Four sessions at the September HIPAA security conference focused on artificial intelligence, including a preview of federal guidance still in draft.

 

What happened

The HHS Office for Civil Rights and the National Institute of Standards and Technology held their annual Safeguarding Health Information conference on September 2 and 3, 2026, at the NIST campus in Gaithersburg, Maryland, according to the event listing. Four sessions covered artificial intelligence directly, including measuring AI tools in healthcare, the NIST framework for managing AI risk, a panel with speakers from several sectors, and a preview of guidance still in development, Meghan Mead of the Network for Public Health Law reported. Speakers came from OCR, NIST, the HHS Administration for Strategic Preparedness and Response, the Federal Trade Commission, and private healthcare organizations. Other sessions covered enforcement updates and encryption methods designed to withstand future computers.

 

Going deeper

Governance about the whole life of an AI tool, from the point of purchase through to daily use, came up across the presentations. Speakers argued that assessing an AI tool means looking at the outcomes it produces rather than only its outputs, which requires judgement as well as numbers. That pushed the discussion toward measurement itself, since an organization cannot judge a tool's accuracy without agreeing in advance what accuracy means and how to test for it. NIST sets out nine characteristics an AI system needs before it can be considered trustworthy, covering validity and reliability, safety, security, explainability, privacy, resilience, fairness with harmful bias managed, accountability, and transparency, in its AI Risk Management Framework. Speakers also raised a practical difficulty, since AI models do not behave identically each time they run, change as they are updated, and can be withdrawn and replaced by their vendor with little notice.

 

What was said

"Robust AI governance is the foundation for ensuring AI tools are used safely and effectively," Mead wrote in her summary of what speakers from both the public and private sectors told the conference, published September 16, 2026. She noted that the United States has no detailed federal law governing AI, and that organizations are asking for guidance on addressing security risks before incidents occur rather than afterwards. Martin Stanley of NIST's AI Standards and Guidelines Group told the conference that managing AI risk is a moving target because the models themselves change over time.

 

In the know

NIST is building a set of AI-specific recommendations on top of its existing cybersecurity guidance rather than writing a separate framework. A preliminary draft of the Cyber AI Profile, published as Internal Report 8596 in December 2025, is organized around three areas: securing AI systems, using AI in cyber defence, and protecting an organization against attacks that use AI, according to the draft. It adds AI considerations to each part of the existing Cybersecurity Framework, so organizations already using that framework extend what they have rather than starting again. NIST has held public workshops on the profile, and its published roadmap indicates another workshop and a further draft are planned, without a date for either. NIST is separately revising the AI Risk Management Framework, also with no announced release date.

 

The big picture

Vendors are where most healthcare AI arrives, and the Health Sector Coordinating Council has published a Third-Party AI Risk and Supply Chain Transparency Guide directed at organizations assessing suppliers that have built AI into their products or plan to. Two speakers from healthcare organizations pointed conference attendees to that material. The guidance now available is voluntary rather than required, which leaves compliance teams deciding how much of it to adopt while the underlying HIPAA obligations stay unchanged. An AI tool processing patient information still needs a risk analysis covering it, access controls limiting what it can reach, and a business associate agreement with whoever supplies it.

 

FAQs

What is the NIST AI Risk Management Framework?

A voluntary framework published in 2023 for identifying and managing risks in AI systems. It is organized around four activities, covering setting up governance, understanding the context an AI system operates in, measuring its risks, and managing them, and it carries no regulatory force on its own.

 

Why does it matter that AI models are not deterministic?

Conventional software, given the same input, returns the same output every time, which makes testing straightforward. AI models can produce different results from identical inputs, and their behaviour shifts as they are retrained, so a test passed at deployment says less about how the tool performs six months later.

 

What does measuring outcomes rather than outputs involve?

An output is what the tool produces, such as a draft note or a risk score. An outcome is what happened as a result, such as whether documentation improved or whether flagged patients actually deteriorated. Measuring the second requires following cases over time rather than checking the tool's immediate response.

 

Does voluntary federal guidance carry any weight in an OCR investigation?

Not as a legal requirement, though regulators do consider whether an organization followed recognized practices when assessing its security programme. Documented use of a recognized framework is easier to defend than an approach with no external reference point.

 

What should a healthcare organization ask an AI vendor?

What data the tool processes and where it goes, whether the vendor will sign a business associate agreement, how the tool's performance is measured after deployment, what happens if the underlying model is updated or withdrawn, and whether the vendor will disclose the AI components inside its own product.