Califf Warns of Regulation of Large Language Models
FDA commissioner Robert Califf is warning that large language models (LLMs) and their use by industry will require regulation because the agency runs the risk of being “swept up quickly by something that we hardly understand.” ChatGPT and other models are becoming more accessible to the public and their use by regulated industry is likely to explode. For example, LLMs are beginning to be used by the pharma industry, according to the blog LabTwin, to predict drug interactions and side effects, and optimizing clinical trials by scanning patients’ medical histories to target eligible participants, among other tasks.
Speaking at the 2023 Science for Patient Engagement Symposium in Washington, DC 5/8, Califf said digital technology advancements are likely to disrupt the status quo. “We have enormous potential as the world enters the fourth industrial revolution in which rich and diverse sources of digital data are available at scale in real time with potentially unlimited storage capacity, and these data are becoming widely available as part of the clinical care system,” he said. “Large language models are the next step that appears to be ushering in the revolution that many of us were hoping for.”
But the promise of technology also comes with some concern. “Algorithms, including the new large language models, evolve after they are put into practice,” Califf cautioned. “It is unlikely that an accurate algorithm will stay accurate when deployed in real life over time. It needs adjustment and measurement of its operating characteristics continuously throughout the life cycle. How do we get that done?”
Califf also said false or misleading information spread on social networks could increase from LLMs. “This new digital era can connect us, but technologies like large language models give almost everyone the potential to produce false narratives or even so-called deep fakes — fabricated images and voices,” he said. “A key part of a successful transition to digital health is an effective regulatory scheme to guide digital technologies to improved human outcomes and interaction… I could go on and on, but I see the regulation of large language models as critical to our future.”
Earlier this year, Califf opined in an online post that it is just a matter of time before the agency receives its first artificial intelligence-assisted” marketing submission (see earlier story) “As we develop much better systems of evidence generation, data science and statistics form the bedrock of generating evidence from data and information,” he said in the post. “The FDA has a talented and influential group of quantitative experts, but to regulate the information-rich, digital world of the future we need to make it a conscious focus. An effective organization for knowledge sharing and professional advancement is needed across the multiple professional disciplines with a role in quantitative analysis, data management and computer science.”