Blog Posts
67% of employees are using AI tools at work, with minimal to no visibility. Learn what Shadow AI is and how to protect your organization against it.
72% of doctors use AI in clinical practice today – up from 48% a year ago. The majority are using tools with no Business Associate Agreement in place. The violation does not trigger an alert. No log entry, no ransomware notice, no stolen laptop report. An open exposure window with no timestamp. Here is exactly what is compliant, what is not, and why.
Every healthcare compliance officer has been asked at some point: "Is it ok for me to use ChatGPT to write this patient note?" The honest answer is rarely a simple yes or no – it depends on which ChatGPT, which plan, whether there is a Business Associate Agreement, what data was entered and what happens to that data after. According to the American Medical Association's Center for Digital Health and AI, Physician Survey on Augmented Intelligence (February 2026), 72% of physicians now report using AI in clinical practice — up from 48% the year before and 38% in 2023.
None of that adoption curve was accompanied by a parallel curve in governance. A December 2025 survey of 518 healthcare providers and administrators, reported by the Association of Health Care Journalists, found that 40% had encountered an unauthorized "shadow AI" tool in their organization, and a further 17% admitted to using one themselves — entirely outside any approved workflow, BAA, or oversight. This is what the industry now calls Shadow AI, and in healthcare it carries a specific, well-defined legal exposure that doesn't exist in most other industries: HIPAA.
Under HIPAA's Privacy Rule, any vendor who develops, receives, maintains or transmits PHI on behalf of a covered entity is a "business associate." A covered entity cannot share PHI with a business associate unless there is a signed Business Associate Agreement (BAA) – a legally binding contract under 45 CFR §164.504(e) that requires the vendor to keep the data safe under HIPAA. No BAA, no PHI, period, no matter how good the privacy settings of the tool look.
The problem in 2026 is that "ChatGPT" is no longer one product — it's at least five, each with a completely different BAA status. OpenAI launched two healthcare-adjacent products on January 8, 2026 with nearly identical names — and confusing them is the single most common compliance failure we now see.
| Product / Tier | BAA Available? | Can Process PHI? | What It Actually Is |
|---|---|---|---|
| ChatGPT Free / Plus / Team | No | Never | Consumer tiers. OpenAI does not offer a BAA for these plans under any circumstance — any PHI entered without a BAA in place is an automatic violation under 45 CFR §164.504(e). |
| ChatGPT Health | No | Never | Consumer wellness product announced Jan 7, 2026. Designed for personal health literacy, not clinical use. OpenAI will not sign a BAA for this product under any configuration. |
| ChatGPT Enterprise / Edu | Conditional | If BAA signed | OpenAI may sign a BAA on request for these tiers — but it must be deliberately executed, and per 45 CFR §164.504(e) the exact deployment must match what's covered under the agreement. |
| ChatGPT API | Conditional | If BAA signed | Healthcare developers can apply for a BAA to embed the API into clinical and operational applications — approval depends on the surrounding application also being HIPAA-compliant. |
| ChatGPT for Healthcare | Yes | Yes (BAA required) | Enterprise product launched Jan 8, 2026 for large health systems. Initial partners: Cedars-Sinai, Stanford Medicine Children's Health, Memorial Sloan Kettering, UCSF, Boston Children's. Not generally available to small practices. |
Vendors describe enterprise tiers as "HIPAA-eligible." That phrase is doing a lot of work — it means the vendor can sign a Business Associate Agreement under specific configurations, not that any deployment of their product is automatically covered. A signed BAA, specifically configured access controls, and the exact tool your staff member opened all need to match. The free-tier tool your front desk coordinator used yesterday is never covered, no matter what the enterprise contract says about a different deployment.
Follow this decision tool with a real case from your organization. It is the same logic that OCR investigators use to assess a complaint.
These are the stories we hear most often when working with our healthcare clients. Each of these happens dozens of times a day in a typical health system.
The HIPAA Safe Harbor method (45 CFR §164.514(b)(2)) provides a list of 18 specific identifiers. If one or more of these is associated with health information, the combination is PHI – period. Most staff think that PHI is name + diagnosis. It is a much longer list and some of the most commonly missed identifiers are in every clinical note that is written.
This one, relatively ordinary progress note contains 7 of the 18 identifiers – and that's without counting the diagnosis itself. The complete list also includes fax numbers, SSNs, account numbers, certificate/license numbers, device identifiers, URLs, IP addresses, biometric identifiers, and "any other unique identifying number, haracteristic, or code" – a catch-all that courts and OCR have read broadly. A clinician who pastes "this note" into an AI tool to "clean up the formatting" has just sent all 7 to whatever vendor is on the other
"Your staff isn't trying to break HIPAA – they're trying to do their jobs. But every time patient data is entered into an AI tool without a signed Business Associate Agreement, a violation has occurred. Quietly. No alert, no log entry, no breach notification trigger. Just an open exposure window with no timestamp."
— Polygraf AI, on what we see in every healthcare AI governance auditIn contrast to a ransomware attack or a lost laptop, an AI-related PHI breach is not usually found in an alert. It is found in a routine audit, a patient complaint or – more and more – in an OCR investigation of something unrelated. Once found, the HIPAA penalty structure is per violation per category, and the amounts are set out in 45 CFR §160.404 and adjusted annually for inflation. The figures below reflect the 2025-adjusted amounts published by HHS in the Federal Register on January 28, 2026.
The organization did not know and could not reasonably have known of the violation. Most accidental AI PHI entry by a well-intentioned employee begins here, if the organization had a reasonable policy and training program.
The breach was not negligent but reasonable – e.g. there was a policy but staff were not trained on how it relates to AI tools.
Leadership knew that staff were using unapproved AI tools with patient data (a known issue) and fixed it in 30 days once it was formally reported.
Leadership was aware of the widespread use of Shadow AI (e.g., from survey data that showed about one in five staff were already using unauthorized AI tools) and did nothing to address it. This is where the largest OCR settlements come from.
These figures are per violation, and a single AI prompt can contain multiple identifiers — each potentially counted separately. A single breach can involve multiple violation categories (privacy, security, breach notification), each penalized separately under 45 CFR §160.404, with combined state attorney general actions sometimes pushing total exposure for a single incident into the tens of millions. The HHS Office for Civil Rights has settled or imposed civil money penalties in 152 cases to date, totaling $144,878,972 — and that's before forensic investigation costs (often $50,000–$500,000+) and breach notification costs.
The aim is not to stop AI – clinicians have made it clear they will use it and productivity gains are real. The aim is to make the easy path the compliant path so staff do not have to choose between doing their job efficiently and following the rules.
Polygraf AI's Desktop Overlay and Behavioral Control Plane are where the staff actually type – looking at prompts in real time for any of the 18 Safe Harbor identifiers before they are sent to an AI tool, whether or not that tool has a BAA. For approved AI tools, Polygraf enforces that PHI is only sent to covered, BAA-approved endpoints. For everything else, sensitive content is redacted or blocked before it leaves the device. Every detection is logged – creating the exact evidence trail that will separate Tier 1 from Tier 4 if OCR ever asks.
Polygraf AI reviews AI prompts at the gate for all 18 HIPAA Safe Harbor identifiers, blocking or redacting PHI before it gets to any AI tool, covered or not BAA. Sub-100ms. On-prem. Full audit trail of every detection.
At Polygraf, we envision a future where AI augments human capabilities without compromising safety, privacy, or ethical standards. Trust in our commitment to building this future with you.
© 2026 Polygraf AI. All rights reserved.
Our Solutions 
Regulations Your download will start now.
Please provide information below and we will send you a link to download the white paper.
Data Privacy 
Data Provenance 
Content Generation