Does the Act cover health inferences produced by AI?

Section 010 includes certain information that a regulated entity, small business or processor derives or extrapolates by algorithms or machine learning to associate or identify a consumer with covered health information. Counsel should map the exact input, inference, consumer link and exemption facts rather than treating every model output as automatically covered.

Does the Act require a log for every model request?

Chapter 19.373 specifies privacy-policy, consent, rights, security, processor, sale and geofence duties. It lacks a universal command to retain every model request. Per-decision records are an implementation pattern that can corroborate data class, purpose, destination, policy and outcome, subject to the organization’s necessity and retention analysis.

Can one consent cover collection and sharing?

Section 030 says consent for sharing must be separate and distinct from consent for collection when consent supplies the legal route. The statute also contains a branch for activity necessary to provide a product or service requested by the consumer. Preserve the branch analysis and the exact choice presented.

What evidence supports a deletion response?

Keep the authenticated request, search method, responsive records, downstream notices, deletion confirmations, backup handling and response. Runtime model records help locate disclosed events when they carry a stable subject reference. Each underlying store still needs its own execution proof.

Where does HIPAA fit?

Section 100 contains data-level and entity-related exemptions, including specified protected health information and other health-care information. The Washington Attorney General describes the Act as addressing consumer health data outside HIPAA’s familiar coverage. Legal should test each dataset and entity against the statutory exemption text.

Washington My Health My Data AI Audit Evidence: Reconstruct Model Disclosures

A symptom checker can assemble age, location and a description of pain into one model request. The resulting evidence problem begins before the HTTPS call leaves the application: whose data was involved, what health status the system inferred, which purpose applied, which provider received it and which consumer choice authorized the flow.

Washington’s My Health My Data Act, Chapter 19.373 RCW, supplies the legal baseline. The statute’s main regulated-entity provisions applied on 31 March 2024, with a 30 June 2024 date for small businesses. This Washington My Health My Data AI audit evidence guide maps the current text to artifacts a reviewer can retrieve and test.

Pin scope to the consumer, entity and data

Section 010 defines a consumer as a Washington resident or a natural person whose consumer health data is collected in Washington, acting in an individual or household context. Employment-context individuals fall outside that definition. The same section defines consumer health data as personal information linked or reasonably linkable to a consumer that identifies past, present or future physical or mental health status.

The definition reaches beyond a medical chart. Its examples include diagnoses, treatment, medication, reproductive or sexual health information, biometric and genetic data, certain precise location information, and data identifying someone seeking health care services. It also includes information processed to associate a consumer with covered health information when derived or extrapolated by algorithms or machine learning.

Start with a scope sheet naming the legal entity, service, consumer connection, input fields, inferred outputs, use purpose, exemptions considered and approving counsel. Keep the model version and prompt template beside that analysis. A white inventory row labeled triage-assistant-07 should point to the exact inference tested, rather than the broad label “wellness data.”

Join the privacy policy to production behavior

Section 020 requires a consumer health data privacy policy that discloses collected categories and purposes, source categories, shared categories, recipient categories and specific affiliates, plus the method for exercising consumer rights. A link must appear prominently on the homepage. Additional undisclosed categories or purposes require disclosure and affirmative consent before collection, use or sharing.

Build a policy exhibit with the rendered page, publication timestamp, version history and approval. Then reconcile its claims against the deployed AI route. The operating export should identify the feature, payload classification, declared purpose, provider endpoint, affiliate or third-party category, policy version and outcome.

That comparison catches a common evidence defect: procurement names one hosted model while a later application release sends traffic to a second endpoint. The privacy policy is legal and product evidence. Observed route records show what happened on selected HTTP calls. A defensible package keeps both, joined through stable service and release identifiers.

Preserve collection and sharing consent separately

Section 030 permits collection with consumer consent for a specified purpose or to the extent necessary to provide a requested product or service. Sharing follows a separate branch. Consent for sharing must be separate and distinct from collection consent, unless sharing is necessary to provide the requested product or service. A consent request must disclose covered data categories, purpose and specific uses, recipient categories, and the withdrawal method.

For each AI flow, preserve the screen shown, consent text, consumer action, timestamp, purpose, data categories, sharing recipients, withdrawal state and release version. Keep the necessity analysis when the organization relies on the requested-service branch. Legal owns that conclusion.

My view is that a single database flag called health_consent=true should fail audit review. It erases the distinction between collection and sharing, hides the approved purpose, and says nothing about the text presented to the consumer. A selected model event should join to the exact consent artifact through a consent receipt ID supplied by the application.

Reconstruct access, withdrawal and deletion requests

Section 040 gives consumers rights concerning collection, sharing and sale, including access to covered data and information about recipients. It also addresses withdrawal of consent and deletion. When a deletion request reaches archived or backup systems, deletion can be delayed until restoration, subject to a six-month limit after authentication under the statutory conditions.

Create a rights case file containing intake, identity authentication, scope query, responsive data, recipient list, decision, response, downstream notices, deletion execution and appeal handling. If consumer health data describes someone other than the caller, retrieval needs a stable data-subject reference. Caller identity alone gives a weak search key.

HTTP model records can identify sampled disclosures and destinations when the application supplies that subject reference. They cannot delete source records, vector indexes, provider-held copies or backups. Preserve execution evidence from each store owner and link it to the request ID. A second reviewer should be able to repeat the search and explain every exclusion.

Test processors and security as operating controls

Section 050 restricts employee, processor and contractor access to what is necessary for consented purposes or a requested product or service. It also requires administrative, technical and physical security practices meeting a reasonable industry standard of care, appropriate to the volume and nature of the data. Processor duties in Section 060 require a binding contract and processing consistent with its instructions.

Retain processor terms, instructions, access roles, approved endpoints, classification policy, security tests, incidents and remediation. Compare provider behavior documented in the contract with the route used by a production sample. An independent provider purpose or out-of-contract processing can change the legal analysis, so procurement language and actual use both matter.

Run three synthetic tests under one change ticket: a health marker sent to an unapproved endpoint, missing identity context, and a withdrawn consent receipt. The evidence should show the configured outcome, policy version, alert and clean retest. These tests are implementation evidence. Washington’s statute sets the access, contract and security duties without prescribing a universal per-request log format.

Package the evidence for repeatable review

Section 070 requires a separate valid authorization before sale and specifies its contents, defects and handling. The geofence restrictions in Section 080 prohibit specified uses around entities providing in-person health care. Section 090 applies Washington’s Consumer Protection Act to violations. Those branches deserve explicit rows even when a service concludes that no sale or geofence occurs.

Use one evidence index:

Scope: entity, consumer connection, covered data, inference and exemption analysis.
Transparency: policy version, homepage placement, categories, purposes and recipients.
Choice: collection consent, separate sharing consent, withdrawal and necessity analysis.
Rights: authenticated request, search query, recipient notices, deletion and appeal.
Operations: processor terms, access controls, route samples, tests and remediation.
Special branches: sale authorization, geofence review and Consumer Protection Act response.

Hash exports, save collection queries and record the reviewer. The Washington compliance checklist turns these folders into owner-based completion tests.

DeepInspect

DeepInspect supplies the HTTP operating-evidence layer for authenticated users or agents calling LLM endpoints. It evaluates application-supplied identity, purpose and consent context, data classification, destination and policy before forwarding a request, then produces a signed, tamper-evident decision record outside the calling application’s write path.

Those records can reconcile approved providers with observed routes, retrieve the policy used for triage-assistant-07, scope disclosures and preserve synthetic control tests. DeepInspect leaves statutory scope, privacy-policy wording, consent presentation, consumer authentication, deletion across data stores, processor contracts, sale authorization and geofence analysis with their named owners. Book a technical deep dive at deepinspect.ai.