Is It HIPAA-Compliant to Use AI for Radiology Report Drafting?

Every radiologist who has thought about pasting a dictation into an AI tool has paused at the same moment: is this allowed? The honest answer is that HIPAA does not prohibit using AI to draft a report. It regulates what happens to protected health information along the way. Whether a given workflow is compliant comes down to three questions: what leaves your workstation, who processes it and under what agreement, and what controls and records exist around it.

This article walks through those three questions from the perspective of an individual radiologist, not a hospital IT department. It is general information, not legal advice; your institution's privacy office has the final word on your own workflow.

What in a dictation is actually PHI

Protected health information is health information that can be tied to a person. The HIPAA Safe Harbor method lists eighteen identifier types that must be removed for data to count as de-identified: names, geographic subdivisions smaller than a state, all elements of dates other than year, phone and fax numbers, email addresses, Social Security numbers, medical record numbers, health plan and account numbers, certificate and license numbers, vehicle and device identifiers, URLs, IP addresses, biometric identifiers, full-face photographs, and any other unique identifying number or code.

Look at where those live in a radiology report. Almost all of them sit in the header: patient name, date of birth, MRN, accession number, study date, referring physician. The body of the report, the part an AI tool actually helps with, usually contains findings, measurements, and comparisons. The usual leaks are the ones that slip into the narrative: “compared with the study of March 14,” a referrer's name in the indication, or an accession number pasted along with the text.

Three ways a workflow can be compliant

1. De-identify before anything leaves. If no PHI is disclosed to the tool, HIPAA's disclosure rules are not triggered. This is the cleanest path for an individual and the one most privacy officers will accept without a meeting.
2. Disclose PHI only to a business associate. If PHI does reach a vendor, HIPAA requires a Business Associate Agreement between the covered entity and that vendor, and the vendor must in turn have agreements with its own subprocessors (hosting, transcription, AI inference).
3. Both. Redact what you can, and use a vendor that is prepared to be a business associate for whatever remains. This is the defensible default for real clinical use.

Note who signs a BAA: the covered entity. For an employed or contracted radiologist, that is typically the hospital or the group, not you personally. That is why individual use of a tool tends to start with de-identified or synthetic cases, and why group or enterprise use is where BAAs get signed.

Why general-purpose chatbots are the wrong tool for this

The convenient option is the chatbot already open in another tab. It is also the option most hospital policies now explicitly prohibit for patient information, for good reasons: consumer AI services generally will not sign a BAA, their terms may allow using your inputs to improve their models, they keep no audit trail you can produce for a compliance review, and nothing in them detects an MRN before you press enter. None of that is a criticism of the underlying models. It is a mismatch between a consumer product and a regulated workflow.

Questions to ask any vendor

Whether you are evaluating a tool for yourself or recommending one to your group, these questions separate a product built for clinical text from one that merely accepts it.

• Will you sign a Business Associate Agreement, and at what tier?
• Which subprocessors handle my data (hosting, transcription, AI inference), and do they have BAAs with you?
• Where is data processed and stored, and can you tell me the region?
• Is my data used to train models? Is that contractually excluded?
• How long is report text retained, and can I delete it?
• Is data encrypted in transit and at rest?
• Is there an audit log of who accessed what, and can it be exported for a compliance review?
• Is multi-factor authentication available, and is it enforced for accounts that handle PHI?
• Does the tool detect and redact identifiers before text reaches the AI, and can I review what it found?
• What is your breach notification process and timeline?

How RadPro AI approaches it

RadPro AI was built to answer these questions up front, so the safeguards are in the product rather than in a policy document. Before any text reaches an AI model, automatic PHI detection screens for the Safe Harbor identifier types and highlights each item it finds in a color for its type so you can confirm or dismiss it; redaction happens before processing. Data is encrypted in transit and at rest, every report and session action is written to an audit log that can be exported, and sessions time out automatically. Multi-factor authentication is optional on the free tier and mandatory on paid plans, because paid plans are where real patient workflows live. Reports are stored in Toronto. The AI steps run on Amazon Web Services in the United States; with identifier protection on, the text sent there carries placeholders instead of the identifiers RadPro AI detected. Executed BAAs are in place with the hosting and AI model infrastructure providers, and a Business Associate Agreement with RadPro AI is available on request for covered entities.

Two honest caveats. PHI detection assists your review; it does not guarantee that every identifier is caught, and you remain responsible for what you submit and for the privacy obligations that apply to you. And RadPro AI is a drafting aid and is not intended for use as a medical device: it does not interpret images, and every draft is reviewed and signed by you.

A compliant workflow for an individual radiologist

1. Dictate the case the way you normally do, in your existing dictation system.
2. Copy the report body only. Leave the header, accession number, and referrer out.
3. Generalize dates in the narrative (“prior study from three months ago”).
4. Paste into the tool with PHI redaction turned on, and review the highlighted items before processing.
5. Verify the structured draft and impression against your own source text. The structuring step must not add or remove findings.
6. Paste the edited result back into your reporting system and sign it there. Nothing about your system of record changes.

If you practice in Canada, the same logic applies under PIPEDA and provincial health information laws such as Ontario's PHIPA: minimize what is disclosed, know where it goes, and keep a record. Using any tool, including this one, does not by itself establish compliance with any privacy law.

FAQ

Can I use the free tier with real patient data?

The free tier has the same PHI detection and redaction as paid plans, but it does not enforce multi-factor authentication, and a BAA is in place only when a covered entity has requested and signed one. The recommended free-tier workflow is synthetic or de-identified text. If you want to run real cases, use a paid plan with MFA, and check your institution's policy first.

Does automatic PHI redaction make me compliant?

No tool can make a workflow compliant on its own. Redaction reduces what is disclosed and gives you a review step, which is valuable, but you remain responsible for reviewing what you submit and for the obligations that apply to you and your employer.

Do I need a BAA if I am an individual radiologist?

A BAA is required whenever PHI is disclosed to a vendor, and it is signed by the covered entity, which is usually your hospital or group rather than you personally. The two individual-friendly paths are to de-identify before disclosure, or to bring the tool to your hospital or group so that, as the covered entity, it can request a BAA. Read more about the workflow side in structured reporting without changing how you dictate.