Clinicians and scientists should work alongside large language models rather than seek to replace human expertise, say the authors of a new Lancet Viewpoint that argues collaboration is the safest path to accelerating discovery.
Clinical researchers should embrace a new model of human–AI collaboration that positions large language models as intellectual partners rather than replacements for scientific expertise, say experts.
Writing in a Viewpoint in The Lancet Digital Health, the international group of researchers argued that a “co-intelligence” framework could help unlock the promise of generative AI in medicine while preserving scientific rigour, accountability, and trust.
They said the most effective future for medical research lay in iterative collaboration between humans and AI systems rather than automation.
“Traditional medical research relies on siloed expertise, with methodological and data siloes restricting researchers’ capacity for innovation,” they wrote.
“Integrating LLMs [large language models] into research workflows can potentially disrupt this status quo.
“For example, LLMs might enhance interoperability by translating different data schemas and terminologies to facilitate comparison and analysis; identifying patterns across large datasets, multimodal data, and across different disciplines; or bridging methodological siloes to generate new approaches inspired by other disciplines to address specific challenges.
“This capacity to synthesise information from disparate domains might aid in overcoming intellectual siloes, catalysing the formulation of novel research hypotheses and innovative interdisciplinary solutions.”
Their comments come amid growing enthusiasm for LLMs, which are increasingly being used to support literature reviews, hypothesis generation, data extraction, and analysis of large volumes of unstructured biomedical information.
While these technologies could process and synthesise information at a scale impossible for humans alone, the researchers warned that they remained vulnerable to hallucinations, bias, and factual inaccuracies.
Under the co-intelligence model, researchers would retain responsibility for critical judgement, ethical oversight, and contextual interpretation, while AI systems contributed computational power, pattern recognition, and rapid information synthesis.
The researchers said that combining these complementary strengths could produce insights that neither humans nor AI could achieve independently. They cited rare disease research as an example of the potential benefits of this approach.
“Rare diseases represent a particularly illustrative and compelling example. Collectively, such diseases affect up to 442 million people globally; however, only less than 3% of these diseases have known treatments,” the researchers wrote.
“Although special fast-track regulatory pathways have been designed to mitigate some of the pain points, the small numbers of patients for each individual disease has not yet inspired sufficient investment in funding or resources necessary to drive breakthroughs in cures.
“A co-intelligent set-up has circumvented this challenge by facilitating an unconventional combination of medications to successfully treat two patients with rare diseases – one with POEMS syndrome and another with Castleman disease.
“The authors have subsequently built on this approach to pioneer a new field of drug repurposing.
“Beyond traditional approaches that require extensive drug development and clinical trials on a disease-by-disease basis, computational pharmacophenomics leverages LLMs, knowledge graphs, and machine learning algorithms to bring together multiple disparate sources of big biomedical data to draw associations among the 74 million extant drug–disease matches.
“A human-in-the-loop system was used to iteratively refine model outputs and apply domain expertise to identify biologically plausible targetable modes of action that could be tested and validated by human collaborators, on a scale several orders of magnitude larger than hitherto seen.”
However, the authors warned against overestimating current AI capabilities. They noted that LLMs remained prone to generating convincing but incorrect information and that inappropriate reliance on AI could threaten scientific validity.
In medical research, such errors had the potential to influence clinical practice, policy decisions, and patient outcomes.
The researchers also raised concerns about “automation bias” and “cognitive surrender”, in which researchers become overly dependent on AI-generated outputs and reduced their own critical evaluation.
Over time, this could erode core research skills and make it more difficult to detect errors or challenge flawed conclusions.
To mitigate these risks, the researchers called for greater investment in AI literacy across the research workforce.
They said clinicians and scientists needed a practical understanding of how LLMs function, their limitations, and the circumstances in which they could and could not be trusted.
Related
Prompt engineering, critical appraisal of AI outputs, and familiarity with emerging AI tools were also increasingly important competencies, they said.
“We do not question the existence of cars or debate their ethics; instead, we teach people to drive – safely, at scale, and within a regulated framework – and this innovation has completely transformed the face of human history,” they wrote.
“In much the same way, equipping clinicians and researchers with the ability to question, adapt, and co-create with LLMs might offer a path to expanding the epistemic range and speed of medical research in an increasingly data-driven landscape, while retaining human judgement and accountability at its core.”
The researchers outlined four broad considerations for organisations integrating LLMs into research workflows – assessing whether AI was appropriate for a particular task; evaluating potential gains in efficiency and quality; ensuring robust validation and reproducibility; and addressing ethical, governance, and regulatory requirements, particularly when patient data were involved.
Looking ahead, the researchers called for the development of standards for reporting human–AI collaborative research, improved explainability of AI-generated outputs, and formal methods to evaluate the effectiveness and safety of co-intelligent systems.
They also advocated for integrating AI fluency into medical and research education in the same way that statistics and research methodology were currently taught.
Rather than debating whether AI should have a role in medicine, they said that the priority should be ensuring clinicians and researchers knew how to use it responsibly.
“By addressing these challenges, we can ensure that LLMs become a catalyst for medical innovation and improved patient outcomes without compromising the foundational principles of scientific inquiry and rigour,” the researchers concluded.
