When AI co-scientists accelerate the pace of science, can academic oversight keep up?
Three University of Melbourne research teams are testing what happens when you give AI a scientific goal, complex data and time to think.
Introduction written by Dr Emerson Keenan, Program Manager, AI for Research Innovation, Melbourne Medical School*
Most people’s experience of generative artificial intelligence begins and ends with a chatbot: type a question, get an answer. Tools like ChatGPT, Claude and Gemini have become remarkably useful for summarising text, drafting emails, even writing code. But they often operate in limited exchanges, one question in, one answer out.
That’s a long way from how the scientific process typically works.
Scientific discovery is iterative and messy. It demands cycling between hypotheses and data, synthesising findings across hundreds of papers, cross-referencing databases, and doing all of this while maintaining the methodological oversight that separates insight from speculation. A chatbot can help with pieces of that process. But until recently, it could not drive it.
A new class of AI system is now emerging that attempts something fundamentally different. Known as “AI co-scientists,” these platforms don’t just respond to individual prompts. When given a research goal and a complex dataset, they can autonomously break the problem down into dozens of sub-tasks, query scientific literature, run statistical analyses, cross-reference datasets, and synthesise their findings into structured reports. A single run can involve hundreds of individual analytical steps and take up to twelve hours to complete.
It is this shift, from AI as a productivity tool to AI as a research partner, that the Melbourne Medical School recently set out to investigate. Earlier this year, the School launched its AI Co-Scientist Pilot Program, selecting three teams to test these tools against real research problems. The program was designed with clear guardrails: all teams were required to work exclusively with publicly available datasets to ensure no private information was ingested by the AI platform, and each team was paired with a mentor to provide support as they scoped their research questions and interpreted the AI's outputs.
The goal was not to showcase the technology, but to honestly assess what it could and couldn’t do, and to begin building the practical knowledge that researchers will need to develop as these tools become more widely used.
Three labs, three experiments
The three teams selected for the pilot program cross fetal medicine, population health, and precision oncology. Each brought a different dataset, a different research question, and a different set of expectations.
One team saw the AI analyse a complex national dataset on area-level childhood disadvantage and, within hours, find new patterns and relationships in the data. Another watched it cross-reference thousands of proteins against online databases, completing work that would have taken a researcher weeks. A third received a clinically sound prioritisation of cancer-driving genes from genomic data, only to find that many of the AI’s cited references pointed to unrelated scientific papers.
All three projects are real. All three are underway at The University of Melbourne. And together, they paint a picture of the future of research that is both promising and nuanced: emerging AI tools can genuinely accelerate discovery, but getting trustworthy results still takes careful, expert-led effort.
Reading the signals in amniotic fluid
Professor Lisa Hui, from the Department of Obstetrics, Gynaecology and Newborn Health, came to the pilot with a dataset her team knew well, a published proteomics study of human amniotic fluid extracellular vesicles.
These tiny membrane-bound packages, shed by fetal tissues into the surrounding fluid, carry protein cargo that may reveal how a developing baby’s organs are maturing.
The team wanted to know whether an AI, given the same data, could find signals they had missed.
In several respects, it delivered. The platform rapidly connected the team’s findings to animal studies, transcriptomic references, fetal tissue atlases, and public extracellular vesicle databases: a breadth of cross-referencing that would have taken weeks by conventional methods. It identified candidate biomarkers the researchers had not prioritised and, most intriguingly, suggested an entirely new research direction, opening a line of inquiry for future experiments.
But some of the most instructive findings were where the AI stumbled. On closer inspection, its headline “discovery”: an apparent gestational age-based sorting mechanism within cells was based on a small number of samples with limited statistical support. The team’s own domain expertise suggested a simpler explanation: the changing protein patterns likely reflected shifting tissue contributions to amniotic fluid over time, not a true intracellular sorting mechanism. It was a plausible and internally coherent hypothesis, but it was speculative rather than proven.
"Using AI in this project felt less like asking a machine for answers and more like working with a very fast, very well-read research assistant,” says Professor Hui. “It helped us see our own data differently and ask new questions. But when we looked more closely, some of the most exciting findings were still inferential and speculative. The value was not that it gave us answers, but that it suggested new questions.”
Mapping early childhood disadvantage
Dr Sarah Gray and the Changing Children’s Chances team from the Centre for Community Child Health at the Murdoch Children’s Research Institute tasked the AI with a different kind of challenge. They provided a complex national dataset on early childhood disadvantage that needed extensive cleaning, exploration, and pattern recognition before any meaningful analysis could begin.
Here, the AI proved most valuable in the early, labour-intensive stages of research. It accelerated the iterative cycle between research questions, data exploration, and preliminary results. It helped reveal area-level patterns of disadvantage in the data that the team could then examine more closely. The speed at which the AI could generate and test exploratory hypotheses compressed weeks of preliminary work into hours.
The experience also highlighted where human judgement remains indispensable. Decisions around data cleaning, handling outliers, managing missing data, and defining variables still required careful input from researchers. A technical (and human) understanding of both the dataset and the communities that the data represents will always be critical.
The sheer volume of outputs the AI generated could, at times, be overwhelming rather than illuminating. It demanded rigorous methodological work by the team to separate genuine insights from statistical noise. And while the tool could find promising patterns, it could not yet produce publication-ready outputs, meaning substantial refinement was still needed before results could be shared.
“The AI didn't replace the work, it changed it.” says Dr Gray. “Using AI sped up exploration and surfaced complex patterns in the data quickly, but knowing which results to trust and what needs validating still relies heavily on expert judgement.”
Prioritising cancer genes at scale
If the first two teams tested the AI’s abilities in data exploration and hypothesis generation, the third posed what might seem like its most natural task. A collaboration of genome curation specialists from the Collaborative Centre for Genomic Cancer Medicine (a joint venture of the University of Melbourne and Peter MacCallum Cancer Centre), led by Dr. Andrew Fellowes (Peter MacCallum Cancer Centre) and Dr. Joep Vissers (Department of Clinical Pathology), asked the AI to sift through published literature to generate lists of genes associated with individual tumour types and therapy response, and evaluate mutations for their clinical significance in diagnosis, treatment, and prognosis.
The AI performed well at generating comprehensive gene lists and producing variant prioritisation rationales: tasks that require integrating information from hundreds of papers. For a field where keeping pace with the literature is a constant challenge, the speed advantage was significant.
But the limitations were equally instructive. The platform could not access a number of paywalled journals, which meant gaps in its evidence base. It made factual errors, and in some cases misapplied the established rules used in clinical gene variant curation: rules where precision is not merely academic but directly affects patient care. Perhaps most tellingly, the team found that formulating simple validation criteria for the AI’s complex outputs was itself a significant intellectual challenge.
The researchers believe that more sophisticated prompt design could mitigate some of these issues, but the experience underscored that deploying AI in a clinical genomics context demands an especially high bar for verification.
What the pilot revealed
Taken together, the three case studies tell a consistent story. The AI co-scientist was genuinely useful, not as a replacement for researchers, but as a force multiplier for the early, exploratory phases of scientific work. It quickly synthesised vast literature, cross-referenced datasets that no individual researcher could hold in their head, and proposed hypotheses that opened new avenues of inquiry.
But in every case, the outputs required careful expert scrutiny before they could be trusted. Spurious citations, overconfident conclusions, and ultimately unsupported hypotheses were observed in all scenarios. The AI’s most dangerous outputs were not the ones that were obviously wrong, but the ones that were almost right: coherent enough to be convincing, yet built on foundations that only a domain expert would think to question.
The pilot also revealed the importance of how questions are framed. Broad, exploratory prompts yielded the most unexpected and interesting directions, while tightly constrained prompts produced more accurate but less novel results.
Perhaps the most profound shift was in the researcher’s role itself. Rather than executing analyses directly, researchers found themselves defining goals, designing prompts, and most critically, validating outputs. The work didn’t disappear; it merely changed form. The bottleneck moved from performing the analysis to knowing whether or not to trust it.
The road ahead
The Melbourne Medical School’s program is now compiling its findings into a practical guide: not only as a manual for how to best use these emerging AI platforms, but also as a framework for the judgement calls that surround them. A key ambition for future scaling is also to expand internal access to these platforms and enable researchers to work with private datasets under appropriate governance frameworks.
The three teams’ experience suggests that AI will not replace the scientific process. But with the right oversight, it could reshape how quickly and broadly that process unfolds. The researchers who learn to direct these tools and validate their outputs, while maintaining the rigour that makes science trustworthy, may be the ones who define the next chapter of medical discovery.
The technology is moving fast. The question now is whether our frameworks for academic oversight can keep pace.
*The author used an AI tool to assist in writing this article, which was carefully reviewed with input from all teams