![\"\"](\"https:\/\/www.newsbeep.com\/il\/wp-content\/uploads\/2026\/09\/d41586-026-02931-5_52851272.jpg\"\/)
<\/p>\nWhy AI cannot do good science without humans<\/p>\n
<\/a><\/p>\n
The process can involve vast quantities of computational power because agents pursue distinct lines of reasoning before converging on workable solutions. And that takes time. Long enough, Pertl jokes, for her to complete a long-distance triathlon, which she has done eight times since she started her PhD at the Whitehead Institute for Biomedical Research in Cambridge, Massachusetts.<\/p>\n
On a rainy Tuesday in July, Pertl put the AI system to work on one of cancer\u2019s most complicated problems. She asked Co-Scientist for non-obvious but practical ways of harnessing the biology of molecular droplets known as condensates<\/a> to shut down MYC, a protein that runs amok in most cancers and has long defied attack attempts<\/a>.<\/p>\n The request built on years of research by Pertl\u2019s supervisor, Whitehead biologist Richard Young. In 2018, Young and his colleagues found that cells switch on crucial genes by gathering regulatory proteins into condensates that cluster at important genome-control regions called super-enhancers1<\/a>. They also worked out how cancer cells hijack these super-enhancers to send the gene encoding MYC into overdrive, a strong hint that condensates help to fuel the high expression rate of the protein2<\/a>.<\/p>\n To Young, the obvious next step is to target the tumour cells\u2019 condensates directly. Dewpoint Therapeutics in Boston, Massachusetts, a company that he co-founded, is now exploring that option. However, because previous attempts to subdue MYC failed<\/a>, Pertl wanted to see whether AI could come up with a fresh line of attack.<\/p>\n Getting to an idea worth pursuing took some work. Pertl and Whitehead bioengineer Kalon Overholt spent close to an hour going back and forth with Co-Scientist before it understood the question. At first, the AI model assumed that the protein clusters at the centre of the research were meant to be the drug rather than the target. The tool also assumed that the clusters always activated gene expression, when in fact some do the opposite.<\/p>\n Once Pertl and Overholt had corrected these misconceptions, they set the system loose. By the next day, Co-Scientist had combed through more than 700 scientific papers and generated 108 possible strategies. It rejected all but one of the approaches as unworkable. The last strategy turned the laboratory\u2019s thinking on its head.<\/p>\n Rather than dissolving the clusters \u2014 the method that Dewpoint is pursuing \u2014 the AI tool proposed gluing them together. Using a molecular linking technology called click chemistry<\/a>, the approach smushes proteins that either activate or repress the MYC gene into one gooey mass, setting off a chain reaction until MYC\u2019s DNA can no longer be read.<\/p>\n Scientists already knew that DNA and the proteins surrounding it could turn from a loose liquid into something firmer, stiffening like gelatin in a refrigerator. But exploiting that transition to shut down a cancer gene was uncharted territory. \u201cIt\u2019s extremely conceptually compelling,\u201d says Overholt. \u201cWe had certainly never thought about anything like this.\u201d<\/p>\n Trust, but verify<\/p>\n This kind of AI-assisted brainstorming is becoming increasingly common. Researchers have already used Co-Scientist, developed by Google in Mountain View, California, to find a drug combination that kills leukaemia cells in a dish3<\/a> and identify a treatment that, in the lab, regenerates liver tissue damaged by disease4<\/a>.<\/p>\n And Co-Scientist isn\u2019t the only game in town<\/a>. Frontier AI labs, including Anthropic and OpenAI, and start-ups such as FutureHouse5<\/a> and Phylo6<\/a> (all four based in or around San Francisco, California) are rolling out systems that can tackle tasks once reserved for human scientists, freeing researchers to focus on the most consequential questions and decisions. \u201cWe imagine it to be like a collaborator \u2014 a partner with you,\u201d says Vivek Natarajan, an AI researcher at Google, who helped to develop Co-Scientist.<\/p>\n Teams of AI agents boost speed of research<\/p>\n <\/a><\/p>\n That could change not just how scientists work, but also how they are valued. For generations, scientific progress has depended on researchers who could pose difficult questions, devise ways to answer them and make sense of the results. As machines take over more of that work, the scarce resource could end up being human scientific judgement: knowing which questions are worth asking and which lines of enquiry worth pursuing. \u201cThe most valuable part now is actually asking the question,\u201d says Ajay Agrawal, an economist at the University of Toronto Rotman School of Management in Canada, who studies AI\u2019s effects on innovation and entrepreneurship.<\/p>\n The goal, says Le Cong, a molecular geneticist at Stanford University in California and scientific co-founder of Phylo, is not to replace researchers but to free them from the routine aspects of discovery. \u201cWe\u2019re trying to move humans up the value chain,\u201d he says.<\/p>\n But the adoption of these AI tools also creates a troubling paradox. Researchers might need deeper knowledge than ever to judge whether an AI-generated idea is sound, even as students get fewer opportunities to build that expertise by doing the work themselves. The tools that promise to accelerate discovery could therefore make it harder to train the scientists needed to oversee them.<\/p>\n \u201cInvestigators with human skills that understand the full picture are going to be more valuable than ever,\u201d concludes Hector Zenil, a biomedical computing researcher at King\u2019s College London.<\/p>\n For Fyodor Urnov, who studies genome editing at the University of California, Berkeley, and was an early adopter of FutureHouse\u2019s Kosmos platform, the promise of AI research assistants comes with a simple rule: don\u2019t take their pronouncements on faith.<\/p>\n Urnov grew up in Moscow and remembers well a phrase popularized during the 1980s nuclear-arms-control negotiations between the United States and the Soviet Union. \u201cMy relationship with [Kosmos] is very much based on that Reagan\u2013Gorbachev proverb,\u201d he says: \u201cTrust, but verify.\u201d<\/p>\n Anna Pertl used Co-Scientist to look for ways to target the cancer-associated protein MYC.Credit: Anna Pertl<\/p>\n A verification step was built into a test of Google\u2019s Co-Scientist by researchers at Imperial College London, who used the platform in 2025 while investigating how bacteria swap snippets of DNA across species7<\/a>. They had already collected experimental data and formed a theory, but rather than feeding those into the system, they held the unpublished results back and asked the AI tool to work from the publicly available literature and data sets alone. Only after it had produced its theory did they reveal the experimental findings and compare the platform\u2019s work with their own.<\/p>\n Co-Scientist\u2019s theory of what was happening closely matched what the researchers, microbiologists Jos\u00e9 Penad\u00e9s and Tiago Costa, had spent years working out: bacterial DNA elements can borrow parts of viruses to spread between distantly related hosts. Thanks to its ability to reason across and connect information from many sources, Co-Scientist reached essentially the same conclusion in about two days7<\/a>. Had the Imperial researchers used the tool from the start, Costa says, it could have saved them the years of exploratory experiments that they conducted before they hit on the right mechanism.<\/p>\n \u201cIf you think about scientific discovery as a 100-metre race,\u201d he says, with an AI assistant \u201cyou start this race at, say, metre 30 instead of metre 0\u201d.<\/p>\n AI assistants generally use several agents to attack the same question from different angles. They can break scientific questions into smaller tasks, with agents tackling different pieces in parallel. The Google system \u2014 and competing platforms by FutureHouse, Huawei Technologies in Shenzhen, China8<\/a>, and Sakana AI in Tokyo9<\/a> \u2014 can explore several possible paths at once, allowing useful connections to emerge from the interplay between agents rather than from a single chain of reasoning.<\/p>\n \u201cYou have all these AI agents with different backgrounds to give different perspectives,\u201d says Kyle Swanson, a computer scientist at Stanford who helped to develop Virtual Lab<\/a>, an early academic AI research-assistant platform10<\/a>. \u201cIt gives you instant access to interdisciplinary experts.\u201d<\/p>\n Importance of good questions<\/p>\n But even a team of AI specialists requires someone to point the tool towards a worthwhile problem and to frame the challenge in such a way that the system can tackle it productively.<\/p>\n For a narrowly defined problem, the goal might be obvious<\/a>. Perhaps this is why decades-old problems posed by the late Hungarian mathematician Paul Erd\u0151s<\/a> and solved by an AI tool in May proved fertile ground for machine logic<\/a>. However, in open-ended science, deciding what to ask, and what would count as a meaningful answer, remains a fundamentally human task.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/il\/wp-content\/uploads\/2026\/09\/d41586-026-02931-5_52456372.jpg\"\/)
<\/p>\n![\"Portrait](\"https:\/\/www.newsbeep.com\/il\/wp-content\/uploads\/2026\/09\/d41586-026-02931-5_53718404.jpg\"\/)
<\/p>\n
![\"\"](\"https:\/\/www.newsbeep.com\/il\/wp-content\/uploads\/2026\/09\/d41586-026-02931-5_52308700.jpg\"\/)
<\/p>\n