{"id":44640,"topic":"ai","source":"Penn LDI","title":"How to Regulate AI Without Stifling Health Innovation - Penn LDI","url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","url_hash":"d60d7e877b93fbb3c47d95e7d12e6b317c806d2a","author":"","summary":"How to Regulate AI Without Stifling Health Innovation Penn LDI","content":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece.\nIn a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach. They argue that the licensing model is particularly important for clinical AI systems that set their own goals, create step-by-step action plans, and execute tasks without human supervision. They also say that easing workforce shortages requires policies that properly support the implementation of agentic AI.\nBoth Bressman and Bergman’s team say this licensing model would require new federal capabilities. Bergman’s piece calls on Congress to establish a national licensing and oversight entity — the Office of Clinical AI Oversight — within the U.S. Department of Health and Human Services.\nLast month, the White House laid out a new federal approach to AI in an executive order, suggesting that the federal government should have prerelease voluntary oversight over some of the most powerful models if they are determined to be “covered frontier models” that affect national security.\nAt least one bill introduced in Congress proposes licensing AI to prescribe medications. At the same time, states are considering more than 250 bills that affect the technology. Bills that have passed range from Utah’s new law, which allows AI prescribing of some drugs, to California’s prohibition on using AI to deny insurance coverage.\nWe asked Bressman and Bergman about the rapidly changing AI environment and their recommendations for what’s next.\nDr. Bergman, can you cite an example to show how this AI licensing policy could work?\nBergman: Take a specific AI model providing primary care in a rural federally qualified health center, the kind of place that has lost most or all of its physicians. Under this proposal, the AI first has to pass standardized competency exams, scoring at or above the median of recent human test-takers on all three U.S. Medical Licensing Examination (USMLE) steps plus any relevant board exams. Then it enters a supervised deployment period in the clinic, like the residency physicians complete, where it has to show adequate performance on a set patient volume with cases reviewed by human clinicians.\nOnly then is it licensed for a defined scope in that setting, so it might take histories and recommend next steps, but not prescribe without a clinician. The license is time-limited, renewed annually or biennially, and subject to ongoing monitoring. For the patient, this looks less like a chatbot and more like a medical visit. Someone who today waits weeks or drives hours could sit down for an online visit with a licensed AI agent that takes a full history and works through the problem.\nRoutine cases might resolve on the spot. When something falls outside the AI’s scope or looks serious, it hands the case off to a human.\nHow can agentic AI be a solution for health care workforce shortages?\nBergman: There are two potential pathways. The first is productivity augmentation: you use AI to stretch the clinicians you already have. That is real, but it is capped by clinician supply. Pushing high volumes of patients can also worsen burnout, which is already a pressing clinician workforce problem. This pathway cannot close a shortfall of tens of thousands of physicians.\nThe second pathway is autonomous substitution. Agentic systems handle a defined slice of care with little or no human supervision. That is the only path that actually expands capacity. Recent trials show that the core cognitive work of primary care now overlaps with what these systems do well.\nThe White House’s Executive Order in December tried to preempt state activity on health AI, presumably to avoid the fragmented environment you describe in your piece. Why wasn’t that enough?\nBergman: An executive order cannot preempt state law on its own. Under the Supremacy Clause, that power belongs to Congress. This is why our proposal calls for a statute, not an executive order. Preemption is fine, but only when it comes with a real federal standard.\nWhat are the implications for health AI from the June executive order, which signals a new pre-release federal review process for certain AI models?\nBressman: The executive order signals something of a shift in the administration’s posture toward AI, which had previously emphasized deregulation in support of innovation in most of its policy. How it will affect clinical AI remains to be seen. The hope is that there will be a similar shift toward a more deliberate regulatory framework, particularly for generative and agentic AI.\nBergman: The order seems to be primarily a national security instrument, and it is not clear what its effects would be on the regulation of AI outside cybersecurity risks, especially in medicine. The review process itself sits at the wrong level for autonomous clinical AI. Review is applied at the model layer, but clinical risk lives at the deployment layer.\nWhat recommendations do you have for Congress in the months ahead?\nBressman: Federal inaction in this space has led to an emerging patchwork of state-level policies, and as long as the Food and Drug Administration (FDA) and Congress stay silent, this fragmentation will only grow. This will serve no one well, not patients, providers, or developers. A well-considered, centralized approach, whether housed within the FDA’s existing infrastructure or through a novel Office of Clinical AI Oversight, is the only way to get ahead of this.\nBergman: Legislate the framework. Congress should create, by statute, an Office of Clinical AI Oversight within the U.S. Department of Health and Human Services (HHS). Give it authority to certify autonomous clinical AI against a real performance standard. Fund it through developer fees, like the FDA model. Authorize a multistakeholder process to set competency benchmarks, drawing on the Accreditation Council for Graduate Medical Education (ACGME), The Joint Commission, and specialty boards. Make federal certification binding on the states. Build clear developer pathways with presubmission consultation, so oversight does not block innovation.\nWhat advice would you give state policymakers right now?\nBergman: I would steer states away from certifying clinical competence themselves. Focus instead on scope of practice, supervision, enforcement, transparency, and coverage rules. California’s ban on using AI to deny coverage is a good example. States should keep pressing Congress. The durable fix is a federal competency floor that state rules can sit atop.\nBressman: I understand the desire to move faster than the speed of Congress here, and the regulatory sandboxes that have emerged are one approach. To the extent that states can align on standards as they build out these sandboxes and consider more durable policy, it will help mitigate the kind of patchwork we would like to avoid. More importantly, they should not let the desire to innovate get ahead of their primary role: ensuring safe, accountable care for patients. As some states experiment with the idea of licensing AI for practice, they should outline clear standards rather than letting model developers set the terms.\nYour licensing frameworks include a role for health systems. What should they focus on for 2026 and 2027?\nBergman: Health systems should not wait for the federal framework. They should build AI governance structures now. Under our proposal, the deploying institution carries real responsibility. I would focus on three things.\nFirst, stand up a real AI oversight function. Govern how tools are acquired, validated, deployed, monitored, and, importantly, retired. Second, treat local validation as nonnegotiable. The lesson of the Epic Sepsis Model is that positive performance reports from the vendor do not guarantee good performance in your patients. Systems need to measure outcomes on their own populations. Third, design and implement triage, escalation, and adverse event reporting processes.\nBressman: I agree that health systems need to continue developing institutional governance frameworks, particularly as they integrate generative and agentic AI into clinical workflows. If models were to be licensed to practice in some form, what would it mean to credential these tools locally? Academic health systems should also be leading the way in generating evidence around the impact of these tools. That includes clinical outcomes, cost, and access to care, to be sure, but also the trade-offs and potential unintended consequences that may be harder to capture – risks like deskilling or never-skilling of clinicians, automation bias, and model sycophancy, to name a few.\n“A Licensure Framework for Autonomous Clinical AI” by Alon Bergman, Robert M. Wachter, and Ezekiel J. Emanuel was published in JAMA on April 29, 2026.\nAuthor\nMore on AI in Health Care\nIn a First-of-Its-Kind Analysis, Researchers Track Industry Payments Tied to AI Medical Devices and Call for Stronger FDA Oversight, Transparency, and Aid for Underserved Hospitals","image_url":"https://d197nivf0nbma8.cloudfront.net/uploads/2026/07/ai-reg-qa.png","lang":"en","published_at":"2026-07-23T17:04:45+00:00","fetched_at":"2026-07-23T17:15:05+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece. In a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach.","cluster_id":null,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9230 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9230,"summary_length":392,"usable_text_length":9230,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9230,"summary_length":392}},"news_item":{"id":44640,"canonical_url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","source_url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","title":"How to Regulate AI Without Stifling Health Innovation - Penn LDI","source_name":"Penn LDI","author":null,"published_at":"2026-07-23T17:04:45+00:00","locale":"en","topic":"ai","tags":[],"rss_summary":"How to Regulate AI Without Stifling Health Innovation Penn LDI","full_text":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece.\nIn a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach. They argue that the licensing model is particularly important for clinical AI systems that set their own goals, create step-by-step action plans, and execute tasks without human supervision. They also say that easing workforce shortages requires policies that properly support the implementation of agentic AI.\nBoth Bressman and Bergman’s team say this licensing model would require new federal capabilities. Bergman’s piece calls on Congress to establish a national licensing and oversight entity — the Office of Clinical AI Oversight — within the U.S. Department of Health and Human Services.\nLast month, the White House laid out a new federal approach to AI in an executive order, suggesting that the federal government should have prerelease voluntary oversight over some of the most powerful models if they are determined to be “covered frontier models” that affect national security.\nAt least one bill introduced in Congress proposes licensing AI to prescribe medications. At the same time, states are considering more than 250 bills that affect the technology. Bills that have passed range from Utah’s new law, which allows AI prescribing of some drugs, to California’s prohibition on using AI to deny insurance coverage.\nWe asked Bressman and Bergman about the rapidly changing AI environment and their recommendations for what’s next.\nDr. Bergman, can you cite an example to show how this AI licensing policy could work?\nBergman: Take a specific AI model providing primary care in a rural federally qualified health center, the kind of place that has lost most or all of its physicians. Under this proposal, the AI first has to pass standardized competency exams, scoring at or above the median of recent human test-takers on all three U.S. Medical Licensing Examination (USMLE) steps plus any relevant board exams. Then it enters a supervised deployment period in the clinic, like the residency physicians complete, where it has to show adequate performance on a set patient volume with cases reviewed by human clinicians.\nOnly then is it licensed for a defined scope in that setting, so it might take histories and recommend next steps, but not prescribe without a clinician. The license is time-limited, renewed annually or biennially, and subject to ongoing monitoring. For the patient, this looks less like a chatbot and more like a medical visit. Someone who today waits weeks or drives hours could sit down for an online visit with a licensed AI agent that takes a full history and works through the problem.\nRoutine cases might resolve on the spot. When something falls outside the AI’s scope or looks serious, it hands the case off to a human.\nHow can agentic AI be a solution for health care workforce shortages?\nBergman: There are two potential pathways. The first is productivity augmentation: you use AI to stretch the clinicians you already have. That is real, but it is capped by clinician supply. Pushing high volumes of patients can also worsen burnout, which is already a pressing clinician workforce problem. This pathway cannot close a shortfall of tens of thousands of physicians.\nThe second pathway is autonomous substitution. Agentic systems handle a defined slice of care with little or no human supervision. That is the only path that actually expands capacity. Recent trials show that the core cognitive work of primary care now overlaps with what these systems do well.\nThe White House’s Executive Order in December tried to preempt state activity on health AI, presumably to avoid the fragmented environment you describe in your piece. Why wasn’t that enough?\nBergman: An executive order cannot preempt state law on its own. Under the Supremacy Clause, that power belongs to Congress. This is why our proposal calls for a statute, not an executive order. Preemption is fine, but only when it comes with a real federal standard.\nWhat are the implications for health AI from the June executive order, which signals a new pre-release federal review process for certain AI models?\nBressman: The executive order signals something of a shift in the administration’s posture toward AI, which had previously emphasized deregulation in support of innovation in most of its policy. How it will affect clinical AI remains to be seen. The hope is that there will be a similar shift toward a more deliberate regulatory framework, particularly for generative and agentic AI.\nBergman: The order seems to be primarily a national security instrument, and it is not clear what its effects would be on the regulation of AI outside cybersecurity risks, especially in medicine. The review process itself sits at the wrong level for autonomous clinical AI. Review is applied at the model layer, but clinical risk lives at the deployment layer.\nWhat recommendations do you have for Congress in the months ahead?\nBressman: Federal inaction in this space has led to an emerging patchwork of state-level policies, and as long as the Food and Drug Administration (FDA) and Congress stay silent, this fragmentation will only grow. This will serve no one well, not patients, providers, or developers. A well-considered, centralized approach, whether housed within the FDA’s existing infrastructure or through a novel Office of Clinical AI Oversight, is the only way to get ahead of this.\nBergman: Legislate the framework. Congress should create, by statute, an Office of Clinical AI Oversight within the U.S. Department of Health and Human Services (HHS). Give it authority to certify autonomous clinical AI against a real performance standard. Fund it through developer fees, like the FDA model. Authorize a multistakeholder process to set competency benchmarks, drawing on the Accreditation Council for Graduate Medical Education (ACGME), The Joint Commission, and specialty boards. Make federal certification binding on the states. Build clear developer pathways with presubmission consultation, so oversight does not block innovation.\nWhat advice would you give state policymakers right now?\nBergman: I would steer states away from certifying clinical competence themselves. Focus instead on scope of practice, supervision, enforcement, transparency, and coverage rules. California’s ban on using AI to deny coverage is a good example. States should keep pressing Congress. The durable fix is a federal competency floor that state rules can sit atop.\nBressman: I understand the desire to move faster than the speed of Congress here, and the regulatory sandboxes that have emerged are one approach. To the extent that states can align on standards as they build out these sandboxes and consider more durable policy, it will help mitigate the kind of patchwork we would like to avoid. More importantly, they should not let the desire to innovate get ahead of their primary role: ensuring safe, accountable care for patients. As some states experiment with the idea of licensing AI for practice, they should outline clear standards rather than letting model developers set the terms.\nYour licensing frameworks include a role for health systems. What should they focus on for 2026 and 2027?\nBergman: Health systems should not wait for the federal framework. They should build AI governance structures now. Under our proposal, the deploying institution carries real responsibility. I would focus on three things.\nFirst, stand up a real AI oversight function. Govern how tools are acquired, validated, deployed, monitored, and, importantly, retired. Second, treat local validation as nonnegotiable. The lesson of the Epic Sepsis Model is that positive performance reports from the vendor do not guarantee good performance in your patients. Systems need to measure outcomes on their own populations. Third, design and implement triage, escalation, and adverse event reporting processes.\nBressman: I agree that health systems need to continue developing institutional governance frameworks, particularly as they integrate generative and agentic AI into clinical workflows. If models were to be licensed to practice in some form, what would it mean to credential these tools locally? Academic health systems should also be leading the way in generating evidence around the impact of these tools. That includes clinical outcomes, cost, and access to care, to be sure, but also the trade-offs and potential unintended consequences that may be harder to capture – risks like deskilling or never-skilling of clinicians, automation bias, and model sycophancy, to name a few.\n“A Licensure Framework for Autonomous Clinical AI” by Alon Bergman, Robert M. Wachter, and Ezekiel J. Emanuel was published in JAMA on April 29, 2026.\nAuthor\nMore on AI in Health Care\nIn a First-of-Its-Kind Analysis, Researchers Track Industry Payments Tied to AI Medical Devices and Call for Stronger FDA Oversight, Transparency, and Aid for Underserved Hospitals","excerpt":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece. In a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach.","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 9230 characters.","diagnostics_url":"/api/diagnose?url=https%3A//ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9230 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9230,"summary_length":392,"usable_text_length":9230,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9230,"summary_length":392}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"How to Regulate AI Without Stifling Health Innovation - Penn LDI","url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","summary":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece. In a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach.","source":"Penn LDI","date":"2026-07-23T17:04:45+00:00","content":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece.\nIn a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach. They argue that the licensing model is particularly important for clinical AI systems that set their own goals, create step-by-step action plans, and execute tasks without human supervision. They also say that easing workforce shortages requires policies that properly support the implementation of agentic AI.\nBoth Bressman and Bergman’s team say this licensing model would require new federal capabilities. Bergman’s piece calls on Congress to establish a national licensing and oversight entity — the Office of Clinical AI Oversight — within the U.S. Department of Health and Human Services.\nLast month, the White House laid out a new federal approach to AI in an executive order, suggesting that the federal government should have prerelease voluntary oversight over some of the most powerful models if they are determined to be “covered frontier models” that affect national security.\nAt least one bill introduced in Congress proposes licensing AI to prescribe medications. At the same time, states are considering more than 250 bills that affect the technology. Bills that have passed range from Utah’s new law, which allows AI prescribing of some drugs, to California’s prohibition on using AI to deny insurance coverage.\nWe asked Bressman and Bergman about the rapidly changing AI environment and their recommendations for what’s next.\nDr. Bergman, can you cite an example to show how this AI licensing policy could work?\nBergman: Take a specific AI model providing primary care in a rural federally qualified health center, the kind of place that has lost most or all of its physicians. Under this proposal, the AI first has to pass standardized competency exams, scoring at or above the median of recent human test-takers on all three U.S. Medical Licensing Examination (USMLE) steps plus any relevant board exams. Then it enters a supervised deployment period in the clinic, like the residency physicians complete, where it has to show adequate performance on a set patient volume with cases reviewed by human clinicians.\nOnly then is it licensed for a defined scope in that setting, so it might take histories and recommend next steps, but not prescribe without a clinician. The license is time-limited, renewed annually or biennially, and subject to ongoing monitoring. For the patient, this looks less like a chatbot and more like a medical visit. Someone who today waits weeks or drives hours could sit down for an online visit with a licensed AI agent that takes a full history and works through the problem.\nRoutine cases might resolve on the spot. When something falls outside the AI’s scope or looks serious, it hands the case off to a human.\nHow can agentic AI be a solution for health care workforce shortages?\nBergman: There are two potential pathways. The first is productivity augmentation: you use AI to stretch the clinicians you already have. That is real, but it is capped by clinician supply. Pushing high volumes of patients can also worsen burnout, which is already a pressing clinician workforce problem. This pathway cannot close a shortfall of tens of thousands of physicians.\nThe second pathway is autonomous substitution. Agentic systems handle a defined slice of care with little or no human supervision. That is the only path that actually expands capacity. Recent trials show that the core cognitive work of primary care now overlaps with what these systems do well.\nThe White House’s Executive Order in December tried to preempt state activity on health AI, presumably to avoid the fragmented environment you describe in your piece. Why wasn’t that enough?\nBergman: An executive order cannot preempt state law on its own. Under the Supremacy Clause, that power belongs to Congress. This is why our proposal calls for a statute, not an executive order. Preemption is fine, but only when it comes with a real federal standard.\nWhat are the implications for health AI from the June executive order, which signals a new pre-release federal review process for certain AI models?\nBressman: The executive order signals something of a shift in the administration’s posture toward AI, which had previously emphasized deregulation in support of innovation in most of its policy. How it will affect clinical AI remains to be seen. The hope is that there will be a similar shift toward a more deliberate regulatory framework, particularly for generative and agentic AI.\nBergman: The order seems to be primarily a national security instrument, and it is not clear what its effects would be on the regulation of AI outside cybersecurity risks, especially in medicine. The review process itself sits at the wrong level for autonomous clinical AI. Review is applied at the model layer, but clinical risk lives at the deployment layer.\nWhat recommendations do you have for Congress in the months ahead?\nBressman: Federal inaction in this space has led to an emerging patchwork of state-level policies, and as long as the Food and Drug Administration (FDA) and Congress stay silent, this fragmentation will only grow. This will serve no one well, not patients, providers, or developers. A well-considered, centralized approach, whether housed within the FDA’s existing infrastructure or through a novel Office of Clinical AI Oversight, is the only way to get ahead of this.\nBergman: Legislate the framework. Congress should create, by statute, an Office of Clinical AI Oversight within the U.S. Department of Health and Human Services (HHS). Give it authority to certify autonomous clinical AI against a real performance standard. Fund it through developer fees, like the FDA model. Authorize a multistakeholder process to set competency benchmarks, drawing on the Accreditation Council for Graduate Medical Education (ACGME), The Joint Commission, and specialty boards. Make federal certification binding on the states. Build clear developer pathways with presubmission consultation, so oversight does not block innovation.\nWhat advice would you give state policymakers right now?\nBergman: I would steer states away from certifying clinical competence themselves. Focus instead on scope of practice, supervision, enforcement, transparency, and coverage rules. California’s ban on using AI to deny coverage is a good example. States should keep pressing Congress. The durable fix is a federal competency floor that state rules can sit atop.\nBressman: I understand the desire to move faster than the speed of Congress here, and the regulatory sandboxes that have emerged are one approach. To the extent that states can align on standards as they build out these sandboxes and consider more durable policy, it will help mitigate the kind of patchwork we would like to avoid. More importantly, they should not let the desire to innovate get ahead of their primary role: ensuring safe, accountable care for patients. As some states experiment with the idea of licensing AI for practice, they should outline clear standards rather than letting model developers set the terms.\nYour licensing frameworks include a role for health systems. What should they focus on for 2026 and 2027?\nBergman: Health systems should not wait for the federal framework. They should build AI governance structures now. Under our proposal, the deploying institution carries real responsibility. I would focus on three things.\nFirst, stand up a real AI oversight function. Govern how tools are acquired, validated, deployed, monitored, and, importantly, retired. Second, treat local validation as nonnegotiable. The lesson of the Epic Sepsis Model is that positive performance reports from the vendor do not guarantee good performance in your patients. Systems need to measure outcomes on their own populations. Third, design and implement triage, escalation, and adverse event reporting processes.\nBressman: I agree that health systems need to continue developing institutional governance frameworks, particularly as they integrate generative and agentic AI into clinical workflows. If models were to be licensed to practice in some form, what would it mean to credential these tools locally? Academic health systems should also be leading the way in generating evidence around the impact of these tools. That includes clinical outcomes, cost, and access to care, to be sure, but also the trade-offs and potential unintended consequences that may be harder to capture – risks like deskilling or never-skilling of clinicians, automation bias, and model sycophancy, to name a few.\n“A Licensure Framework for Autonomous Clinical AI” by Alon Bergman, Robert M. Wachter, and Ezekiel J. Emanuel was published in JAMA on April 29, 2026.\nAuthor\nMore on AI in Health Care\nIn a First-of-Its-Kind Analysis, Researchers Track Industry Payments Tied to AI Medical Devices and Call for Stronger FDA Oversight, Transparency, and Aid for Underserved Hospitals","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 9230 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9230 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9230,"summary_length":392,"usable_text_length":9230,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9230,"summary_length":392}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/44640","export_markdown":"/api/items/44640/export?format=markdown","export_json":"/api/items/44640/export?format=json","diagnose":"/api/diagnose?url=https%3A//ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/"},"formats":{"full":{"id":44640,"title":"How to Regulate AI Without Stifling Health Innovation - Penn LDI","url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","source":"Penn LDI","author":null,"published_at":"2026-07-23T17:04:45+00:00","locale":"en","topic":"ai","tags":[],"excerpt":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece. In a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach.","full_text":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece.\nIn a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach. They argue that the licensing model is particularly important for clinical AI systems that set their own goals, create step-by-step action plans, and execute tasks without human supervision. They also say that easing workforce shortages requires policies that properly support the implementation of agentic AI.\nBoth Bressman and Bergman’s team say this licensing model would require new federal capabilities. Bergman’s piece calls on Congress to establish a national licensing and oversight entity — the Office of Clinical AI Oversight — within the U.S. Department of Health and Human Services.\nLast month, the White House laid out a new federal approach to AI in an executive order, suggesting that the federal government should have prerelease voluntary oversight over some of the most powerful models if they are determined to be “covered frontier models” that affect national security.\nAt least one bill introduced in Congress proposes licensing AI to prescribe medications. At the same time, states are considering more than 250 bills that affect the technology. Bills that have passed range from Utah’s new law, which allows AI prescribing of some drugs, to California’s prohibition on using AI to deny insurance coverage.\nWe asked Bressman and Bergman about the rapidly changing AI environment and their recommendations for what’s next.\nDr. Bergman, can you cite an example to show how this AI licensing policy could work?\nBergman: Take a specific AI model providing primary care in a rural federally qualified health center, the kind of place that has lost most or all of its physicians. Under this proposal, the AI first has to pass standardized competency exams, scoring at or above the median of recent human test-takers on all three U.S. Medical Licensing Examination (USMLE) steps plus any relevant board exams. Then it enters a supervised deployment period in the clinic, like the residency physicians complete, where it has to show adequate performance on a set patient volume with cases reviewed by human clinicians.\nOnly then is it licensed for a defined scope in that setting, so it might take histories and recommend next steps, but not prescribe without a clinician. The license is time-limited, renewed annually or biennially, and subject to ongoing monitoring. For the patient, this looks less like a chatbot and more like a medical visit. Someone who today waits weeks or drives hours could sit down for an online visit with a licensed AI agent that takes a full history and works through the problem.\nRoutine cases might resolve on the spot. When something falls outside the AI’s scope or looks serious, it hands the case off to a human.\nHow can agentic AI be a solution for health care workforce shortages?\nBergman: There are two potential pathways. The first is productivity augmentation: you use AI to stretch the clinicians you already have. That is real, but it is capped by clinician supply. Pushing high volumes of patients can also worsen burnout, which is already a pressing clinician workforce problem. This pathway cannot close a shortfall of tens of thousands of physicians.\nThe second pathway is autonomous substitution. Agentic systems handle a defined slice of care with little or no human supervision. That is the only path that actually expands capacity. Recent trials show that the core cognitive work of primary care now overlaps with what these systems do well.\nThe White House’s Executive Order in December tried to preempt state activity on health AI, presumably to avoid the fragmented environment you describe in your piece. Why wasn’t that enough?\nBergman: An executive order cannot preempt state law on its own. Under the Supremacy Clause, that power belongs to Congress. This is why our proposal calls for a statute, not an executive order. Preemption is fine, but only when it comes with a real federal standard.\nWhat are the implications for health AI from the June executive order, which signals a new pre-release federal review process for certain AI models?\nBressman: The executive order signals something of a shift in the administration’s posture toward AI, which had previously emphasized deregulation in support of innovation in most of its policy. How it will affect clinical AI remains to be seen. The hope is that there will be a similar shift toward a more deliberate regulatory framework, particularly for generative and agentic AI.\nBergman: The order seems to be primarily a national security instrument, and it is not clear what its effects would be on the regulation of AI outside cybersecurity risks, especially in medicine. The review process itself sits at the wrong level for autonomous clinical AI. Review is applied at the model layer, but clinical risk lives at the deployment layer.\nWhat recommendations do you have for Congress in the months ahead?\nBressman: Federal inaction in this space has led to an emerging patchwork of state-level policies, and as long as the Food and Drug Administration (FDA) and Congress stay silent, this fragmentation will only grow. This will serve no one well, not patients, providers, or developers. A well-considered, centralized approach, whether housed within the FDA’s existing infrastructure or through a novel Office of Clinical AI Oversight, is the only way to get ahead of this.\nBergman: Legislate the framework. Congress should create, by statute, an Office of Clinical AI Oversight within the U.S. Department of Health and Human Services (HHS). Give it authority to certify autonomous clinical AI against a real performance standard. Fund it through developer fees, like the FDA model. Authorize a multistakeholder process to set competency benchmarks, drawing on the Accreditation Council for Graduate Medical Education (ACGME), The Joint Commission, and specialty boards. Make federal certification binding on the states. Build clear developer pathways with presubmission consultation, so oversight does not block innovation.\nWhat advice would you give state policymakers right now?\nBergman: I would steer states away from certifying clinical competence themselves. Focus instead on scope of practice, supervision, enforcement, transparency, and coverage rules. California’s ban on using AI to deny coverage is a good example. States should keep pressing Congress. The durable fix is a federal competency floor that state rules can sit atop.\nBressman: I understand the desire to move faster than the speed of Congress here, and the regulatory sandboxes that have emerged are one approach. To the extent that states can align on standards as they build out these sandboxes and consider more durable policy, it will help mitigate the kind of patchwork we would like to avoid. More importantly, they should not let the desire to innovate get ahead of their primary role: ensuring safe, accountable care for patients. As some states experiment with the idea of licensing AI for practice, they should outline clear standards rather than letting model developers set the terms.\nYour licensing frameworks include a role for health systems. What should they focus on for 2026 and 2027?\nBergman: Health systems should not wait for the federal framework. They should build AI governance structures now. Under our proposal, the deploying institution carries real responsibility. I would focus on three things.\nFirst, stand up a real AI oversight function. Govern how tools are acquired, validated, deployed, monitored, and, importantly, retired. Second, treat local validation as nonnegotiable. The lesson of the Epic Sepsis Model is that positive performance reports from the vendor do not guarantee good performance in your patients. Systems need to measure outcomes on their own populations. Third, design and implement triage, escalation, and adverse event reporting processes.\nBressman: I agree that health systems need to continue developing institutional governance frameworks, particularly as they integrate generative and agentic AI into clinical workflows. If models were to be licensed to practice in some form, what would it mean to credential these tools locally? Academic health systems should also be leading the way in generating evidence around the impact of these tools. That includes clinical outcomes, cost, and access to care, to be sure, but also the trade-offs and potential unintended consequences that may be harder to capture – risks like deskilling or never-skilling of clinicians, automation bias, and model sycophancy, to name a few.\n“A Licensure Framework for Autonomous Clinical AI” by Alon Bergman, Robert M. Wachter, and Ezekiel J. Emanuel was published in JAMA on April 29, 2026.\nAuthor\nMore on AI in Health Care\nIn a First-of-Its-Kind Analysis, Researchers Track Industry Payments Tied to AI Medical Devices and Call for Stronger FDA Oversight, Transparency, and Aid for Underserved Hospitals","reading_time_min":7,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 9230 characters.","diagnostics_url":"/api/diagnose?url=https%3A//ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9230 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9230,"summary_length":392,"usable_text_length":9230,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9230,"summary_length":392}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9230 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9230,"summary_length":392,"usable_text_length":9230,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9230,"summary_length":392}},"actions":{"read":"/item/44640","export_markdown":"/api/items/44640/export?format=markdown","export_json":"/api/items/44640/export?format=json","diagnose":"/api/diagnose?url=https%3A//ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/"}},"digest":{"id":44640,"title":"How to Regulate AI Without Stifling Health Innovation - Penn LDI","url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","source":"Penn LDI","topic":"ai","published_at":"2026-07-23T17:04:45+00:00","excerpt":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece. In a recent…","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 9230 characters.","reading_time_min":7,"cluster_id":null},"card":{"display_title":"How to Regulate AI Without Stifling Health Innovation - Penn LDI","subtitle":"Penn LDI · 2026-07-23","summary":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric…","badges":["quality:high"],"links":{"read":"/item/44640","original":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","diagnose":"/api/diagnose?url=https%3A//ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/"},"quality_warning":null},"export":{"title":"How to Regulate AI Without Stifling Health Innovation - Penn LDI","url":"https://ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","summary":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece. In a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach.","source":"Penn LDI","date":"2026-07-23T17:04:45+00:00","content":"The best way to regulate generative AI in health care is to treat it like a human clinician, who gets licensed and supervised, so argues LDI Senior Fellow and Co-Chair of the LDI AI in Medicine Working Group Eric Bressman in a 2025 JAMA Internal Medicine piece.\nIn a recent piece, also in JAMA, LDI Senior Fellow Alon Bergman and co-authors Zeke Emanuel and Bob Wachter endorse that approach. They argue that the licensing model is particularly important for clinical AI systems that set their own goals, create step-by-step action plans, and execute tasks without human supervision. They also say that easing workforce shortages requires policies that properly support the implementation of agentic AI.\nBoth Bressman and Bergman’s team say this licensing model would require new federal capabilities. Bergman’s piece calls on Congress to establish a national licensing and oversight entity — the Office of Clinical AI Oversight — within the U.S. Department of Health and Human Services.\nLast month, the White House laid out a new federal approach to AI in an executive order, suggesting that the federal government should have prerelease voluntary oversight over some of the most powerful models if they are determined to be “covered frontier models” that affect national security.\nAt least one bill introduced in Congress proposes licensing AI to prescribe medications. At the same time, states are considering more than 250 bills that affect the technology. Bills that have passed range from Utah’s new law, which allows AI prescribing of some drugs, to California’s prohibition on using AI to deny insurance coverage.\nWe asked Bressman and Bergman about the rapidly changing AI environment and their recommendations for what’s next.\nDr. Bergman, can you cite an example to show how this AI licensing policy could work?\nBergman: Take a specific AI model providing primary care in a rural federally qualified health center, the kind of place that has lost most or all of its physicians. Under this proposal, the AI first has to pass standardized competency exams, scoring at or above the median of recent human test-takers on all three U.S. Medical Licensing Examination (USMLE) steps plus any relevant board exams. Then it enters a supervised deployment period in the clinic, like the residency physicians complete, where it has to show adequate performance on a set patient volume with cases reviewed by human clinicians.\nOnly then is it licensed for a defined scope in that setting, so it might take histories and recommend next steps, but not prescribe without a clinician. The license is time-limited, renewed annually or biennially, and subject to ongoing monitoring. For the patient, this looks less like a chatbot and more like a medical visit. Someone who today waits weeks or drives hours could sit down for an online visit with a licensed AI agent that takes a full history and works through the problem.\nRoutine cases might resolve on the spot. When something falls outside the AI’s scope or looks serious, it hands the case off to a human.\nHow can agentic AI be a solution for health care workforce shortages?\nBergman: There are two potential pathways. The first is productivity augmentation: you use AI to stretch the clinicians you already have. That is real, but it is capped by clinician supply. Pushing high volumes of patients can also worsen burnout, which is already a pressing clinician workforce problem. This pathway cannot close a shortfall of tens of thousands of physicians.\nThe second pathway is autonomous substitution. Agentic systems handle a defined slice of care with little or no human supervision. That is the only path that actually expands capacity. Recent trials show that the core cognitive work of primary care now overlaps with what these systems do well.\nThe White House’s Executive Order in December tried to preempt state activity on health AI, presumably to avoid the fragmented environment you describe in your piece. Why wasn’t that enough?\nBergman: An executive order cannot preempt state law on its own. Under the Supremacy Clause, that power belongs to Congress. This is why our proposal calls for a statute, not an executive order. Preemption is fine, but only when it comes with a real federal standard.\nWhat are the implications for health AI from the June executive order, which signals a new pre-release federal review process for certain AI models?\nBressman: The executive order signals something of a shift in the administration’s posture toward AI, which had previously emphasized deregulation in support of innovation in most of its policy. How it will affect clinical AI remains to be seen. The hope is that there will be a similar shift toward a more deliberate regulatory framework, particularly for generative and agentic AI.\nBergman: The order seems to be primarily a national security instrument, and it is not clear what its effects would be on the regulation of AI outside cybersecurity risks, especially in medicine. The review process itself sits at the wrong level for autonomous clinical AI. Review is applied at the model layer, but clinical risk lives at the deployment layer.\nWhat recommendations do you have for Congress in the months ahead?\nBressman: Federal inaction in this space has led to an emerging patchwork of state-level policies, and as long as the Food and Drug Administration (FDA) and Congress stay silent, this fragmentation will only grow. This will serve no one well, not patients, providers, or developers. A well-considered, centralized approach, whether housed within the FDA’s existing infrastructure or through a novel Office of Clinical AI Oversight, is the only way to get ahead of this.\nBergman: Legislate the framework. Congress should create, by statute, an Office of Clinical AI Oversight within the U.S. Department of Health and Human Services (HHS). Give it authority to certify autonomous clinical AI against a real performance standard. Fund it through developer fees, like the FDA model. Authorize a multistakeholder process to set competency benchmarks, drawing on the Accreditation Council for Graduate Medical Education (ACGME), The Joint Commission, and specialty boards. Make federal certification binding on the states. Build clear developer pathways with presubmission consultation, so oversight does not block innovation.\nWhat advice would you give state policymakers right now?\nBergman: I would steer states away from certifying clinical competence themselves. Focus instead on scope of practice, supervision, enforcement, transparency, and coverage rules. California’s ban on using AI to deny coverage is a good example. States should keep pressing Congress. The durable fix is a federal competency floor that state rules can sit atop.\nBressman: I understand the desire to move faster than the speed of Congress here, and the regulatory sandboxes that have emerged are one approach. To the extent that states can align on standards as they build out these sandboxes and consider more durable policy, it will help mitigate the kind of patchwork we would like to avoid. More importantly, they should not let the desire to innovate get ahead of their primary role: ensuring safe, accountable care for patients. As some states experiment with the idea of licensing AI for practice, they should outline clear standards rather than letting model developers set the terms.\nYour licensing frameworks include a role for health systems. What should they focus on for 2026 and 2027?\nBergman: Health systems should not wait for the federal framework. They should build AI governance structures now. Under our proposal, the deploying institution carries real responsibility. I would focus on three things.\nFirst, stand up a real AI oversight function. Govern how tools are acquired, validated, deployed, monitored, and, importantly, retired. Second, treat local validation as nonnegotiable. The lesson of the Epic Sepsis Model is that positive performance reports from the vendor do not guarantee good performance in your patients. Systems need to measure outcomes on their own populations. Third, design and implement triage, escalation, and adverse event reporting processes.\nBressman: I agree that health systems need to continue developing institutional governance frameworks, particularly as they integrate generative and agentic AI into clinical workflows. If models were to be licensed to practice in some form, what would it mean to credential these tools locally? Academic health systems should also be leading the way in generating evidence around the impact of these tools. That includes clinical outcomes, cost, and access to care, to be sure, but also the trade-offs and potential unintended consequences that may be harder to capture – risks like deskilling or never-skilling of clinicians, automation bias, and model sycophancy, to name a few.\n“A Licensure Framework for Autonomous Clinical AI” by Alon Bergman, Robert M. Wachter, and Ezekiel J. Emanuel was published in JAMA on April 29, 2026.\nAuthor\nMore on AI in Health Care\nIn a First-of-Its-Kind Analysis, Researchers Track Industry Payments Tied to AI Medical Devices and Call for Stronger FDA Oversight, Transparency, and Aid for Underserved Hospitals","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//ldi.upenn.edu/our-work/research-updates/how-the-u-s-can-regulate-ai-without-stifling-health-care-innovation/","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 9230 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 9230 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":9230,"summary_length":392,"usable_text_length":9230,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":9230,"summary_length":392}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}