In research labs at Harvard Medical School, artificial intelligence isn't just a tool for occasional help. Many scientists now treat it as a partner - a co-scientist that helps plan experiments, analyze data, and predict treatment outcomes. A 2024 survey by Ithaka S+R found that 63% of biomedical researchers had tried generative AI for work, though only 7% used it regularly at that time, a figure that has likely grown since.
"They serve as collaborators," said Marinka Zitnik, PhD, of the Zitnik Lab at HMS. "Sometimes we refer to them as co-scientists because they perform science alongside human researchers, by helping them carry out tasks and analysis, like reading study results, synthesizing scientific literature, and analyzing biological data."
The biggest barriers to adoption, according to the Ithaka S+R report, remain concerns about accuracy and a lack of clarity around best practices for ethical use. Many scientists start with publicly available chatbots for basic literature searches. But Zitnik warns that "having a chatbot is not enough for medical research," since these tools pull from the open internet where "there are lots of things on the internet that are not to be trusted." A growing number of labs now build customized AI agents that connect directly to vetted clinical data, computational tools, and laboratory workflows that chatbots cannot reach.
Designing studies with fewer dead ends
AI helps researchers answer "what would happen if?" questions with more speed and broader information than before. Daniel Raftery, PhD, director of Northwest Metabolomics Research Center at the University of Washington School of Medicine, uses AI to generate research ideas he might not have considered. "Sometimes [the ideas] are off the mark, but sometimes there's a seed of an idea there, based on the prior literature and [other information], that I just hadn't seen or thought of or didn't search for," he said.
When digging into a project, AI tools can search studies to surface which experimental designs have been tried for similar treatments, what variables were tested, and what outcomes emerged. "If you have 10 variables you want to examine, you don't need to do 100 experiments," Raftery said. AI can "help pinpoint the experiments you really need to do, help you come up with a better experimental design so you don't waste time." This kind of computational narrowing is part of a broader shift toward AI for Science & Research, where automation handles the combinatorial heavy lifting that once consumed months of manual work.
Predicting treatment results before trials begin
Custom agents are moving beyond design support into predicting how treatments will affect real populations. Researchers at the Zitnik Lab, working with the University of Oxford and the Broad Institute, built an agent called ATHENA that performs treatment reasoning. Given complex patient case information, ATHENA assesses likely responses to specific treatments - including adverse events - by querying vast datasets on diseases, patient cohorts, and all drugs developed by the FDA since 1939.
In one case, ATHENA predicted that patients with a certain disease and medication history would face higher risk for an adverse event if they were on an ACE inhibitor for blood pressure. The lab then checked electronic health records and confirmed that those patients were indeed reporting the adverse event at higher rates. "We are seeing examples of how AI scientists can provide useful insights that can steer experiments in the lab, can steer the work of clinical researchers who might be designing clinical trials," Zitnik said.
Sandeep Sahani, MD, chair of radiology at UWSOM, pointed to the analytical scale now available. "Now we can look at metadata from 100,000 patients, but also parse out the data into cohorts - such as specific ethnicities, genders, and ages - to guide future courses of action," he said.
Automating routine lab work
Much of a lab scientist's workload - cataloging samples, preparing reports, managing data - is administrative and time-consuming. Researchers now use AI to consolidate findings from literature searches, maintain task lists, and draft first-pass manuscripts. "It's like having an extra pair of hands to handle some of the more routine aspects of research," said Carolyn Glass, MD, PhD, principal investigator at the Carolyn Glass Lab at Duke University School of Medicine.
On a more advanced level, lab leaders envision AI agents working with robotic systems to run tasks with human oversight. This could include preparing and analyzing thousands of biological samples and deciding what to test next without direct human instruction. "It's possible to envision an entire lab with robotic systems running experiments continuously," said Hai-Quan Mao, PhD, director of the Institute for NanoBioTechnology at Johns Hopkins University. "Students and researchers can establish the experimental framework, define the decision criteria, and oversee the results." For scientists looking to build these skills systematically, an AI Learning Path for Research Scientists covers data modeling, lab automation, and experimental design workflows.
Keeping humans in the loop
Researchers at academic health systems emphasize that human oversight remains non-negotiable. "We have to be rigorous in evaluating what AI provides," Mao said. "There has to be a check, a validation, because you cannot just simply trust whatever AI is telling you." His lab tests AI-generated predictions experimentally to confirm reliability. Glass's lab at Duke establishes "governance and guardrails" on how AI is used.
Scientists also stress that human judgment remains central to defining research questions. "This kind of research still relies heavily on our own expertise and creativity," Mao said. "Scientists remain responsible for defining the important questions and weighing which ideas to pursue. AI frees up the scientist to do more of that high-level thinking."
Why this matters for science and research professionals
The shift from using AI as a search tool to treating it as a customizable research agent changes which skills labs will value. Scientists who can build or direct custom AI workflows - connecting agents to internal datasets, vetted clinical resources, and robotic systems - will be able to test hypotheses and narrow experimental variables far faster than those relying on manual methods alone. The bottleneck is no longer data availability. It is knowing which questions to ask and how to configure the tools that answer them.
Your membership also unlocks:
