AI in Healthcare: Familiar Liabilities, New Causes

Key takeaways:

• AI is changing how familiar healthcare liabilities arise, not creating entirely new ones. Because AI is prediction-based rather than rule-based, traditional governance and safety approaches may no longer be sufficient.
• Recent litigation suggests AI may change how organizations investigate adverse events. Reconstructing AI's role in care, and the evidence needed to understand it, may become substantially harder.
• The liability insurance market is still defining its approach to AI. Healthcare organizations face growing uncertainty about what AI-related exposure remains insurable.
• Organizations best positioned for this environment treat AI adoption as a risk management discipline, not a technology implementation. Their AI governance has real authority, runtime controls, and evidence of what AI systems actually did when questions arise.

Healthcare spent decades redesigning systems to prevent human error. AI raises a new challenge: familiar healthcare liabilities may increasingly arise through technologies whose behavior is not solely human-directed.

AI is no longer a future consideration for healthcare organizations; it is already embedded in clinical documentation, diagnostic support, patient communication, and care management workflows across the industry. The governance and liability implications of that reality are only beginning to come into focus.

As healthcare enters the era of AI-enabled care, organizations are only now beginning to understand the liabilities associated with these technologies. Early experience suggests that healthcare’s biggest litigation risks remain remarkably stable: informed consent gaps, privacy and confidentiality violations, documentation gaps, and discrimination.

What is changing is not the litigation landscape itself, but the paradigm for preventing harm and understanding failure. Existing safety strategies and governance models were built to manage errors arising from humans and humans interacting with well-understood technology. AI tests whether those same approaches will remain sufficient as healthcare organizations increasingly rely on probabilistic systems whose behavior can vary across settings, populations, and workflows, and may not always be directly observable or fully explainable.

Several recent cases, drawn from active litigation, settled disputes, and emerging regulatory enforcement, illustrate how familiar liabilities are materializing through AI-enabled technologies may introduce new risks into the healthcare system. They help shine a light on the practical difficulties organizations may face when investigating AI-related events, including reconstructing what technology was used, what outputs were generated, what information was available to clinicians, and whether sufficient evidence exists to understand what role technology played in care when outcomes are questioned.

Insurance carriers have also been impacted, as a rapidly evolving liability coverage environment has driven insurers to reassess whether traditional insurance models adequately address AI related exposure. As some carriers narrow coverage, introduce exclusions, or withdraw from segments of the market altogether, healthcare organizations now growing uncertainty regarding what liabilities remain insurable and under what conditions.

Across every category of risk this article examines, one conclusion holds: the organizations best positioned to navigate this environment will be those that approach AI adoption not as a technology implementation but as a risk management discipline supported by governance structures with real authority, runtime controls that monitor AI behavior and enforce safeguards as care is delivered, and the evidence to establish what systems actually did when questions arise.

Informed Consent Gaps

Ambient AI systems that record, transmit, and process confidential clinical conversations through third-party infrastructure are generating class action litigation under legal theories whose application to this technology is still being tested. The central issue is whether patients understood that ambient AI would be used during their encounter and had a meaningful opportunity to consent. Two class action suits filed within months of each other, one against a health system serving over a million patients annually, one against two of California's largest health systems, highlight how plaintiffs' firms are beginning to apply consent and authorization frameworks to ambient AI clinical workflows at scale. Courts have not yet ruled on the core theories, some cases may settle, and legislative or regulatory responses could shift the landscape. Exposure in a class-action context scales with every recorded encounter.

Saucedo v. Sharp HealthCare (November 2025), filed in California state court, alleges that Sharp deployed an ambient AI clinical documentation platform that recorded entire clinical encounters through microphone-enabled clinician devices, transmitted audio to a third-party vendor’s cloud infrastructure, and generated draft clinical notes inserted into the EHR. The complaint further alleges that vendor personnel could access recordings and transcripts for quality assurance, troubleshooting, and model improvement.

The complaint alleges Sharp deployed the system without a standardized consent process, pre-visit notice, standardized scripts, or reliable opt-out and deletion workflows. Of particular significance, plaintiffs allege Sharp configured templated EHR language falsely documenting that patients had been advised of and consented to recording when no such advisement or consent allegedly occurred.

Plaintiffs rely on established California privacy and medical confidentiality statutes, including the California Invasion of Privacy Act (CIPA), which prohibits recording confidential communications without all-party consent, and the California Confidentiality of Medical Information Act (CMIA), which plaintiffs argue required patient authorization before physician-patient conversations and related medical information were disclosed to third-party vendors. The complaint also invokes California’s Unfair Competition Law (UCL), seeking disgorgement tied to operational and financial benefits allegedly derived from deployment, and pleads negligent misrepresentation based on allegedly false consent language embedded in the medical record.

The complaint also raises a practical operational issue. According to the plaintiffs, Sharp could not compel the deletion of encounter audio after patient objection because vendor retention practices controlled the process. Plaintiffs frame this not simply as an inconvenience, but as an ongoing failure to preserve confidentiality once concerns were raised. In addition to monetary damages, plaintiffs seek injunctive relief directed at how ambient AI is operationalized, including encounter-level consent, deletion capabilities, vendor oversight, staff training, and correction of allegedly inaccurate medical records. The complaint further suggests scrutiny may extend beyond the institution itself, naming managing agents, informatics personnel, and vendor-facing staff generically as defendants, signaling that workflow design, training materials, and system configuration decisions would be included in case discovery.

Washington v. Sutter Health and MemorialCare (April 2026), filed in federal court, against two of California’s largest health systems, broadens the theory of exposure. The complaint applies to the federal Electronic Communications Privacy Act (ECPA) to a proposed nationwide class of patients whose clinical conversations were allegedly recorded, transmitted, and processed through ambient AI systems without consent. California patients form a subclass asserting additional claims under CIPA, CMIA and UCL.

The allegations against Sutter largely mirror those in Sharp -- inadequate notice, lack of a standardized consent process, insufficient authorization for vendor disclosure, limited opt-out mechanisms, and ineffective deletion-on-demand capability. Plaintiffs similarly allege that defendants deployed the system to obtain operational and financial benefits, including reduced documentation burden and improved capture of billable services, without implementing the consent, authorization, and deletion safeguards plaintiffs contend were required. Unlike Sharp, the complaint adds a federal ECPA theory based on contemporaneous interception of physician-patient communications through third-party infrastructure but does not include the allegedly false EHR consent documentation that distinguishes Sharp. Correspondingly, its requested injunctive relief is narrower.

Taken together, these cases suggest litigation involving ambient AI may increasingly focus not only on patient harm, but on whether organizations designed defensible workflows for consent, documentation, vendor oversight, and response once concerns were raised.

These lawsuits suggest organizations should adapt longstanding informed consent and privacy frameworks to the operational realities of AI-enabled clinical recording and documentation. Organizations deploying ambient AI should be prepared to demonstrate a defensible consent architecture, one that includes encounter-level informed consent documented in the EHR, clear vendor disclosure and authorization workflows, opt-out and deletion-on-demand processes, vendor contract terms governing retention and secondary use, staff training, and governance over system configuration and workflow design.

The allegations in Sharp and Washington suggest courts will likely ask whether organizations built practical safeguards before deployment, and whether they can document how those safeguards were designed, implemented, and governed. These cases also signal that scrutiny may extend beyond the institution to those responsible for deployment of decisions, making workflow records, training materials, and system configuration evidence highly relevant to discovery.

Patient Privacy and Third-Party Tracking Tools

Healthcare organizations increasingly rely on vendor-supplied technologies embedded in websites, patient portals, appointment scheduling systems, and other patient-facing digital tools. Often implemented for analytics, marketing, or operational purposes, these technologies may collect patient-related information outside of traditional clinical encounters. A growing wave of privacy lawsuits involving website tracking technologies raises governance questions directly relevant to health systems deploying AI in care delivery.

A tracking pixel is a small, invisible piece of code embedded in a webpage or application that automatically transmits information about digital activity, searches performed, pages visited, buttons clicked, or forms submitted, to third parties such as Meta or Google. In healthcare settings, this may connect patient identity or status with health-related interests, appointments sought, or providers researched. A widely cited 2022 investigation by The Markup found Meta Pixel installed on one-third of the top 100 U.S. hospital websites and documented transmissions to Facebook that allegedly included patient names, appointment details, and medication information without patient knowledge or consent. Even after the Mass General Brigham settlement, pixels reportedly remained active on affiliated sites, suggesting legal awareness had not consistently translated into operational remediation.

Doe v. Partners Healthcare / Mass General Brigham, filed in 2019, was among the first major cases to expose this risk. Plaintiffs alleged MGB and 38 affiliated entities deployed tracking technologies across patient-facing websites and care-access tools without adequate consent, transmitting health-related search activity to third-party advertising ecosystems. Plaintiffs were existing MGB patients interacting with digital care tools, including MGB’s secure patient portal, reinforcing expectations of confidentiality. Brought entirely under Massachusetts state law, invasion of privacy, interception of wire communications, and breach of fiduciary duty, the case settled for $18.4 million without alleging a medical record breach or relying on a traditional HIPAA PHI disclosure theory. Instead, plaintiffs relied heavily on MGB’s own privacy and confidentiality assurances, illustrating that healthcare organizations may face substantial litigation exposure when patient-facing technologies appear inconsistent with organizational privacy commitments.

Beginning in 2022, cases involving authenticated patient portals, appointment scheduling systems, and health-related webpages expanded scrutiny toward alleged disclosures more directly tied to care interactions. Doe v. Kaiser Foundation Health Plan, filed in 2023, produced a settlement fund of up to $47.5 million covering approximately 13.4 million Kaiser Permanente members who accessed authenticated webpages and mobile applications between 2017 and 2024. The case illustrates how a single technology governance failure may generate simultaneous exposure under overlapping federal and state privacy frameworks.

In December 2022, OCR issued guidance taking the broad position that tracking technologies on authenticated patient portals involve PHI and that even unauthenticated webpages could implicate HIPAA. Portions of that guidance have since been challenged, and OCR has narrowed aspects of its position. Yet private litigation has continued largely independent of regulatory uncertainty, grounded in overlapping federal and state privacy claims.

The litigation continues to expand rapidly. Since 2022, more than 200 lawsuits involving third-party tracking technologies have been filed against healthcare organizations, spanning academic medical centers, regional health systems, children’s hospitals, fertility clinics, mental health platforms, telehealth providers, and specialty practices nationwide. Court-approved settlements involving Advocate Aurora ($12.25 million), Henry Ford Health (up to $12.2 million), BJC Health System (up to $9.25 million), MarinHealth ($3 million), Mount Nittany ($1.8 million), and Eisenhower Medical Center ($875,000) reflect a litigation wave showing little sign of slowing. Although defendants consistently denied wrongdoing, the scale of settlements underscores the magnitude of perceived governance risk. Some lawsuits have also named Meta and Google, raising the possibility that liability may eventually extend beyond providers to technology platforms.

The broader lesson extends beyond website tracking. Organizations assumed meaningful legal and reputational risk from inexpensive, vendor-supplied technologies peripheral to care delivery, yet capable of transmitting patient-related information in ways not fully understood by leadership. The governance questions exposed in these cases mirror those raised by ambient documentation tools, AI copilots, predictive models, virtual assistants, and clinical decision support systems: What data are captured? Where do they go? Who can access them? How are vendors constrained? Are patients informed? Is use monitored over time? And critically, what are the vendor’s downstream data practices?

For healthcare organizations, the pixel tracking litigation illustrates that governance gaps in vendor-supplied technologies, even those peripheral to clinical care, can generate significant legal and reputational exposure. The questions these cases raised about data flows, vendor practices, patient notice, and ongoing monitoring are the same questions organizations must now apply to AI tools embedded in clinical workflows.

Discrimination Risks in AI-Enabled Care Delivery

Algorithmic discrimination in clinical settings is often inadvertent, emerging through biased training data, proxy variables, or unequal performance across patient populations rather than deliberate design. When patient care decision support tools are trained on datasets that underrepresent certain populations, or rely on variables correlated with race, disability, income, or age, they may perform less accurately or generate systematically different recommendations across patient groups while appearing clinically neutral. In practice, these disparities may arise through risk stratification tools, predictive analytics, or diagnostic support systems that perform differently across populations without clear attribution.

Evidence of differential algorithmic performance is now well documented. Clinical algorithms have demonstrated unequal performance across racial and ethnic groups in contexts ranging from kidney disease assessment and transplant eligibility to illness severity scoring and care management, often not because of intentional design, but because models were trained on historical data reflecting existing disparities in access, utilization, and treatment. The clinical and legal implications of those performance gaps are increasingly difficult to ignore.

These disparities carry litigation and regulatory risk under longstanding federal civil rights protections. Title VI of the Civil Rights Act, Section 504 of the Rehabilitation Act, and the Age Discrimination Act prohibit discrimination based on race, national origin, disability, and age in federally funded health programs. Section 1557 of the Affordable Care Act (ACA) incorporated these protections into a healthcare-specific framework, and 2024 regulations under 45 C.F.R. § 92.210 require covered health programs to make reasonable efforts to identify and mitigate discrimination risks in patient care decision support tools, including AI-enabled predictive analytics, risk stratification algorithms, diagnostic support systems, and other clinical decision support tools. The rule imposes an ongoing duty to identify and mitigate discrimination risks, and while OCR did not prescribe a specific compliance model, it contemplates processes capable of evaluating, monitoring, and responding to discriminatory performance over time. Compliance with these requirements was required by May 1, 2025. By July 2025, covered entities were also required to implement broader written Section 1557 policies and procedures, reinforcing that bias assessment increasingly sits within an established civil rights compliance framework.

The civil rights enforcement landscape around clinical AI is still developing. The clearest judicial precedents do not yet involve bedside clinical AI. Instead, courts in Arkansas, Idaho, and other states found due process violations when opaque Medicaid allocation algorithms made consequential decisions about home care and disability benefits without meaningful transparency or appeal rights. Litigation squarely testing whether AI-enabled clinical tools violate Title VI or Section 1557 on disparate impact grounds has not yet produced definitive rulings, though the regulatory framework and case law trajectory suggest the exposure is real and building.

The enforcement picture has been further complicated by shifting federal priorities. At the federal level, enforcement priorities have shifted away from AI bias and algorithmic fairness initiatives, and a December 2025 executive order directed the Department of Justice to challenge state AI laws inconsistent with federal policy. That creates uncertainty about the pace of federal enforcement, but it does not eliminate underlying legal obligations, Section 1557 regulations remain in place, OCR continues to investigate complaints, and state-level protections in many jurisdictions operate independently of federal enforcement priorities.

Organizations navigating these risks need a clear picture of which patient care decision support tools they use in clinical and coverage decision-making, how those tools perform across patient populations, and whether their programs include not only bias assessment before deployment, but ongoing monitoring for discriminatory performance over time. Organizations best positioned as enforcement and litigation mature will be those that can document what tools they use, how disparities were evaluated, what mitigation steps were taken when disparities or concerns emerged, and how those decisions were governed.

Determining Causation in AI-Assisted Care Events

Determining causation has long been central to healthcare litigation, but AI may make that task substantially harder. Mracek v. Bryn Mawr Hospital (E.D. Pa. 2009, aff'd 3d Cir. 2010) illustrates the evidentiary burden even when technology visibly fails. In that case, the da Vinci robotic surgical system displayed error messages during surgery and could not be made operational despite efforts by the surgical team and a da Vinci representative. The surgeon completed the procedure using laparoscopic equipment. The Third Circuit affirmed summary judgment for the defendants, concluding that evidence of device malfunction alone was insufficient to establish causation. Without evidence linking the device failure to the plaintiff's injuries rather than some other factor, the causation element failed as a matter of law. The lesson remains important: even when technology appears to fail in an observable way, causation cannot simply be assumed.

The ongoing TruDi litigation against Acclarent (2024–2025) illustrates how that burden may become even more complex when harm is alleged to arise from AI-assisted outputs rather than a visible device malfunction. Plaintiffs allege Acclarent's TruDi AI-assisted surgical navigation system misidentified anatomy during sinus procedures, contributing to strokes and other catastrophic injuries. Unlike traditional device failure cases, the central question is not simply whether the technology malfunctioned, but whether AI-generated information materially influenced clinical decision-making under real-world conditions and patient-specific anatomy.

As AI becomes embedded in clinical workflows, causation disputes may increasingly require organizations to establish what role the technology played in a specific encounter. When outcomes are questioned, investigators, regulators, and litigants may seek to understand what technology was used, what information was available to the system, what output was generated, how that output was presented to clinicians, and how clinicians responded. Establishing those facts may provide important evidence regarding whether and how AI-generated information influenced care.

This distinction is important. Determining what happened is not the same as determining why harm occurred. A complete operational record may show that a clinician received and relied upon an AI-generated recommendation, yet causation may still depend on additional questions regarding clinical judgment, patient-specific factors, competing causes, and the reasonableness of relying on the technology under the circumstances. In other cases, operational evidence may provide a clearer causal chain by demonstrating that inaccurate AI-generated information influenced clinical actions that contributed to the outcome.

The TruDi litigation also highlights a related challenge. In traditional malpractice cases, information relevant to causation often resides in the medical record. AI-related disputes may require examination of additional evidence concerning system performance, validation, design limitations, warnings, or other manufacturer-controlled information needed to evaluate competing causation theories. As AI-assisted care continues to evolve, questions of causation may increasingly extend beyond clinician decision-making alone.

For healthcare organizations, understanding the role AI-enabled technology played in an adverse event remains important regardless of whether litigation follows. Clearly, determining what happened is not the same as determining why harm occurred, but the ability to accurately reconstruct the role technology played in care provides a foundation for both patient safety investigation and legal analysis. Organizations deploying AI-enabled technologies should consider whether their existing event investigation and safety review processes are capable of evaluating the role technology played in care. Traditional clinical documentation may not always capture the information needed to understand how AI-generated outputs influenced decision-making, whether safeguards functioned as intended, or whether technology contributed to an adverse outcome.

An emerging category of runtime assurance infrastructure is beginning to address this gap by monitoring AI system behavior as care is delivered, enforcing safeguards in real time, and generating contemporaneous, verifiable records of what output was generated, what safeguards executed, and when. These records establish the factual foundation any causation analysis requires, and they serve patient safety purposes that do not depend on litigation: a root cause analysis needs to know whether an alert fired regardless of whether a court ever asks. As AI adoption expands, the ability to investigate these questions systematically may become an increasingly important component of both patient safety and organizational oversight.

The Insurance Safety Net May Be Narrowing

AI is prompting insurers to reconsider how emerging technology risks are evaluated, priced, and covered. As AI becomes embedded in increasingly consequential clinical functions, insurers are reassessing how these risks fit within traditional coverage structures. New Insurance Services Office (ISO) generative AI exclusions became available beginning in 2026, while insurers and underwriters are increasingly using endorsements, exclusions, and standalone products to address AI-related exposures more explicitly. Coverage that once may have been addressed implicitly is increasingly becoming the subject of specific policy language.

From an actuarial standpoint, AI presents challenges that differ from many traditional risks. Loss history remains limited; models evolve through updates and retraining, and organizations often rely on common platforms, vendors, and foundation models that create the potential for correlated losses across multiple insureds. Gallagher Re describes these as AI-native risks that do not fit neatly within existing insurance structures, and that may be difficult to address through traditional underwriting approaches alone.

The insurance market's response has not been uniform. Some insurers are introducing exclusions or narrowing coverage in certain areas, while others are developing standalone AI liability products designed to address exposures such as hallucinations, model drift, algorithmic discrimination, and other AI-specific failure modes. Gallagher Re characterizes the current environment as fragmented, with traditional cyber, technology errors and omissions (E&O), product liability, and commercial general liability policies each addressing only portions of the overall risk landscape.

For healthcare organizations, one of the more significant observations may involve liability allocation. Gallagher Re observes that courts and regulators may increasingly treat AI-related failures as the responsibility of the organization deploying the technology rather than the technology vendor. Vendor contracts frequently contain liability caps, limited indemnification provisions, and broad warranty disclaimers, potentially leaving deployers with greater exposure than they may expect. Organizations may find that responsibility for AI-related outcomes remains with the entity using the technology, even when key aspects of that technology are controlled by third parties.

The underwriting conversation may also be changing. Gallagher Re identifies governance frameworks such as NIST AI RMF and ISO 42001 as important components of managing AI-related risk. As insurers work to better understand and evaluate AI-related exposures, governance practices such as inventory management, oversight, monitoring, auditability, and incident response may become increasingly relevant to conversations about risk assessment and insurability.

Healthcare organizations should evaluate AI governance, vendor contracting, and insurance coverage as interconnected components of a broader risk management strategy. As AI adoption expands, understanding how liability is allocated among providers, vendors, and insurers may become as important as understanding the technology itself. Organizations should assess whether their insurance programs, indemnification provisions, liability caps, and governance controls align with the risks their AI initiatives create, and whether responsibility, authority, and risk are aligned across those initiatives.

What’s Next for Healthcare Orgs

Healthcare’s biggest litigation risks remain remarkably stable, informed consent gaps, privacy violations, discrimination and documentation gaps are not new legal theories. What is changing are the technologies, workflows, and failure modes through which these liabilities increasingly emerge. AI is not creating an entirely new category of legal exposure so much as introducing new ways longstanding risks may materialize and spread at scale.

At the same time, AI may challenge whether traditional approaches to safety investigation, governance, and liability management remain sufficient. In deterministic systems, organizations could often reconstruct what happened through audit logs, software rules, and observable failures. As healthcare increasingly relies on probabilistic systems whose behavior may vary across settings, populations, and workflows, determining what role technology played in a specific event may become substantially harder.

To meet their responsibility for delivering safe care, healthcare organizations may need to rethink how they govern, monitor, and manage these evolving risks not only through stronger oversight at deployment, but through operational monitoring, reconstructability, contractual protections, and insurance strategies capable of keeping pace with an increasingly AI-enabled healthcare environment. Organizations best positioned to navigate this environment will be those that approached AI adoption not as a technology implementation but as a governance challenge. That challenge requires sustained attention to how liability is allocated, how performance is monitored, and how organizational accountability is maintained when the technology driving care decisions is neither fully transparent nor fully within the organization's control.