AI-driven clinical decision support tools are increasingly embedded in U.S. hospital workflows, yet a significant portion operates outside federal regulatory review. According to a new viewpoint published in The Lancet Digital Health, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory and the Jameel Clinic warn that this regulatory gap undermines patient safety and transparency. The authors call for a pragmatic framework to bring visibility and accountability to the four billion dollar market of clinical software without stifling innovation. Currently, approximately 65 percent of U.S. hospitals rely on AI or predictive models to identify high-risk patients and guide treatment. Despite their clinical impact, many of these tools have never undergone Food and Drug Administration evaluation. The regulatory uncertainty stems from policy shifts initiated by the 2016 21st Century Cures Act, which exempted certain clinical decision support software from medical device classification under specific conditions. The FDA’s 2019 draft guidance interpreted these exemptions permissively, prompting widespread adoption of AI tools. When the agency issued tighter final guidance in 2022, enforcement remained minimal, creating a prolonged period of regulatory ambiguity. A January 2026 policy update further expanded enforcement discretion, potentially reducing oversight for certain high-volume tools. This uneven regulatory landscape creates direct inconsistencies in patient care. For example, the widely deployed Epic Sepsis Model and Epic Deterioration Index lack FDA clearance, whereas competing tools like the Sepsis ImmunoScore and PeraTrend have received it. Similarly, the Tyrer-Cuzick breast cancer risk model, integrated into national clinical guidelines and used by insurers to authorize MRI screening, has faced no FDA review despite documented concerns regarding its predictive accuracy and disproportionate underestimation of cancer risk in Black women. Researchers note that hospital-developed or embedded software often bypasses scrutiny that commercial device manufacturers must face, placing validation burdens inversely to clinical impact. The authors emphasize that a blanket regulatory crackdown is neither feasible nor advisable given the FDA’s resource constraints and the technology’s therapeutic potential. Instead, they propose a three-part framework to bridge the gap between innovation and oversight. First, health systems should establish a public registry disclosing which clinical decision support tools are in use and how they were validated, without triggering immediate regulatory penalties. Second, the FDA should implement structured, nonbinding consultation pathways to foster proactive dialogue between regulators, developers, and clinicians. Third, the agency must refine its guidance to resolve interpretive disputes and align standards with current AI capabilities. As artificial intelligence continues to integrate into electronic health records, the researchers argue that transparency must precede unrestricted deployment. The proposed measures aim to ensure that tools influencing patient outcomes are rigorously evaluated, particularly for historically underserved populations. By balancing accountability with innovation, the medical technology sector can advance clinical decision-making while safeguarding patient trust. The evolving regulatory landscape will ultimately determine how AI is deployed, validated, and trusted in American healthcare for years to come.