As AI continues to reshape medicine and biomedical research, the Cancer Center is taking a thoughtful approach to its potential—and its limitations. Below, Julian Hong, MD, the Cancer Center’s new head of AI, discusses the role AI can play in improving patient care, accelerating discovery, and enabling new research, while emphasizing the importance of data quality, privacy, reproducibility, and real-world usefulness. Dr. Hong also shares how faculty, trainees, and early-career investigators can help shape the cancer center’s AI future.
Q: This is a new role in the cancer center. How do you see its purpose, and what do you hope to accomplish?
AI is playing an increasingly important role in medicine, and I see the role as helping the cancer center develop a strategy for AI. It’s important for us to make sure that we are making AI and its capabilities useful across our clinical and research communities. Too often, AI can be a solution looking for a problem, and the real-world gains can be more modest than the excitement around the technology might suggest. My goal is to make sure we focus on where AI is genuinely useful, whether that means improving care for our patients, accelerating our research, or enabling us to tackle questions we couldn’t address before.
Q: What opportunities do you see for AI to make the biggest difference in cancer research and care over the next five to ten years?
A lot of attention in AI has been focused on language models and administrative tasks in the short term since these feel like solvable problems, but I think the bigger opportunity is our ability to learn from the enormous amount of data generated through clinical care and research. Advances in computational methods have already improved prediction, clinical decision-making, and hypothesis generation, and AI gives us an opportunity to take those approaches further.
Q: What kinds of infrastructure, resources, or support do you think researchers will need to make AI a useful part of their research programs?
A big part of this is lowering the barriers to bringing AI to important use cases—whether that means implementing a tool clinically, designing a trial around AI-enabled care, or helping researchers identify tools that fit their needs. Some of that will involve building purpose-built tools internally, and some will involve identifying and supporting the best external options. Another major opportunity is building on UCSF’s strong data and analytical infrastructure while making it more usable and accessible to the broader research community.
Q: How do you think about balancing innovation with concerns about data quality, bias, privacy, and reproducibility?
I’ll start with privacy, since we have a particular responsibility to our patients who trust us to safeguard their information and to use it responsibly, so that will always come first.
Data quality, bias, and reproducibility are all major issues and have been for years. I generally think that meaningful innovation is not possible without addressing them. AI tools can’t be useful if the underlying data are poor, the results are biased, or the findings aren’t reproducible.
Q: If a faculty member or trainee has an idea for an AI-enabled research project, what should they do? How can they get started, initiate a collaboration, or apply for pilot funding?
Reach out to us! Even if it’s still early. There are several opportunities for collaboration across the different AI entities on campus like the Bakar Computational Health Sciences Institute, Division of Clinical Informatics and Digital Transformation (DoC-IT), or Computational Precision Health. Many of us are interested in helping however we can. Depending on the project, many of the intra- or extramural funding mechanisms can be good opportunities (many curated on the Cancer Center website!).
Q: What misconceptions about AI in cancer research do you encounter most often? In what ways is it overhyped or undervalued?
I think you see both overhyping and undervaluation, and unfortunately, overhyping typically leads to disenchantment and consequent undervaluing. AI carries so much promise, but so much of it is overstated in the everyday discussion today. I think a big contributor is the personification of large language models, which is the typical “AI” people think of today.
The reality is these are still mathematical representations of language (which can be small enough to fit on your personal device!) and are faced with the limitations that come with that. AI is also not monolithic. We’re in this era of “agents” and “multimodal” models, which have opened great capabilities, but behind the scenes AI has become a system of linked tools, so it’s not all-powerful as some would have you believe.
Q: How can AI help accelerate discovery without replacing the creativity and expertise of scientists?
AI is a great iterative tool and can help scientists verbalize and think through problems, explore existing knowledge, and develop new hypotheses. But it is fundamentally dependent on existing data and knowledge, so we must think about what the results mean and what the right next steps are. AI can hopefully make us more efficient in the questions we ask and the problems we solve.
Q: What role do trainees and early-career investigators have in shaping the cancer center's AI future?
Their involvement is essential. In general, our community should be the driver of our AI future. As I said above, usefulness is the most important thing, and it will take our community, especially our trainees and early-career investigators, to sort out how AI can be useful for us together. We love having junior folks as part of our team and one of my favorite parts of my job here at UCSF in general is supporting the next generation.
Q: Are there areas of cancer research—from basic science to population health—where you think AI has especially untapped potential?
Rather than point to one specific area, I think one of the biggest untapped opportunities is building AI tools that are truly purpose-built for particular scientific and clinical problems. Many current tools are designed to be broadly capable, but realizing their value in any field often requires deep domain knowledge and highly specialized (and high quality) data. Bringing AI expertise together with domain experts to build and adapt tools for specific problems could have an enormous impact across basic science, clinical research, and population health.