Russ Altman is a senior fellow (by courtesy) and science fellow at the Hoover Institution. He is the Kenneth Fong Professor of Bioengineering, Genetics, Medicine, and Biomedical Data Science, and a professor of computer science, by courtesy, at Stanford University. He holds an AB from Harvard College, an MD from Stanford University School of Medicine, and a PhD in medical information sciences from Stanford. Altman recently spoke with Donna Obeid, senior research program manager at Hoover’s Technology Policy Accelerator.

Donna Obeid: Russ, you host The Future of Everything podcast and help lead Stanford’s One Hundred Year Study on Artificial Intelligence. In medicine, AI’s early gains may come from helping clinicians reclaim time. Where is it already useful in practice, and where does the promise still run ahead of the evidence?

Russ Altman: Clinical documentation is the early “killer app.” Physicians value it because it saves them hours each week. Medical imaging is promising, but the next step is to move beyond impressive demonstrations and integrate these tools into clinical workflows in ways that increase radiologists’ productivity.

Early disease detection is an especially promising area of research, but it has not yet translated widely into practice, partly because of reimbursement and insurance issues. Insurers may hesitate to pay for tests whose savings accrue to a future insurer. If a patient switches plans after roughly three years, for example, one insurer may pay for early detection while another benefits from the later savings.

Donna Obeid: The promise of medical AI depends on access to representative, high-quality health data. What have you found about the representativeness of current datasets, and what role could the federal government play in improving access while protecting privacy?

Russ Altman: This is a critical issue. We wrote a paper showing that a large fraction of medical AI datasets come from just three states: New York, Massachusetts, and California. That raises a legitimate question about whether the resulting models will work well in other settings.

The federal government could help facilitate access to data from federally financed health care programs, including Medicaid and Medicare. Such data could help researchers train and validate models across a broader range of populations. Access would need to respect privacy and confidentiality, but I believe we can work through those issues.

Donna Obeid: Much of your research examines why people respond differently to the same medication. What is needed to make genetic information a routine part of prescribing?

Russ Altman: The science is there for pharmacogenomics—the use of genetic information to guide drug prescribing. The main barriers have been reimbursement and regulation.

Before the pandemic, some coverage decisions by United Health in favor of pharmacogenomics generated excitement. That momentum faded when the pandemic redirected attention. We need Medicare to approve pharmacogenomic services. The FDA also needs to provide clearer guidance about what is required for approval. It has not fully recognized some well-regarded resources, such as the Clinical Pharmacogenetics Implementation Consortium recommendations, and that uncertainty makes some companies nervous about bringing products to market.

Finally, some payers expect large trials to prove efficacy. Those trials would be expensive, and there are no obvious sponsors: pharmacogenomics applies across many drugs, many of them off-patent.

Donna Obeid: AI systems may change as they encounter new data and clinical environments. How should they be evaluated over time, and how should responsibility be shared when they fail?

Russ Altman: The FDA has floated some interesting models. One good idea is to let a product sponsor propose a reasonable strategy for longitudinal evaluation rather than requiring a single, one-size-fits-all approach. That places responsibility with the companies that have the greatest stake in solving the problem.

There are also established principles for evaluation, including benchmarks and “unit” tests. Basic tests can be run repeatedly to make sure that updates do not undermine performance on earlier tests.

In my view, responsibility should be shared by physicians, who use the tools and remain responsible for their medical decisions, and the companies that make money from them and advise physicians. Companies should not be able to pass the consequences of bad advice entirely to practitioners.

Donna Obeid: You testified before the Senate HELP Committee about AI’s potential to improve patient care and accelerate drug discovery. What did you emphasize, and what concerns did the senators raise?

Russ Altman: I focused on two opportunities: AI’s potential to improve health care, especially in underserved regions, and its potential to accelerate drug discovery.

The senators’ questions reflected concerns about the fairness of AI-driven decisions, the ability of clinicians to disregard AI advice, and the effects of AI on the workforce.

Donna Obeid: AI could accelerate drug discovery, but laboratory validation and clinical trials remain essential. Where is AI already helping move drugs through development, and where are its limits still clear?

Russ Altman: People often ask me, “Has AI invented a drug yet?” I think that is the wrong question. A better question is: Are companies using AI effectively to shorten the time from discovery to regulatory approval?

From conversations with colleagues and through my consulting work, I see companies actively looking for ways AI can speed up the process. Current uses include areas where AI has shown promise: predicting the toxicity of new drugs, identifying patients for trials, designing small molecules for a target pocket, and supporting clinical study design and monitoring.

Other areas that require deep knowledge of biology are not there yet. Current large language models are not very good at retaining all the details of biological systems—and many of those details remain unknown. Deep biological reasoning is still a challenge.

Donna Obeid: Looking a decade ahead, where do you expect AI to make the greatest difference in discovery and clinical care, and what safeguards should policymakers put in place, including against biological misuse?

Russ Altman: I expect a great deal of activity in AI-augmented discovery: AI systems will generate hypotheses, test them with robotic laboratory equipment, and report the results to human drug developers. It will be a joint effort, combining human experience, intuition, and biological knowledge with AI’s tireless ability to generate and test hypotheses. I do not think we should leave the work entirely to AI. People should design the approach and use AI as an intelligent assistant that can carry it out and report back.

In clinical care, we need ways to evaluate and validate AI tools to make sure they add value and improve the quality and length of patients’ lives. That means ensuring that systems work well for the populations they are meant to serve, are updated appropriately, and can be monitored for poor performance, errors, and changes in the populations they encounter. For clinical AI, policymakers will need to address liability, human control, alignment with human values, professional responsibility, ethical performance, transparency about the data used to build systems, and respect for privacy, security, and confidentiality.

One issue that also needs attention is the potential for AI-enabled discovery tools to be used to create pathogens for state or nonstate actors, including individual perpetrators. There needs to be a federal and global framework for monitoring the use of biological AI tools alongside relevant purchasing activity, so that concerning networks can be identified and disrupted when necessary. We addressed these issues at the Spirit of Asilomar meeting in 2025. I hope those conversations continue. There is still much work to do to prevent the misuse of AI-enabled biological tools.

Expand
overlay image