Margaret Lozovatsky, MD, FAMIA, has spent years helping health systems adopt digital tools designed to improve patient care. As a pediatric hospitalist, she has also seen what can go wrong. In one case, a predictive artificial intelligence (AI) model flagged a healthy infant as having an 89% mortality risk — not because the child was critically ill, but because the algorithm had never been designed or tested for pediatric patients.

Failures like this have raised alarms as AI becomes more deeply embedded in clinical workflows. High-profile examples, from biased diagnostic models to systems that misinterpret imaging data, have underscored concerns about how reliably these tools perform once they move into everyday care.

“AI offers significant opportunities in healthcare,” said Lozovatsky, vice president of Digital Health Innovations at the American Medical Association (AMA) since early 2024. “As these tools roll out across care settings, we need a thoughtful approach to how AI is integrated into workflows. Standardized evaluation, monitoring, and safety protocols are essential to ensure that rapid implementation does not put patients or clinicians at risk.”

AI is already reshaping how physicians work, from analyzing medical images and flagging early signs of stroke or sepsis to easing documentation burdens and supporting clinical decision-making. Between 2023 and 2024, physician use of AI tools rose sharply, from 38% to 66%, according to AMA survey data. As of January 2026, the FDA has authorized 1300 AI-enabled medical devices, up from fewer than 10 approvals in 2013. Most are concentrated in image analysis, with radiology accounting for the largest share. 

The Legal Bind

The rapid adoption of AI in clinical care has pushed questions about accountability to the forefront. When an AI-informed decision contributes to patient harm, responsibility can be difficult to parse: Does liability fall on the physician, the health system, the technology vendor, or some combination of all three?

So far, the legal system has offered limited guidance. As of early 2026, there have been no US malpractice jury verdicts in which AI itself was the central basis of a claim. Instead, courts have largely relied on traditional malpractice principles to evaluate cases involving algorithmic tools.

“Under existing malpractice law, the physician is still the focal point of liability,” said Christopher Robertson, JD, professor of law and health policy at Boston University School of Law, Boston. “Courts are likely to focus on whether the clinician acted reasonably in relying on the system, questioning it, or overriding it.”

That framework reflects the enduring role of the standard of care, the legal benchmark used to assess negligence. Even when AI plays a meaningful role in patient care, said Stacey Lee, JD, professor at the Johns Hopkins Carey Business School with a joint appointment at the Bloomberg School of Public Health in Baltimore, there is no doctrine that assigns shared legal responsibility to the technology itself.

“In malpractice law, the central question hasn’t changed,” Lee said. “The question is whether the provider acted as a reasonably prudent provider. Saying ‘the AI got it wrong’ generally isn’t a defense.”

This creates a double bind for clinicians. They may be criticized for relying too heavily on flawed AI output yet also questioned for “failing” to consider tools shown to reduce diagnostic error and improve patient safety.

“We’re in a Wild West moment,” Robertson said. “If AI tools become widely adopted and reasonably relied upon, courts may eventually view their use, or at least consideration, as part of competent care. Right now, physicians are being asked to use these tools without clear structures to support that use.”

Physicians in the Middle

Adam Rodman, MD, MPH, a practicing internist and AI researcher at Beth Israel Deaconess Medical Center, Boston, approaches AI from a different angle. While malpractice risk looms large in legal debates, Rodman is more focused on how the technology is already reshaping clinical encounters.

In his practice, AI enters the conversation before the visit even begins. With many patients consulting chatbots in advance, they often arrive with AI-generated interpretations of symptoms, lab results, or imaging. Rather than discourage the behavior, Rodman engages with it directly.

“When patients show me their ChatGPT transcripts, it actually strengthens the doctor-patient relationship,” he said. “We talk through what the AI got right and what it got wrong,” turning the technology into a catalyst for a “three-way interaction” between clinician, patient, and AI.

Rodman is cautiously optimistic about AI’s potential to support clinical reasoning. In a 2024 JAMA Network Open study he led, ChatGPT-4 outperformed both residents and attending physicians on simulated diagnostic and reasoning tasks. Yet clinicians given access to the chatbot performed only slightly better than those who were not, suggesting that simply introducing AI does not automatically improve decision-making.

The findings highlight a paradox. Experienced clinicians tend to use decision-support tools conservatively, while trainees may lean on them too heavily.

Rodman’s biggest concern is “deskilling” — the risk that physicians-in-training may rely on AI before developing core clinical reasoning skills. As a co-director of the Digital Education Track for the internal medicine residency at Harvard Medical School in Boston, he sees near-universal use of AI reference tools among residents, many of whom rely on Health Insurance Portability and Accountability Act-compliant platforms such as OpenEvidence.

“If you rely on AI to generate differential diagnoses before you’ve learned how to think through them yourself, you miss out on an essential part of training,” Rodman said.

Over time, that reliance can affect how clinicians articulate their clinical reasoning when decisions are later reviewed. To maximize safety, he advocates for training models that treat AI as a “human-on-the-loop” assistant, allowing technology to process complex data while physicians retain oversight and responsibility.

“Discomfort and supervised mistakes are how you learn to be a doctor,” he said. “AI should support that process, not replace it.”

Tyler Berzin, MD, gastroenterologist at Beth Israel Deaconess Medical Center and associate professor at Harvard Medical School, studies and uses AI-assisted polyp detection during live colonoscopy, an area where automation has shown measurable benefit without displacing physician judgment.

Computer-vision systems flag subtle lesions in real time by highlighting suspicious areas on the screen, while the physician retains full control over interpretation and removal.

“Polyps can be subtle,” Berzin said. “You’re navigating the scope, monitoring the patient, and interpreting what you’re seeing all at once.”

Studies suggest these systems boost detection rates by roughly 25%, offering meaningful gains in colorectal cancer prevention. Because missed polyps are a common source of malpractice claims in gastroenterology, Berzin views AI-assisted detection as a safety aid rather than a liability risk.

He describes the technology as “level-one automation” — akin to a lane-departure warning in a car — providing alerts that clinicians may accept or ignore.

Professional societies have not yet mandated these tools, Berzin said, in part to avoid disadvantaging physicians who cannot access them. But if guidelines eventually incorporate AI-assisted detection, liability expectations could shift.

“When guidelines change, the liability conversation will change,” Berzin said. “At that point, not using a tool that improves detection could become harder to defend.”

Governance and Guardrails

While courts continue to address AI liability case by case, regulatory activity is accelerating elsewhere, most notably at the state level. Over the past 2 years, dozens of states have introduced or passed healthcare-specific AI laws, many taking effect in 2026. These measures generally require disclosure when AI is used in patient care and reaffirm that final clinical decision-making authority rests with licensed professionals.

That state-by-state approach — now facing legal challenges and renewed federal efforts to limit or preempt state regulation — has created significant uncertainty for clinicians and health systems operating across jurisdictions.

“The result is a growing patchwork of rules that leaves physicians and health systems to interpret compliance largely on their own,” said Lee, who advises health systems on legal risk and liability related to AI use in clinical care.

In that gap, professional organizations have stepped in. The AMA deliberately uses the term “augmented intelligence” to emphasize that AI tools are intended to support — not replace — clinical reasoning.

“Clinicians still need to understand what a tool is doing, where it performs well, and where it can fail,” said the AMA’s Lozovatsky. “That requires evaluation, training, and ongoing oversight, not blind trust.”

In 2025, the AMA released its “Governance for Augmented Intelligence” toolkit, which outlines practical steps for evaluating AI tools before deployment, defining appropriate use cases, training clinicians, and establishing monitoring and feedback processes once systems are live.

Lozovatsky stressed that innovation must advance alongside safeguards that anticipate human error and support clinicians when uncertainty is highest. That includes systems designed to flag unreliable inputs, prompt second looks, and surface uncertainty rather than mask it behind confident-sounding outputs.

Trust, she said, remains the central challenge. AMA surveys consistently show that clinicians’ top concerns about AI involve bias, safety, and the potential impact on patient trust. Without clear ways to evaluate performance, document appropriate use, and obtain informed consent, many physicians remain cautious about adopting tools that could otherwise improve care.

“These tools are expensive, complex, and deeply intertwined with workflow,” Lozovatsky said. “If you don’t do the work up front — local testing, clinician training, ongoing monitoring — you can’t deploy AI safely at the speed it’s arriving.”

Calls are growing for a shared accountability model that allocates responsibility across developers, health systems, and clinicians according to control over an AI tool’s lifecycle — a shift that legal experts say would likely require federal regulation.

“We don’t yet have all the legal answers,” Lozovatsky said. “But we do know this: governance is what determines whether AI becomes an added layer of safety — or a new source of risk.”