Stop Treating AI Like Magic: A HIPAA-Safe Field Guide

Your AI project is a HIPAA violation waiting to happen

Everyone tells you to jump on the AI bandwagon, but nobody warns you that the real problem isn't the algorithm—it's the data. You're a health informatician, and you've been handed a mandate: "Use AI to improve outcomes." But before you feed a single patient record into a model, you need to understand that the most common advice—"just start small"—is exactly what will get you fined. Here's the contrarian truth: starting small with real patient data is the riskiest move you can make. You need to start with the regulations, not the code.

Imagine you're the informatics lead

Picture yourself at a mid-sized clinic. The CEO heard that AI can predict no-shows and reduce costs. She wants a pilot in two months. You're the one who has to make it work without triggering an audit. The first thing you realize is that your EHR is certified by ONC—so is 96% of the U.S. hospital market (ONC / HHS (Report to Congress)). But certification doesn't mean you're ready for AI. It means your data is structured, but it's fragmented across HL7 v2 messages, FHIR resources, and maybe some old CDA documents. You need to integrate all of it, and that's where the trouble begins.

Know your data standards before you know your model

You might be tempted to pull everything into one big data lake. Don't. Instead, you need to map your data to standards that AI can actually consume. FHIR is your friend—it's HL7's modern standard using RESTful APIs, and it represents clinical concepts as discrete resources like Patient, Encounter, and Observation (HL7 International). If you're in the U.S., you'll want to use US Core, which is based on FHIR R4 and defines the minimum constraints for patient data access (US Core Implementation Guide). That's your baseline. But don't forget the older stuff: HL7 v2 is still used by more than 95% of U.S. healthcare organizations (HL7 International). You'll need to bridge both worlds. And when it comes to terminology, you need SNOMED CT for clinical terms, LOINC for lab identifiers, and ICD-10 for diagnoses (HL7 International). If your data isn't coded in these, your AI will be garbage in, garbage out.

Here's where HIPAA gets real

Now, here's the part that most AI tutorials skip: HIPAA. You're a covered entity because you transmit health information electronically for claims or other transactions (45 CFR 160.103). That means every piece of patient data you touch is protected health information (PHI), and you have to protect it. The Security Rule requires three categories of safeguards: administrative, physical, and technical (ONC / HHS (Health IT)). For AI, the technical safeguards are critical—you need access control, audit controls, and encryption (45 CFR 164.312). But there's more. If you use a cloud vendor to train your model, that vendor is a business associate, and you need a contract that ensures they protect the data (45 CFR 160.103). And remember the minimum necessary standard: your AI should only use the minimum amount of PHI needed to accomplish its purpose (45 CFR 164.502(b)). That's not just a good idea—it's the law.

Breach notification is a nightmare you can avoid

Let's talk about what happens when things go wrong. Under the HIPAA Breach Notification Rule, if unsecured PHI is breached, you must notify each individual within 60 days of discovery (45 CFR 164.404). That's a tight window. And the fines are not pocket change: the calendar-year cap is now $2,190,294 (Federal Register (2026 HIPAA CMP Adjustment)). That's per year, not per breach. So a single mistake could wipe out your entire AI budget. The only way to avoid this is to encrypt your data and limit access. But encryption is an addressable specification under the Access Control standard (45 CFR 164.312(a)), which means you have to decide if it's reasonable and appropriate. For AI, it's not optional—if you're using cloud GPUs, you encrypt everything.

So, what's the actual playbook?

Here's the blunt advice you came for. Don't start with a pilot on live patient data. Start with synthetic data or de-identified data. That way, you're not triggering HIPAA at all. Once your model works on synthetic data, then move to a small set of real data that you've carefully de-identified. But even de-identification is tricky—HIPAA's expert determination method requires that the risk of re-identification be very small. If you're not confident, use a limited data set with a data use agreement. And when you do use real PHI, make sure you have a robust business associate agreement with your AI vendor—they're a business associate under HIPAA (45 CFR 160.103).

Don't forget the algorithm transparency rules

One more thing: if your AI is part of certified health IT, you're now subject to the ONC HTI-1 rule, which took effect March 11, 2024 (ONC / HHS (HTI-1 Final Rule)). That rule requires transparency for AI and predictive algorithms. That means you have to document how your algorithm works, how it was trained, and what its limitations are. This is a big deal—it's the first-of-its-kind requirement. And NIST's AI Risk Management Framework, released January 26, 2023, is voluntary, but you should use it as a guide (NIST AI Risk Management Framework). It will help you address trustworthiness throughout the AI lifecycle.

Quick tip: Before you write a single line of code, write a data flow diagram showing every place PHI touches your AI system. Then get it reviewed by a HIPAA expert.

Bottom line

AI in healthcare is not a software project—it's a data governance project. The single best move you can make is to invest in a solid data governance framework that maps every piece of data to a standard, documents your HIPAA safeguards, and plans for algorithm transparency. Do that first, and your AI will actually improve care. Skip it, and you'll be the next cautionary tale at a health informatics conference.

Sources

• ONC / HHS (Report to Congress) - https://healthit.gov/news/onc-outlines-health-it-interoperability-progress-report-congress/
• HL7 International - https://www.hl7.org/fhir/
• eCFR 45 CFR Part 160 (HIPAA Definitions) - https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-160
• eCFR 45 CFR Part 164 Subpart C (Security Rule) - https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C
• Federal Register (2026 HIPAA CMP Adjustment) - https://www.federalregister.gov/documents/2026/01/28/2026-01688/annual-civil-monetary-penalties-inflation-adjustment
• ONC / HHS (HTI-1 Final Rule) - https://healthit.gov/regulations/hti-rules/hti-1-final-rule/