Anushka Gupta, Yug Marwaha, and Matthew Montgomery
A recent article, titled, “AI co-scientists are revolutionizing how research is done,” published on the online platform, Nature, describes how researchers are beginning to use AI “co-scientists” to help with scientific research. The article follows molecular geneticist Philine Guckelberger as she uses Kosmos, an AI tool developed by FutureHouse, to search millions of rows of data for patterns. We chose this story to illustrate on one hand, how AI might help researchers make discoveries and on the other hand, to highlight the reasons why human interventions is still needed. As students studying the fields of computer science, data science, finance, and molecular and cellular biology, we’re interested in how AI could change scientific processes and research.
Initially, the use of AI in scientific research was limited to summarizing articles and lengthy papers, to finding relevant resources on a given topic. However, presently AI can be used to design scientific experiments and researchers have even begun building AI systems specifically for scientific research. Robin, a multi-agent system that supports end-to-end scientific discovery and Google Co-Scientist, a multi-agent system that works alongside scientists as a research collaborator, are two examples of this growing approach. For example, AlphaFold, an AI system that was developed by Google Deepmind to predict highly accurate protein structures has transformed protein research and structural biology. What started as a tool to visualise and predict proteins easily has now transformed to predicting new protein structures with desired properties, mutations and structures. As the article, AlphaFold2 turns five mentions, “for biologists, it has established a new infrastructure for accurate, large-scale protein structure prediction, addressing a long-standing challenge in structural biology.” This Nobel prize winning tool has helped scientists to back their hypotheses and optimize their experimental conditions by allowing them to visualize crucial interactions of their protein of interest. In fact, the newer and better AlphaFold 3 has even incorporated predictions of nucleic acids and molecular complexes, allowing scientists to view interactions not just between different proteins, but also between proteins and DNA (the string that stores all our genetic information).
AI is deeply integrated within the process of scientific research now; the robots used have been automated and made user friendly with AI interface. Liquid handling robots are an essential part of research because they help us reduce human error. They are used to move around liquids in the microliter scale, i.e., smaller than a droplet. The accuracy with which these robots transfer quantities as small as 1 uL is incomparable to any experienced scientist. These robots are still programmed using coding languages, and we still cannot use the English language to write programs for robots to make them functionable. But what has already been set in motion and will hopefully gain full potential in the near future is the use of AI to program in the way which will allow students and non-programmers to operate these robots, find errors, and optimize them according to their needs.
One good example of this is PyLab robot, an open-source framework that provides a common interface for different types of robots that handle liquids in labs. This means researchers won’t have to spend time learning a different programming interface or language. They can automate varied levels of programming expertise with AI and in turn focus on their research questions and on improving the quality of research. The question is: where, when, and how do humans intervene and withhold their agency? Can we expect that artificial intelligence will do all the research in a few decades, or is it just about asking the right questions, because AI can hallucinate, make up fake resources and articles, or totally summarize an article incorrectly?
We believe that AI is still not proficient, reliable or smart enough to do research all on its own. In the same Nature article mentioned earlier, where we learned how Kosmos was used to analyse huge amounts of data, we also learned that AI hallucinated, again. It needed human input to fix the mistakes it was making. Guckelberger had to guide the AI in the right direction and explain what it needed to be looking for. AI needed human guidance. After being corrected, Kosmos ended up identifying a novel and unexpected pattern in cancer cells. This pattern was also confirmed by the scientist independently.
What happened here is a great example of the use of AI in advanced research methods. Sure, Kosmos accelerated the discovery process for Guckelberger by drastically taking millions of data points and analysing them in hours, when the same job would have taken the molecular geneticist days. The data was confirmed independently and the finding was meaningful. AI was built to help humans. It works on inputs from humans, where everything it has learnt, built, and done has been through inputs, from its own developers and users.
There are researchers who believe that AI lacks the creativity or motivation that humans possess, when solving real world problems through research. Most of the time, we question existing theories and pathways because we have a personal connection to the research we do. The lack of creativity and personal connection inhibits AI from generating original hypotheses and predicting unexpected outcomes. As this article demonstrates, there are scientists who argue that what years of practical expertise of researchers and the combined wisdom a team provides to the research process, might make it impossible for machines or AI to replicate.
We recognize that the use of AI in scientific research is new and debatable. There are qualms, as with everything. There are bigger questions about data protection, accountability, and privacy. Can AI be trusted with sensitive data from academia? Is it possible to build a safe, foolproof system to use AI and trust it with data? Should AI companies be allowed to take credit for another researcher’s work? Privacy in science is another important question or argument we can think of. We would need stricter rules, guardrails, and policies to protect the participants, researchers, and the data being used in the research. Another question that needs asking is whether the benefits of using AI in research are worth taking risks entailing privacy, ownership, and the integrity of scientific research.
These concerns do not imply that we should stop using automation in the process of scientific research. We believe that science will likely use more AI tools and robots in the future for research compared to other fields, where AI will assist scientists in carrying out repeated, as well as boring lab-related tasks. These tools can help us save more time by accelerating research speed.
AI and robots can be utilized as “co-scientists” only when the researchers are asking the right questions and deciding whether a task is worth delegating to AI. Verifying these results is another important part of this process because AI can sometimes come up with fake results, aside from hallucinating. We can use it to make research faster, but scientists and researchers have to take responsibility and establish human agency for the work that has been completed. They also have to make sure that the results are supported with evidence. We can use machines to do more work in less time, but the decision-making part stays with us humans.
Course Instructor: Dr. Ainehi Edoro
Associate Professor, Dept. of English
email: aedoro@wisc.edu
Editor: Shrinjita Biswas
PhD Candidate, Interdisciplinary Theatre Studies
WORKS CITED:
Co-Scientist Team. “Co-Scientist: A multi-agent AI partner to accelerate research.” Google DeepMind. 19 May 2026.
https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/
Dolgin, Elie. “AI co-scientists are revolutionizing how research is done:Artificial-intelligence systems can generate hypotheses, design experiments and analyse data — but humans still need to decide what makes sense.” Nature. 21 September 2026.
https://www.nature.com/articles/d41586-026-02931-5
Editorial, “AI firms must play fair when they use academic data in training.” Nature. Vol. 632, 953, 2024.
doi: https://doi.org/10.1038/d41586-024-02757-z
Editorial, “AlphaFold2 turns five.” Nature Computational Science, Volume 6, July 2026, 673–674.
https://doi.org/10.1038/s43588-026-01031-8
Editorial, “Why AI cannot do good science without humans: With the arrival of ‘AI scientists’, it’s as well to remember that human wisdom, empathy and sheer messiness are as much part of progress as are process and efficiency.” Nature. May 19 2026.
https://www.nature.com/articles/d41586-026-01551-3
Hinks, Michaela. Ghareeb, Ali. Chang, Benjamin. Mitchener, Ludovico. “Demonstrating end-to-end scientific discovery with Robin: a multi-agent system.” Future House, published by Sam Rodriques. 19 May 2026.
https://www.futurehouse.org/research/demonstrating-end-to-end-scientific-discovery-with-robin-a-multi-agent-system
OpenAI. “How ChatGPT and our foundation models are developed.” (last update: 2 months ago).
https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed
University of Pennsylvania, “Artificial Intelligence (AI) in Human Research.” Human Research Protections Program.
https://irb.upenn.edu/homepage/biomedical-homepage/guidance/types-of-biomedical-research/ai-research/
Walters, William H. Wilder, Esther Isabelle. “Fabrication and errors in the bibliographic citations generated by ChatGPT.” Scientific Reports, Vol. 13, 14045, 2023. https://doi.org/10.1038/s41598-023-41032-5
Image: Night Moth by Brassaï 1935
Reproduction in Minotaure No. 7
Kunstbibliothek
Image Source: Art Blart
