Generative AI complicates this framework. In the event of an adverse outcome, the question of who bears responsibility—provider, health system, AI tool developer, or product manufacturer — becomes genuinely difficult. Time will tell how courts will allocate liability, but plaintiffs’ attorneys are exploring theories grounded in negligent adoption, inadequate oversight, and failure to verify AI outputs.

Critically, providers cannot delegate clinical judgment to an algorithm and expect it to absorb liability. AI does not hold professional licensure and is not subject to ethical standards that bind licensed practitioners. State licensing and scope-of-practice rules increasingly require licensed professionals to review and approve outputs, and expanding healthcare AI laws reinforce that AI cannot be the sole basis for clinical decisions. Providers who over-rely on outputs without applying their own expertise may face claims that they abdicated professional responsibility.

Beyond malpractice, health systems and AI developers face products liability exposure. Traditional products liability distinguishes between a product, its user, and the patient. AI blurs those boundaries; the chain of liability runs from hardware to software to developer to manufacturer to human provider, making fault allocation far more complex than in conventional medical device cases. Strict liability theories — including manufacturing or design defect, failure to warn, negligence, and breach of warranty — may all be viable. An open question is whether the standard of care will be higher for AI-enabled products because the software is arguably more “intelligent” than a reasonably prudent person, potentially raising expectations about what constitutes a defective or substandard output or a “reasonable” clinical provider.

Perhaps the most underappreciated risk is the proliferation of unapproved AI tools. Staff may turn to publicly available AI platforms to summarize patient records, draft clinical notes, or research treatment options. These tools are often adopted without institutional knowledge or authorization and lack safeguards like encryption and closed-loop data handling.

Shadow AI creates legal and compliance exposure on multiple fronts. Unapproved tools have not been vetted for clinical accuracy and may violate organizational policies, regulatory requirements, and contractual obligations. If an adverse outcome is linked to reliance on an unapproved tool, the organization may face liability for both the clinical error and its failure to establish adequate AI governance.

Enforcement actions have also targeted process failures — including inadequate vendor diligence and failures of transparency — not just adverse outcomes. State attorneys general are not waiting for demonstrable patient harm before investigating alleged false and misleading product claims.

There is a fundamental tension between familiar privacy principles — minimum necessary and data minimization — and AI’s operational demands for data retention to support explainability, bias analysis, and transparency. Generative AI tools require large volumes of data to function effectively. Organizations must evaluate whether AI tools transmit patient data to external servers, data is used to train third-party models, or adequate encryption and access controls are in place.

Under HIPAA and state privacy laws, organizations that fail to safeguard personal information processed by AI systems face significant penalties, potential algorithmic disgorgement, and litigation. A pivot to deidentified data does not solve these risks. The process of deidentifying data is itself a “use” of data, and certain state laws require consumer consent before deidentifying data and specific contractual flow-down terms with recipients.

Providers do not need to avoid AI to manage these risks, but they must be deliberate. Organizations can take several steps to reduce liability exposure:

• Establish formal AI governance committees that include clinical, legal, compliance, and information security leadership. AI tools should not be deployed without institutional review and approval.
• Develop standards addressing patient notification, clinician obligations to independently verify AI-generated outputs, and permissible use of AI-enabled products, supported by training.
• Conduct due diligence on AI vendors: understand how models are trained, what data rights vendors retain, how outputs are validated, and what ongoing monitoring the vendor supports. Use the HHS AI transparency rule “nutrition label” questions as a guide and update commercial contract terms, BAAs, and DPAs to address AI-specific concerns.
• Implement ongoing monitoring protocols. AI systems change over time. Monitor for output drift, bias emergence, data quality degradation, and scope creep. Documentation creates critical evidence of diligence in any future enforcement action or litigation.
• Treat AI incidents like safety events, not merely IT glitches, as these incidents can be a catalyst for class action litigation, regulatory enforcement, and lasting reputational harm.

The integration of AI into healthcare is not a question of whether but how. Organizations that invest now in governance, training, and oversight will be better positioned to capture AI’s benefits while managing the accompanying risks.