HIPAA Compliance for AI Tools

HIPAA COMPLIANCE CHECKLIST FOR AI TOOLS IN HEALTHCARE
A Practical Guide for Practice Administrators
========================================================================

Prepared by ReasonNotes | reasonnotes.com

TABLE OF CONTENTS
1. Understanding HIPAA as It Applies to AI Tools
2. Penalty Tiers and Real Enforcement Cases
3. Business Associate Agreement (BAA) Requirements
4. Technical Safeguard Requirements
5. Questions to Ask AI Vendors Before Signing
6. Compliance Checklist Summary
7. Key Regulatory References

========================================================================
SECTION 1: UNDERSTANDING HIPAA AS IT APPLIES TO AI TOOLS
========================================================================

Any AI tool that creates, receives, maintains, or transmits electronic
protected health information (ePHI) on behalf of a covered entity is
subject to HIPAA. This includes:

- AI-powered clinical note generation and medical scribes
- AI diagnostic tools analyzing patient imaging, labs, or records
- NLP tools processing clinical transcripts
- AI chatbots interacting with patients about health conditions
- Voice recognition and transcription services
- AI-based clinical decision support systems

KEY REGULATORY DEVELOPMENTS (2025):

On January 6, 2025, HHS OCR proposed the first major update to the HIPAA
Security Rule in 20 years. Key changes include:

- ELIMINATION of the distinction between "required" and "addressable"
  implementation specifications. ALL safeguards become mandatory.
- MANDATORY multi-factor authentication (MFA) for all systems accessing ePHI.
- MANDATORY vulnerability scanning at least every six months.
- MANDATORY penetration testing at least annually.
- MANDATORY patch management: critical patches within 15 days, high-risk
  patches within 30 days.
- MANDATORY 24-hour notification from business associates to covered entities
  upon activation of contingency plans.

On January 10, 2025, OCR issued a "Dear Colleagues" letter confirming it
will enforce Section 1557 nondiscrimination protections as applied to AI
use in healthcare (effective May 1, 2025).

========================================================================
SECTION 2: PENALTY TIERS AND REAL ENFORCEMENT CASES
========================================================================

CURRENT HIPAA CIVIL PENALTY TIERS (2025-2026, inflation-adjusted):

Tier 1 - Did Not Know:       $145 - $73,011 per violation
Tier 2 - Reasonable Cause:   $1,461 - $73,011 per violation
Tier 3 - Willful Neglect,
         Corrected:          $14,602 - $73,011 per violation
Tier 4 - Willful Neglect,
         NOT Corrected:      $73,011 - $2,190,294 per violation

Annual Cap: $2,190,294 per tier
Criminal penalties: up to $250,000 and 10 years imprisonment.

REAL ENFORCEMENT CASES INVOLVING TECHNOLOGY:

Blackbaud, Inc. (2023-2024)
- Ransomware attack on cloud computing vendor serving healthcare
- Fine: $49.5 million multistate + $6.75 million California settlement

Advocate Aurora Health (2024)
- Meta Pixel tracking code transmitted PHI to Facebook without consent
- Fine: $12.25 million settlement; 3 million patients affected

Mass General Brigham (2024)
- Cookies and analytics tools shared patient data with third parties
- Fine: $18.4 million class action settlement

BetterHelp (2023)
- Telehealth platform shared mental health data via tracking technologies
- Fine: $7.8 million FTC settlement

GoodRx (2023)
- Health app exposed prescription data through third-party trackers
- Fine: $25 million class action settlement

Montefiore Medical Center (2024)
- Employee stole and sold patient data; inadequate access controls
- Fine: $4.75 million; 12,517 individuals affected

TOTAL from tracking technology violations alone: Over $100 million in
combined penalties 2023-2025. OCR wrote to nearly 130 healthcare
organizations in July 2023 warning about tracking technology risks.

========================================================================
SECTION 3: BUSINESS ASSOCIATE AGREEMENT (BAA) REQUIREMENTS
========================================================================

Any AI vendor handling PHI is a Business Associate. A BAA is LEGALLY
REQUIRED before any PHI is shared.

Regulatory basis: 45 CFR 164.314(a), 45 CFR 164.504(e), 45 CFR 164.308(b)(1)

MANDATORY BAA PROVISIONS:

[ ] Permitted and Required Uses/Disclosures
[ ] Safeguard Requirements (BA must comply with Security Rule)
[ ] Subcontractor Requirements (all sub-BAs must also have BAAs)
[ ] Breach Notification (current: 60 days; proposed: 24 hours)
[ ] Individual Rights Support (access, amendment, accounting)
[ ] HHS Access (BA must allow HHS compliance reviews)
[ ] Return or Destruction of PHI upon termination
[ ] Term and Termination provisions

AI-SPECIFIC BAA PROVISIONS YOU SHOULD ADD:

[ ] PHI must NOT be used for model training without authorization
[ ] PHI must NOT be aggregated with other customers' data
[ ] PHI must NOT be retained beyond the service period
[ ] Specify data residency (geographic restrictions)
[ ] Right to inspect vendor's security practices
[ ] AI-specific incident response (hallucinations, model errors)
[ ] Subcontractor chain fully documented (cloud providers, APIs)

VENDORS CURRENTLY OFFERING BAAs FOR AI SERVICES:
- OpenAI (API/Enterprise - must be requested)
- Microsoft Azure (Azure OpenAI Service)
- Google Cloud (Vertex AI)
- Amazon Web Services (Amazon Bedrock)

WARNING: A vendor offering a BAA does NOT mean they are HIPAA compliant.
The BAA is a legal agreement, not a certification. Due diligence required.

========================================================================
SECTION 4: TECHNICAL SAFEGUARD REQUIREMENTS
========================================================================

Per 45 CFR 164.312 and proposed 2025 Security Rule updates:

ENCRYPTION:
[ ] Data at Rest: AES-256 encryption minimum
[ ] Data in Transit: TLS 1.2 or TLS 1.3 (TLS 1.0/1.1 unacceptable)
[ ] Perfect Forward Secrecy (PFS) enabled
[ ] Key management and rotation policies documented

SAFE HARBOR: If ePHI is encrypted per NIST standards and a breach occurs,
breach notification may not apply (NIST SP 800-111, 800-52).

ACCESS CONTROLS:
[ ] Unique User Identification - no shared accounts (164.312(a)(2)(i))
[ ] Multi-Factor Authentication for all users (proposed 2025 rule)
[ ] Role-Based Access Control with least privilege
[ ] Automatic Logoff after inactivity (164.312(a)(2)(iii))
[ ] Emergency Access Procedures documented (164.312(a)(2)(ii))

AUDIT CONTROLS:
[ ] Log who accessed data, when, what, and actions taken
[ ] Logs tamper-resistant and retained per policy
[ ] Regular review of audit logs
[ ] For AI: log all queries containing PHI and all AI outputs with PHI

VULNERABILITY MANAGEMENT (Proposed 2025 Rule):
[ ] Vulnerability scanning every 6 months minimum
[ ] Penetration testing annually minimum
[ ] Critical patches deployed within 15 days
[ ] High-risk patches deployed within 30 days

========================================================================
SECTION 5: QUESTIONS TO ASK AI VENDORS BEFORE SIGNING
========================================================================

DATA HANDLING:
1. Exactly what PHI will your system access, process, and store?
2. Where is PHI stored geographically? Any data outside the US?
3. How is PHI encrypted at rest and in transit? Specific algorithms?
4. Do you use PHI to train or improve your AI models?
5. What happens to our PHI if we terminate the contract?
6. What is your data retention policy?

SECURITY AND COMPLIANCE:
7. Will you sign a BAA?
8. Do you have a current SOC 2 Type II report or HITRUST certification?
9. When was your last independent security audit and penetration test?
10. Do you support MFA for all users?
11. What is your patch management process and timeline?
12. Do you have a dedicated security team?

INCIDENT RESPONSE:
13. What is your breach notification timeline?
14. Do you have a documented incident response plan? Last tested?
15. What was the last security incident you experienced?
16. Do you carry cyber liability insurance? Coverage limits?

AI-SPECIFIC:
17. How do you prevent AI hallucinations generating false patient info?
18. What safeguards prevent the AI from memorizing patient data?
19. How do you validate AI outputs for clinical accuracy?
20. What is your process for AI bias/discrimination? (Section 1557)
21. Can you provide audit logs of all AI interactions involving our PHI?
22. If you use third-party AI models (OpenAI, Anthropic, Google), do
    you have BAAs with those providers?

SUBCONTRACTORS:
23. What cloud provider(s) do you use? (AWS, Azure, GCP?)
24. Do you have BAAs with all infrastructure providers?
25. What other subcontractors have access to our PHI?

RED FLAGS - WALK AWAY IF:
- Vendor refuses to sign a BAA
- Cannot produce a recent independent security audit
- Uses PHI for model training without clear de-identification
- Cannot explain encryption standards
- Stores PHI outside the US without disclosure
- Has no documented incident response plan
- Cannot identify subcontractors that access PHI

========================================================================
SECTION 6: COMPLIANCE CHECKLIST SUMMARY
========================================================================

BEFORE ADOPTING ANY AI TOOL:
[ ] Determine if tool creates, receives, maintains, or transmits ePHI
[ ] Conduct vendor due diligence
[ ] Execute BAA with all required provisions
[ ] Update organization's risk analysis to include the new AI tool
[ ] Document evaluation and decision-making process

TECHNICAL REQUIREMENTS:
[ ] AES-256 encryption at rest
[ ] TLS 1.2+ encryption in transit with PFS
[ ] Multi-factor authentication for all users
[ ] Unique user identification (no shared accounts)
[ ] Role-based access control with least privilege
[ ] Automatic session logoff
[ ] Comprehensive audit logging
[ ] Vulnerability scanning every 6 months
[ ] Penetration testing annually
[ ] Critical patches within 15 days

ADMINISTRATIVE REQUIREMENTS:
[ ] Designated security officer for AI tool compliance
[ ] Written policies for AI tool use with PHI
[ ] Workforce training on AI tool use and PHI handling
[ ] Incident response plan covering AI-specific scenarios
[ ] Regular risk assessments (at least annually)

PRIVACY REQUIREMENTS:
[ ] Minimum necessary standard applied (45 CFR 164.502(b))
[ ] Notice of Privacy Practices updated to reflect AI use
[ ] Patient right of access preserved for AI-generated records
[ ] Accounting of disclosures includes AI vendor disclosures

AI-SPECIFIC REQUIREMENTS:
[ ] PHI not used for model training without authorization
[ ] De-identification verified per Safe Harbor or Expert Determination
[ ] AI bias and discrimination risk assessed (Section 1557)
[ ] AI-generated content distinguished from source clinical data
[ ] Subcontractor chain fully documented with BAAs at every level
[ ] Data deletion/return procedures defined for contract termination

========================================================================
SECTION 7: KEY REGULATORY REFERENCES
========================================================================

HIPAA Privacy Rule:
  45 CFR 164.502    Uses and disclosures
  45 CFR 164.504(e) BAA requirements
  45 CFR 164.514    De-identification standards
  45 CFR 164.524    Individual right of access
  45 CFR 164.530(j) Documentation retention (6 years)

HIPAA Security Rule:
  45 CFR 164.308    Administrative safeguards
  45 CFR 164.310    Physical safeguards
  45 CFR 164.312    Technical safeguards
  45 CFR 164.314    Organizational requirements (BAA standards)

Breach Notification Rule:
  45 CFR 164.404    Notification to individuals (60 days)
  45 CFR 164.410    BA notification to covered entity

Other:
  HITECH Act, Section 13401 - Security provisions for BAs
  Section 1557 of the ACA   - Nondiscrimination, including AI
  NIST SP 800-111           - Storage Encryption Technologies
  NIST SP 800-52            - TLS Implementation Guidelines

DISCLAIMER: This guide is for informational purposes only and does not
constitute legal advice. Consult qualified legal counsel for advice
tailored to your organization's specific circumstances.

========================================================================
Prepared by ReasonNotes | reasonnotes.com
========================================================================