The AI Risk Inside the Clinic

By Keavy Murphy, Vice President, Security, Net Health

Nearly every conversation about artificial intelligence (AI) in healthcare centers on what the technology returns. Is the output accurate? Will the model fabricate something? Can a clinician rely on it?

These are reasonable questions, and they are being asked constantly. From a security and privacy perspective, there is another question healthcare organizations should ask before an AI system ever produces an answer:

How much patient information did the system need in the first place?

Industry attention has narrowed onto model accuracy, output quality, and hallucination. Protected health information (PHI), meanwhile, is entering AI tools in substantial volume, frequently without guardrails or approval.

Oversharing Risk

Healthcare employees routinely work with information-rich records. When someone asks an AI tool for help, providing additional context can feel both natural and useful. The result is a tendency to give the system everything that might help rather than only what it needs.

Consider a few examples:

• A therapist wants assistance summarizing recent progress and provides an entire patient chart when only a defined episode of care is needed.
• A billing employee troubleshooting a claim uploads a spreadsheet containing hundreds of patient records when the problem concerns one claim.
• A wound care physician wants to compare healing rates for surgical debridement patients and includes all patient data for the clinic

In each case, the intended task may be legitimate. The security problem is the scope of the data provided. No rule was broken, and no one intended to expose anything, but in each case the volume of PHI in play far exceeds what the task truly required.

This pattern predates AI. Clinical and administrative staff working under time pressure make security decisions by instinct rather than analysis, and the more thought a secure choice requires, the less likely it is to be made. What AI changed is the scale. These tools accept whatever you give them and never ask why you gave them so much, so a single action now moves far more information than any prior workflow allowed.

What those examples describe is a failure of data minimization, which is the principle that you use the smallest amount of information needed to accomplish a task and nothing more.

In security practice, it limits exposure, since data that was never shared cannot be logged in a system you do not control, retained past its usefulness, or surfaced somewhere unintentionally.

In healthcare, it is also a regulatory requirement. HIPAA’s minimum necessary standard obligates covered entities to limit uses, disclosures, and requests for PHI to what is reasonably required for the purpose at hand. The standard has always applied to how staff share records with each other and with vendors. An AI prompt is a disclosure like any other, and the obligation does not change because the recipient is a model instead of a person.

A Practical Test

Data minimization does not mean depriving clinicians or applications of information they legitimately need. The goal is not “use less data at all costs.” The goal is avoiding unnecessary data exposure.

A useful question for staff is:

Would this task still work if I provided less patient information?

If the answer is yes, input less data.

Instead of a complete schedule, provide the text that needs to be rewritten. Instead of an entire chart, provide the relevant timeframe or documentation. Instead of a multi-patient spreadsheet, isolate the record necessary to investigate the issue.

The same test can be applied when healthcare organizations evaluate technology. If an AI feature performs a narrowly defined function, ask why it needs access to particular categories of data and whether that access can be constrained.

Practical Guidance

Most organizations find that a ban on AI tools is difficult to enforce and easy to work around. The practical moves are about visibility and scope.

1. Find out where it is happening: Ask your scheduling, therapy, and billing teams which tools they are using and for what. The more useful follow-up is how much they are putting in, since that is where potential PHI overexposure occurs.
2. Give employees an approved path: If employees have a legitimate productivity need for AI, but no sanctioned way to meet it, they may turn to publicly available services. Approved tools should have appropriate security, privacy, contractual, access, and data-handling controls for the information they are expected to process.
3. Teach employees to minimize, not simply prohibit: “Never use AI” is difficult guidance to apply in an environment where AI capabilities are increasingly embedded into everyday software. “Only provide the information necessary for the task” is an ethos that employees can use whether they encounter AI today or three years from now.

Net Health builds AI features to operate on the minimum data a task requires, with minimum-necessary PHI access enforced at each step of the pipeline, human review before any AI-generated value is committed to the record, and end-to-end audit logging of AI interactions and PHI access.

That design sits inside an AI management system certified to ISO/IEC 42001, independently validated, so the controls are audited rather than asserted.