NCI leaders: AI offers promise for cancer research but requires better data, controls; new "Cancer AI Conversations" series announced

Leaders at the National Cancer Institute said generative artificial intelligence could accelerate cancer research but that its benefits will depend on higher-quality, better-governed data and models.

"We don't have that really control in any model yet," said Jeff Schilling, NCI chief information officer, describing the institute's current technical safeguards for computation but noting a gap around validating and assuring model outputs. Tony Kurlavage, director of NCI's Center for Biomedical Informatics and Information Technology, urged private, auditable models and clearer documentation of inputs so outputs are less likely to produce misinformation or 'hallucinations.'

The officials said the NCI faces two linked problems: many legacy datasets are incomplete or nonrepresentative, and data are stored in institutional silos that make standardization and reuse difficult. "Their clinical centers where they're doing treatment are completely disconnected from their research centers even within the same institution," Kurlavage said, describing how electronic health records and research systems were designed for different purposes.

To address interoperability and harmonization, both leaders pointed to ongoing efforts and partnerships. Kurlavage highlighted NCI engagement with ARPA‑H's biomedical data fabric work as one avenue to test new technical approaches for mapping and repurposing legacy data; he suggested language models may help accelerate labor-intensive harmonization tasks. Schilling said partnering with commercial vendors and using off‑the‑shelf platforms can be a pragmatic way to bring advanced tools into research settings while leveraging NCI data and expertise.

The NCI also announced a new series, Cancer AI Conversations, to surface timely technical and policy questions about AI in cancer work. The first session is scheduled for Jan. 23 at 11 a.m. and will examine prompt engineering for generative systems; NCI said recordings will be posted on its website. Organizers named presenters from Microsoft Research, the University of Texas at Austin and Stanford University and said the series will mix short presentations with moderated discussion.

The officials repeatedly emphasized the need to design studies with representative samples, consistent nomenclature and sufficient sample sizes so AI tools produce reliable, equitable results. They also stressed the importance of documenting system architecture and APIs to allow reproducible, auditable use.

Next steps identified in the conversation include continued internal discussions about validation standards, engagement with ARPA‑H–funded projects to prototype data‑harmonization approaches, and the NCI's public series to convene experts and share findings. Recordings and links to the announced series will be made available on the NCI website.