HIPAA & AI Compliance

Overview

The Health Insurance Portability and Accountability Act (HIPAA) is the primary United States federal law governing the privacy and security of protected health information (PHI). Enacted in 1996 and continuously updated through the HITECH Act and subsequent rules, HIPAA establishes national standards for the protection of individually identifiable health information. For organizations developing AI in healthcare, HIPAA compliance is mandatory. Civil penalties are tiered by level of culpability and adjusted for inflation each year, and at the top tier the annual cap for violations of an identical provision exceeds US$2 million.

HIPAA's scope covers all "covered entities" (health plans, healthcare clearinghouses, and healthcare providers) and their "business associates", meaning any organization that creates, receives, maintains or transmits PHI on behalf of a covered entity. When an AI vendor processes patient records, clinical notes, diagnostic images, or any data that could identify a patient on a covered entity's behalf, it becomes a business associate subject to HIPAA's regulatory requirements. This has profound implications for how AI training pipelines are designed and where patient data is allowed to flow.

The regulation is built on three main rules: the Privacy Rule, the Security Rule, and the Breach Notification Rule. The Privacy Rule establishes who can access PHI and under what conditions. The Security Rule specifies safeguards for electronic PHI (ePHI), including access controls, audit controls, integrity controls, and transmission security. The Breach Notification Rule mandates that covered entities notify affected individuals, HHS, and sometimes the media when unsecured PHI is compromised.

AI-Specific Requirements

AI systems in healthcare face heightened HIPAA scrutiny because they often require large volumes of patient data for training. The Security Rule's technical safeguard requirements (45 CFR 164.312) demand access controls that restrict system access to authorized users, audit controls that record and examine activity in systems containing ePHI, integrity controls that protect ePHI from improper alteration or destruction, and transmission security that guards against unauthorized access during electronic transmission.

The minimum necessary standard is particularly relevant for AI training data preparation. Under this principle, organizations must limit PHI use and disclosure to the minimum necessary to accomplish the intended purpose. For AI teams, this means scoping training datasets to the clinical data elements the model's specific use case requires, rather than including entire patient records. Data de-identified under the HIPAA Safe Harbor method (removing 18 specific identifiers) or the Expert Determination method is no longer PHI and can be used outside HIPAA's restrictions, but the de-identification must be thorough and verifiable.

Business Associate Agreements (BAAs) are required whenever a service provider creates, receives, maintains or transmits PHI on behalf of a covered entity. For AI development, that covers any cloud platform, annotation service, or model hosting provider that touches PHI. Not every AI service will sign a BAA, and some that do limit which of their services may be used with PHI. That leaves healthcare organizations two practical options for AI work: de-identify data before it reaches an outside provider, or keep PHI-bearing workloads inside infrastructure they already control.

How Ertas Helps

Ertas Studio trains on Ertas's cloud GPUs, so every training example is uploaded to Ertas and stored so the run can be repeated. Treat Studio as a place for de-identified data: remove the 18 Safe Harbor identifiers, or obtain an expert determination, before you upload, so the training set is no longer PHI. Data Craft, Studio's dataset builder, lets you edit each example and track its approval status before training, and you can delete an uploaded dataset at any time. If your use case needs identifiable PHI in training, raise it with us before uploading anything, so the contractual position, including whether a BAA is needed, is settled first.

Inference is where PHI usually appears, because clinicians and staff put real records into prompts. Export the trained model as GGUF or safetensors and run it inside your own environment, and that inference traffic never reaches Ertas. You own the model, and the version you validated stays fixed until you choose to retrain. If you want the model run for you, the adaptive model endpoint and enterprise builds can run it in your own cloud or on-premise.

The HIPAA program stays yours: the Security Rule risk analysis, BAAs with your own vendors, access controls, audit controls and encryption in the environment where the model runs, breach response, and workforce training. Ertas does not provide those controls for your environment.

Compliance Checklist