By admin 
The pivot toward a "competency-based approach" represents a fundamental reimagining of how the federal government validates medical technology. Traditionally, the FDA’s gold standard for device clearance—the 510(k) pathway—relies on a "predicate" system where a new device proves it is "substantially equivalent" to one already on the market. However, generative AI, characterized by Large Language Models (LLMs) and multimodal systems like GPT-4 or Gemini, does not fit neatly into this box. Unlike a digital thermometer or even a standard AI imaging tool that identifies a lung nodule, generative AI is non-deterministic; it can produce different outputs for the same input, and its internal logic is often a "black box" even to its creators. By focusing on "competency," the FDA is exploring a model that evaluates the underlying capabilities and the rigorousness of the developer’s quality management systems, rather than just the static output of a specific version of the software.
This shift is a direct response to the explosion of AI in the clinical space. As of 2024, the FDA has authorized over 900 AI and machine-learning-enabled medical devices, but the vast majority of these are "locked" algorithms used in radiology, cardiology, and neurology. These tools are designed to do one thing—such as flag a potential stroke on a CT scan—and they do not change once they are deployed. Generative AI is a different beast entirely. It is designed to be versatile and, in many cases, to learn or adapt. Abramson’s vision, as shared with STAT, suggests that the agency will release a series of formal policy guidance documents. These will likely include broad overviews of how generative AI fits into the "Software as a Medical Device" (SaMD) category, as well as "narrowly constructed specialty guidance" for high-stakes applications where the complexity of the AI poses unique risks to patient safety.
The stakes of this regulatory endeavor cannot be overstated. In the healthcare sector, the primary concerns regarding generative AI are "hallucinations"—the tendency of LLMs to confidently state false information—and "algorithmic drift," where a model’s performance degrades over time as the data it encounters in the real world shifts away from its training set. If a generative AI tool used for oncology treatment recommendations "hallucinates" a dosage or a drug interaction, the results could be fatal. Consequently, the FDA’s focus on competency involves looking at how a company monitors its AI in the "wild." A competency-based approach might require developers to demonstrate that they have the infrastructure to detect drift in real-time and the "competence" to retrain or pull a model before it causes harm.
This regulatory evolution is also heavily influenced by the political climate. In October 2023, President Biden issued a landmark Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. The order specifically tasked the Department of Health and Human Services (HHS), which oversees the FDA, with establishing a safety program to receive reports of—and act to remedy—harms or unsafe healthcare practices involving AI. Abramson’s comments reflect the FDA’s role as the frontline enforcer of this mandate. The agency is no longer just looking at the software code; it is looking at the entire lifecycle of the product, from data curation and training to post-market surveillance.
Industry experts have noted that the FDA’s Digital Health Center of Excellence is essentially trying to build a bridge while walking on it. The "competency-based" language echoes elements of the FDA’s earlier "Software Pre-Certification (Pre-Cert) Pilot Program." That program, which ran from 2017 to 2022, attempted to look at the "culture of quality" within a company rather than just the product. While the Pre-Cert pilot ultimately showed that the FDA might need new legislative authority from Congress to fully implement such a system, the lessons learned are clearly informing the current strategy for generative AI. By emphasizing "competency," the FDA is signaling that it wants to trust the developer’s process, provided that process is transparent and subject to rigorous auditing.
One of the most complex aspects of regulating generative AI is the concept of "intended use." Under current law, the FDA regulates products based on what the manufacturer claims the product does. If a company markets a generative AI tool as a "scribe" to help doctors take notes, it may fall under a lower level of scrutiny or even be exempt from certain regulations under the 21st Century Cures Act. However, if that same tool is used to suggest a clinical diagnosis, it becomes a high-risk medical device. The line between "administrative aid" and "clinical decision support" is becoming increasingly blurred as LLMs become more capable. Abramson’s promise of "narrowly constructed specialty guidance" suggests the FDA will soon draw much firmer lines around these use cases, potentially creating specific rules for AI used in mental health chatbots, surgical robotics, and personalized genomic medicine.
Furthermore, the FDA is grappling with the "black box" problem. Traditional validation requires understanding why a device failed. With generative AI, even the engineers may not be able to explain why a model reached a specific conclusion. To combat this, the agency is expected to emphasize "explainability" and "transparency" in its upcoming guidance. This could mean requiring developers to provide "nutrition labels" for their AI—clear documentation of what data the model was trained on, its known biases, and its performance metrics across different demographic groups. This is particularly crucial given the well-documented history of AI perpetuating racial and gender biases in healthcare, such as algorithms that underestimated the health needs of Black patients because they relied on historical cost data as a proxy for health status.
The global context also puts pressure on the FDA to act. The European Union’s AI Act is moving toward implementation, categorizing medical AI as "high-risk" and requiring strict conformity assessments. If the FDA’s rules are too divergent from international standards, it could create a fragmented market that slows down the adoption of life-saving tools. Conversely, if the rules are too lax, the U.S. risks a public backlash that could set the field back by decades. Abramson’s focus on providing "clarity" is a nod to the venture capital and startup ecosystem, which has seen billions of dollars poured into healthcare AI. These companies need to know the "rules of the road" before they can move from pilot programs to full-scale clinical deployment.
As the FDA wades deeper into these waters, the "competency-based approach" will likely be met with both praise and skepticism. Proponents argue it is the only way to keep pace with a technology that evolves every week. Skeptics, however, worry that focusing on a company’s "competence" rather than a specific product’s performance could lead to a "regulatory capture" scenario where large, well-funded tech giants like Google, Microsoft, and Amazon have an easier path to clearance than smaller, more innovative startups.
In the coming months, the medical community will be watching for the first drafts of these guidance documents. The FDA’s Digital Health Center of Excellence has a massive task ahead: it must foster an environment where generative AI can reduce physician burnout and improve patient outcomes while simultaneously serving as a bulwark against the inherent unpredictability of the technology. As Abramson noted, the ecosystem is expecting clarity. Whether the "competency-based approach" can provide that clarity without stifling innovation—or compromising safety—remains the central question for the future of digital medicine. The transition from "locked" to "generative" is not just a software update; it is a total paradigm shift, and the FDA’s new rulebook will define the healthcare landscape for the next generation.
