What if we tried testing this question?¡± What changes the speed of science is not the wording of the final sentence, but the direction of the first question.
A New Assistant That Finds a Path Through Piles of Papers
Modern science is not slow because there is too little knowledge. If anything, it is slow because there is too much knowledge. New papers are published every day, while data on genes, proteins, diseases, drugs, and clinical outcomes pile up endlessly. It is difficult enough to keep up with just one field, yet truly important discoveries often emerge at the boundaries between several fields. One clue may be in a cancer research paper, another in immunology data, and another hidden inside the mechanism of action of an already approved drug.
Human researchers connect these clues intuitively. But human time and concentration have limits. No matter how outstanding a researcher may be, no one can read every paper. No one can compare every database at the same time. That is why much of science is, in reality, spent on the act of searching: finding what is already known, finding what research is missing, and finding what relationships have not yet been tested.
Systems like Co-Scientist show their strength precisely at this point. AI can scan literature and data across a much broader range than humans can, and it can do so quickly. It can also suggest possible connections among scattered clues. For example, one study may show that a certain protein pathway is overactive in a particular disease, while another paper may contain evidence that an already approved drug can affect that pathway. Another dataset may show that the same pathway is especially prominent in a certain patient group. If these pieces remain separate, they are merely fragments of information. Once connected, they become a research hypothesis.
Of course, the fact that AI has connected them does not mean it has made a discovery. This point is extremely important. What AI proposes is not a treatment, but a treatment candidate; not a conclusion, but a question to be tested. In science, a hypothesis is only the starting point. Only after passing through cell experiments, animal experiments, clinical trials, and reproducibility checks can it become meaningful knowledge. But if the starting point changes, the research timetable can change as well. If more candidates can be reviewed more quickly, researchers can abandon paths that are likely to fail earlier and direct more resources toward paths with higher potential.
In this sense, an AI co-scientist is not a magician. It is more like an extremely diligent and fast research assistant. It reads papers through the night, gathers relevant clues, organizes possible hypotheses, and brings several pages of notes to the morning meeting. Among those notes, there may be absurd ideas, and there may be proposals that do not yet have enough evidence. But one of them may make a researcher stop and think. ¡°This might really be worth testing.¡± That moment matters.
Not One Brilliant AI, but AIs That Challenge One Another
The distinctive feature of Co-Scientist is that it does not rely on one AI producing a polished answer all at once. The system operates through multiple AI agents with different roles. One agent generates new hypotheses, another looks for weaknesses in those hypotheses, and another strengthens the evidence or revises the idea into a better form. It resembles a research lab meeting: one person presents an idea, another says, ¡°The evidence for that is weak,¡± and someone else adds, ¡°Then perhaps it becomes plausible if we add this condition.¡±
This structure resembles the way science actually works. Good research is not created by a single flash of inspiration. The first idea that comes to mind is usually rough. It becomes stronger little by little as it faces questions from colleagues, survives objections, is examined for experimental feasibility, and is checked against existing research. Inspiration matters in science, but the process of doubting and refining that inspiration matters even more.
The same is true of AI. The first answer AI gives may sound plausible, but plausibility is not enough. In science especially, the smoothness of a sentence guarantees nothing. Something may look logical and still turn out to be wrong in an experiment. Something may seem to have evidence behind it and still miss an essential condition. That is why the core of Co-Scientist lies not only in the fact that AI generates hypotheses, but also in the fact that AI critiques and revises the hypotheses it has generated.
This is different from simple automated writing. Automated writing is closer to filling in blanks. Hypothesis generation and criticism, however, create a competition among possibilities. They involve lining up multiple ideas and asking which one has stronger evidence, which one is more experimentally feasible, and which one better targets a gap in existing research. Nature introduced these AI research-assistance systems as tools that could accelerate research by generating hypotheses, interpreting data, and proposing directions for drug development.
The important point here is not that AI thinks like humans. Nor does it mean that AI has acquired the intuition and responsibility of human researchers. Rather, it means that some parts of scientific thinking, such as generating candidates, comparing them, and critiquing them, can now be repeated by machines at much higher speed. That alone can change the research site. A list of candidate ideas that once required a research team to discuss for weeks may now be produced by AI in a much shorter time.
The First Step of Drug Development Is Changing
One reason Co-Scientist is drawing particular attention is that it first demonstrated its potential in the biomedical field. Drug development is one of the most time-consuming and expensive areas of science and technology. Researchers must understand the cause of a disease, find a therapeutic target, select candidate compounds, test toxicity and efficacy, and pass clinical trials. If the wrong candidate is chosen at the early stage of this long process, enormous amounts of money and time disappear.
That is why what matters in drug development is not ¡°finding the correct answer quickly,¡± but ¡°selecting better candidates worth trying.¡± An AI co-scientist may be especially useful in this early stage. It can help identify the possibility that an existing drug might be used for an entirely different disease, propose new biological targets for attacking a specific disease, or explain hidden mechanisms behind antibiotic resistance.
According to the Nature paper, Co-Scientist proposed new uses for existing drugs and combination-treatment candidates related to acute myeloid leukemia, and some of those proposals were validated through in vitro experiments. This point is important. It means AI did not merely exercise imagination, but produced candidates that could be checked through actual experiments. Of course, this does not mean that these immediately became treatments for patients. But as a starting point for research, it represents a meaningful change.
For a general audience, this change can be explained as follows. If earlier AI was a librarian, Co-Scientist is closer to a research assistant who joins the meeting. A librarian finds the materials you need. A research assistant reads those materials and says, ¡°If we look at this paper and that paper together, there may be a possibility of testing this drug again for this disease.¡± ¡°If this pathway is the actual cause, we could also consider blocking this protein.¡± ¡°However, the evidence for this hypothesis is weak, so we should first confirm it with this kind of experiment.¡±
This shift could affect not only pharmaceutical companies but also university labs, hospital research teams, and biotech startups. Large companies already have massive research workforces and data infrastructure. Smaller research teams, however, may have good ideas but lack the time to search the literature broadly and narrow candidates systematically. If AI co-scientists are used properly, smaller teams may also be able to look at a much wider map of knowledge. This could help widen the starting line of scientific research.
At the same time, the gap could also grow. The speed difference between teams that use AI well and those that do not may widen. Institutions with high-quality data and experimental infrastructure can quickly test AI-generated hypotheses, while institutions without such capacity may struggle to turn even good hypotheses into experiments. In the end, future research competitiveness will not be determined by a single AI tool. It will also require experimental capacity to test AI-generated hypotheses, a culture of organizing data, and a system that preserves even failed experiments as assets for learning.
Scientists¡¯ Work Will Not Disappear. It Will Become More Difficult
The claim that AI can generate scientific hypotheses is easily exaggerated. It sounds as though AI will soon replace scientists. But the actual change is more complex, because what AI does well and what humans must still do are different.
AI is good at broad searching. It can quickly read many papers, compare many possibilities at once, and suggest connections that humans may easily miss. It is also good at repeatedly generating and evaluating candidates. It does not get tired and can continue trying different combinations. In this respect, AI expands the range of exploration available to researchers.
But AI does not take responsibility for why a question matters. It remains the task of humans to decide which research is urgent for patients, which experiments are ethically acceptable, and which results could lead to socially dangerous interpretations. Because AI depends heavily on existing literature and data, it can also bring along the biases embedded in them. It is strong in areas with many studies, English-language papers, and abundant data, but weaker in areas such as rare diseases or health issues in low-income countries, where data may be scarce.
That is why, in the age of AI co-scientists, the role of human scientists does not shrink. It changes. In the past, much time was spent finding and organizing materials. In the future, choosing which question to pursue will become more important. The more candidates AI produces, the more strictly researchers must select among them. The faster AI generates hypotheses, the more calmly researchers must doubt them. The wider the map of knowledge AI lays out, the more carefully researchers must judge which path truly matters.
The value of scientists may move further from the quantity of knowledge toward the quality of judgment. People who know a great deal will still matter, but in the future, people who know what not to trust may matter even more. The ability to see gaps in AI¡¯s seemingly well-supported sentences, to judge whether AI-recommended experiments are actually feasible, and to notice ethical problems that AI has missed will become core scientific capabilities.
This will also affect science education. Future researchers will need more than training in how to use AI. They will need training in how to evaluate AI-generated hypotheses. It will become more important to check whether the literature evidence is sufficient, whether an experimental design is biased, whether the data are skewed toward a particular group, and whether the results are reproducible. More important than the skill of writing good prompts will be the attitude of asking good skeptical questions.
In an Age of Faster Discovery, Slower Verification Becomes More Important
The greatest change an AI co-scientist may bring is speed. The speed of generating hypotheses, narrowing candidates, and connecting literature may all increase. In fields such as cancer, dementia, infectious diseases, climate technology, and new materials, where time means both cost and life, this speed has great significance. Finding possibilities sooner means experiments can begin sooner, and failures can also be identified sooner.
But science is originally a system of slow verification. A hypothesis must pass through multiple experiments, be reproduced by other researchers, and remain stable over time. If AI rapidly produces vast numbers of hypotheses, science may become faster, but it may also become more crowded. Distinguishing good hypotheses from merely plausible ones may become more difficult.
The particular danger is when an ¡°abundance of hypotheses¡± hides a ¡°shortage of verification.¡± Just because AI generates hundreds of research ideas every day does not mean science moves forward by hundreds of steps every day. Time in the laboratory, research funding, equipment, and personnel remain limited. Ultimately, what matters is not the ability to produce many hypotheses, but the ability to choose hypotheses worth testing.
There is also the problem of research concentration. Funding may flow toward fields where AI suggests high potential, diseases with abundant data, and technologies with strong commercial value. If that happens, already prominent fields may advance even faster, while areas that lack data but are socially important may fall further behind. AI could help democratize science, but it could also intensify inequalities in science.
The question of responsibility also remains. If human researchers test a hypothesis proposed by AI and achieve a breakthrough, who contributed the idea? Was it the company that created the AI system, the scientist who provided the research goal, or the research team that carried out the experiment? Conversely, if AI suggests a flawed hypothesis and causes time and research funding to be wasted, or if it encourages a dangerous experimental direction, where does responsibility lie? As AI becomes a colleague in science, the institutions of science must change as well.
The fact that major technology companies such as Google DeepMind provide core tools for scientific discovery also requires separate discussion. Science has long built knowledge around universities, public research institutions, and academic communities. But if key hypothesis-generation tools and research-automation platforms come to operate on AI models owned by a small number of large companies, the structure of scientific knowledge production may also change. Who can access those tools, what data were used for training, and which research directions are recommended first are all questions connected to the fairness of science.
That is why the central question in the age of AI co-scientists is not performance alone. How fast the system is matters, but who uses it matters more. How intelligent it is matters, but how it is verified matters more. As the speed of science increases, the mechanisms of trust that support science must become slower and stronger.
Science Is Moving from the Solitude of Genius to Collaboration with Hybrid Intelligence
The larger message of this change is that the image of the scientist is changing. For a long time, the public has imagined scientists as solitary geniuses. One person remains in the laboratory late into the night, gains a decisive insight, and reaches a discovery that changes the world. Of course, actual science has always been the product of collaboration, but within the myth of scientific discovery, there still stands a figure who shines alone.
In the age of AI co-scientists, this image may change more quickly. The researcher of the future may be less someone who thinks of everything alone and more someone who directs humans, AI, automated experimental equipment, massive databases, and simulation systems together. The laboratory will no longer be a space occupied only by human researchers. It will become a space of hybrid intelligence where many intelligent tools move together.
This change does not reduce the authority of scientists. It makes their work more strategic. Researchers may no longer need to think up every candidate by themselves. Instead, they must decide what problem to solve, which candidates to test, and what to learn from each failure. The scientist becomes not a warehouse of knowledge, but a conductor of discovery.
Corporate research and development may also change. Pharmaceutical companies, battery makers, semiconductor materials firms, and energy-technology companies are all engaged in the struggle to find ¡°possible combinations.¡± They must identify which materials are more stable, which catalysts are more efficient, which therapeutic targets are safer, and which processes are more economical. AI co-scientists could greatly accelerate this early exploration.
This trend also sends an important signal to South Korea. South Korea invests heavily in research and development in semiconductors, batteries, biotechnology, medical AI, and advanced materials. Yet in the field, much time is still spent reviewing papers, analyzing patents, searching for candidate materials, and designing experiments. If AI co-scientists are used properly, they could strengthen the exploratory capacity of entire research teams rather than replace individual researchers. The opportunity may be especially significant for university labs and small to mid-sized biotech companies that lack sufficient personnel and time.
However, innovation does not happen simply because a tool is introduced. Equipment and personnel capable of testing AI-generated hypotheses, a culture of organizing and sharing data, the habit of recording failed experiments, and research ethics that do not blindly trust AI results must all be present together. An AI co-scientist is not a magician that creates discovery alone. It is a catalyst that becomes more powerful in a prepared research ecosystem.
Humans Must Answer the Questions AI Raises
The Co-Scientist study is fascinating because it forces us to rethink what AI means. We too often understand AI only as a machine that generates answers. We see it as a tool that produces an answer when a question is entered and creates an output when a command is given. But in science, the truly important AI may not be AI that gives answers quickly, but AI that helps create better questions.
Science is not the work of neatly organizing what is already known. It is the work of moving toward what is not yet known. In that journey, a hypothesis is a compass. A good hypothesis changes the direction of experiments, changes how research funding is used, and shakes the framework through which we understand disease. Co-Scientist shows that AI can enter the process of making that compass.
But the final responsibility still remains with humans. The more possibilities AI presents, the better choices humans must make. The faster AI generates hypotheses, the more rigorously humans must verify them. The wider the literature AI scans, the more deeply humans must judge its meaning.
The scientist of the future may not be someone who knows everything alone. Instead, that scientist may be someone who chooses the most meaningful path through a forest of possibilities created together by humans and machines. The protagonist of discovery is neither a single genius nor a single algorithm. New knowledge is born when human questions, AI exploration, and experimental verification work together.
A more important change than the age of AI writing papers for us has begun. AI has started asking questions from the seat beside the scientist: ¡°Might this question be worth testing?¡± The future of science will not be determined by admiring that suggestion. It will be determined by the human ability to doubt it, verify it, and ask again.
Reference
Nature. ¡°Accelerating Scientific Discovery with Co-Scientist.¡± Published May 19, 2026.
Nature. ¡°Teams of AI Agents Boost Speed of Research.¡± Published May 19, 2026.