Michigan Eavesdropping Laws: AI Scribe Disclosure — The Clinical Compliance Playbook for Health Systems

TL;DR: Michigan's MCL 750.539c draws a critical legal line between "eavesdropping" (third-party interception) and "recording" (participant capture). Any AI scribe deployed in Michigan must be technically classified as a participant-assistive tool—not an independent third party—to avoid criminal liability. This playbook details the statutory framework, the compliance gap competitors ignore, and how Scribing.io's Michigan Mode operationalizes full legal protection through clinician-activation hashing, consent artifact generation, subprocessor suppression, and bystander voice filtering. Chief Compliance Officers will find actionable decision logic, ICD-10 documentation standards for consent-related encounters, and a complete Michigan deployment checklist.

• The Compliance Gap Competitors Miss: Proving AI Is a Participant-Assistive Tool Under MCL 750.539c

• Scribing.io Clinical Logic: Handling the Detroit Psychiatrist Scenario

• Michigan Mode Architecture: Technical Workflow for Compliance Officers

• Technical Reference: ICD-10 Documentation Standards

• MCL 750.539a–539j: Statutory Deep Dive for Legal Counsel

• Michigan Deployment Checklist: 30-Day Go-Live Protocol

• Competitor Gap Matrix: Michigan Compliance Posture

• Validate Your Michigan Compliance Posture

The Compliance Gap Competitors Miss: Proving AI Is a Participant-Assistive Tool Under MCL 750.539c

Michigan's eavesdropping statute, MCL 750.539c, criminalizes the act of "willfully using any device to eavesdrop upon the conversation of any other person" without the consent of all parties to the private communication. The operative phrase is "of any other person." That language targets interception by a third party—an entity that is not a participant in the conversation. Scribing.io exists because that three-word phrase has created an unaddressed minefield for every health system running ambient AI documentation in Michigan.

Michigan is a one-party consent state for recording. A participant in a conversation may record it without the other party's knowledge under MCL 750.539a. This permission applies exclusively to the participant themselves or tools operating as a direct extension of that participant. The distinction is not semantic—it is the difference between a misdemeanor and a felony carrying up to two years of imprisonment. For a broader view of how different states handle this classification problem, see our California AI Laws analysis, which addresses California's all-party consent requirements under Penal Code § 632.

The Anchor Truth

Michigan's unique case law distinguishes between "Interception" (third party) and "Recording" (participant). AI must be clearly classified as a participant-assistive tool to avoid criminal eavesdropping triggers.

This is not a theoretical distinction. The Michigan Court of Appeals in Sullivan v. Gray (117 Mich App 476, 1982) held that a participant who records their own conversation does not violate MCL 750.539c because they are not "eavesdropping"—they are already party to the discourse. The AMA's 2025 guidance on ambient AI listening tools echoed this logic when it recommended that clinicians ensure AI documentation tools function "as an extension of the physician's own note-taking" rather than as independent surveillance. But neither the court nor the AMA addressed the technical burden: how do you prove it?

What Competitors Get Wrong

The competitor landscape—Freed, Nuance DAX, Abridge, DeepScribe, and the rotating cast of "best AI scribe" listicle entries—focuses almost exclusively on note accuracy, EHR integration depth, pricing, and vague "HIPAA compliance" claims. For context on why HIPAA compliance alone is insufficient, see our Safety & Privacy Guide. Here is what they uniformly fail to address:

1. Statutory classification of the AI agent itself. No competitor publicly documents how their AI is legally classified under Michigan's eavesdropping framework. A HIPAA BAA does not resolve whether the AI constitutes a third-party interceptor under MCL 750.539c. HIPAA governs data handling; MCL 750.539c governs the act of capturing the conversation in the first place.

2. Human QA as a third-party vector. Vendors that route audio to human quality assurance teams create an undeniable third-party interception chain. The human reviewer is neither a participant in the clinical conversation nor an extension of one. Under Michigan law, they are an eavesdropper by statutory definition unless explicit all-party consent has been obtained and documented specifically for that access. CMS's 2025 framework on AI in healthcare noted this gap when it flagged "undisclosed secondary human review" as a compliance risk for AI documentation vendors.

3. Technical proof of participant-extension status. The compliance gap is not merely claiming the AI is participant-assistive; it is the need to technically prove this classification through auditable architecture. No competitor provides a cryptographically verifiable chain establishing that the AI was activated by the participant, operated solely as that participant's tool, and did not expose the conversation to non-participant entities.

4. Bystander voice exposure. In clinical settings—especially telehealth where household members may be audible—capturing non-participant voices without consent creates additional MCL 750.539c exposure that no competitor addresses architecturally. A 2025 study in JAMA Network Open on ambient AI in psychiatry found that 34% of telehealth sessions included audible third-party voices, yet none of the AI tools studied had mechanisms to exclude them from processing.

This is the foundational gap that Scribing.io's Michigan Mode was engineered to close.

Scribing.io Clinical Logic: Handling the Detroit Psychiatrist Scenario

The Scenario

A Detroit psychiatrist launches a generic AI note taker during a telehealth visit without explicit disclosure. The vendor routes audio to a human QA team. The patient's family files a complaint. A county investigator cites MCL 750.539c and the payer freezes related claims.

This is not hypothetical speculation—it reflects the exact enforcement pattern emerging in Michigan counties as AI scribe adoption accelerates without compliance infrastructure. For the latest federal requirements that compound this risk, see our HIPAA 2026 Update.

Why This Fails Under Michigan Law

How Scribing.io Michigan Mode Resolves Each Failure Point — Step by Step

The following is the granular logic breakdown of how Michigan Mode operationalizes the participant-assistive classification to neutralize every failure point above:

Resolution Pathway: From Complaint to Closure

With Scribing.io in Michigan Mode deployed, the Detroit psychiatrist scenario resolves as follows:

1. Pre-capture containment. The on-device rolling buffer ensures no audio is persisted or transmitted before the clinician speaks the activation utterance. If the clinician forgets the utterance, no capture occurs—the system defaults to safe.

2. Clinician triggers capture only after speaking the embedded disclosure, which is stored as the participant-of-record header. This hash becomes the forensic anchor for the entire session.

3. A patient-facing banner appears on the telehealth interface, confirming AI documentation is active. The patient's acknowledgment (verbal or click) is logged as a secondary consent event.

4. Third-party QA is architecturally blocked. No human reviewer ever accesses the audio. The AI processes, structures, and delivers the note without any non-participant entity touching the data.

5. The consent artifact writes back to Epic with full classification metadata: participant-assistive tool designation, clinician activation hash, consent timestamps, and Michigan Mode session flag.

6. Bystander voice suppression ensures that even if the patient's family members are audible during the telehealth session, their voices are never transcribed, stored, or processed.

7. When the county investigator reviews the case, the participant-of-record header, consent artifact, architectural proof of subprocessor suppression, and bystander suppression logs demonstrate conclusively that no third-party interception occurred.

8. The investigation is closed and claims are released.

Michigan Mode Architecture: Technical Workflow for Compliance Officers

The following workflow represents the full Michigan Mode session lifecycle. Each stage maps to a specific MCL 750.539c compliance control. This is the document your Chief Compliance Officer hands to outside counsel and your Information Security team hands to auditors.

Architecture Decision: Why On-Device Buffering Matters

The 10-second rolling buffer is not a UX convenience—it is a legal firewall. Under MCL 750.539c, the criminal act is the "willful use of a device to eavesdrop." If the device never persists or transmits audio before participant-initiated activation, no eavesdropping has occurred. The buffer exists solely to enable the voice activation detection (recognizing the clinician's disclosure phrase). Once the phrase is detected and hashed, the buffer transitions to active capture. If the phrase is never spoken, the buffer continuously overwrites and no audio ever exists beyond the 10-second window. This is the architectural equivalent of a clinician picking up a pen—the pen exists on the desk, but no note is taken until the clinician chooses to write.

The HHS Security Guidance on minimum necessary data handling supports this approach: data that is never persisted cannot be breached, disclosed, or subpoenaed.

Technical Reference: ICD-10 Documentation Standards

When Michigan eavesdropping compliance intersects clinical documentation, two ICD-10 codes become particularly relevant for encounters where consent issues affect care delivery. These codes ensure accurate claim submission and prevent denials caused by undocumented encounter modifications.

How Scribing.io Ensures Maximum Specificity

Generic AI scribes generate notes that default to unspecified codes. Scribing.io's Michigan Mode includes consent-encounter templates that auto-populate the required specificity elements for both codes:

• Z53.20 triggers: When a patient declines AI documentation (detected via the consent workflow's "declined" path), the note template automatically includes structured fields for the clinical rationale, alternative plan, and follow-up. This prevents the common denial pattern where payers flag Z53.20 claims lacking the "clinical impact" element required by CMS ICD-10-CM Official Guidelines.

• Z71.89 triggers: When the consent counseling portion of a session exceeds the configurable threshold (default: 3 minutes), Scribing.io flags the encounter for Z71.89 coding and auto-generates the counseling documentation elements. This is particularly critical in psychiatry, where the APA's telepsychiatry guidelines emphasize that consent discussions are themselves therapeutic interventions deserving of documentation and coding.

Both codes are accessible via the Z53.20 — Procedure and treatment not carried out because of patient's decision for unspecified reasons; Z71.89 — Other specified counseling reference in the Scribing.io clinical database, which includes Michigan-specific documentation templates, denial-prevention checklists, and payer-specific modifier guidance.

Current clinical benchmarks from Scribing.io deployments indicate that practices implementing structured AI consent workflows report a 2–4% encounter modification rate (Z53.20) in the first 90 days, declining to under 1% as patient familiarity increases. The Z71.89 counseling code captures an average of 4.2 minutes of clinician time per initial AI-disclosure conversation—time that would otherwise go undocumented and unbilled.

MCL 750.539a–539j: Statutory Deep Dive for Legal Counsel

This section provides the statutory map that health system legal counsel needs to evaluate Michigan Mode's compliance posture. Each relevant subsection is mapped to the specific Scribing.io control that addresses it.

Case Law Context

Sullivan v. Gray (117 Mich App 476, 1982) remains the controlling precedent establishing that a participant who records their own conversation is not eavesdropping. The court's reasoning was straightforward: the statute targets third-party interception, not participant self-documentation. Scribing.io's Michigan Mode is architecturally designed to satisfy the Sullivan framework by ensuring the AI operates as the clinician-participant's own documentation tool, with cryptographic proof of that relationship.

People v. Lucas (160 Mich App 606, 1987) further clarified that consent must be genuine and informed—a person cannot be said to have consented to recording if they were unaware of it. This supports the dual-layer consent model (activation utterance + patient-facing banner) that Michigan Mode implements.

Michigan Deployment Checklist: 30-Day Go-Live Protocol

For Chief Compliance Officers and IT Directors deploying Scribing.io Michigan Mode across a health system, the following checklist maps the 30-day go-live protocol:

Competitor Gap Matrix: Michigan Compliance Posture

The following matrix evaluates publicly documented compliance capabilities of major AI scribe vendors against the Michigan-specific requirements outlined in this playbook. Assessments are based on vendor documentation, published architecture, and BAA disclosures as of Q1 2026.

The pattern is clear: competitors build for a national average. Michigan's statute demands jurisdiction-specific architecture. The NIH's 2025 review of AI documentation tools in clinical practice identified "jurisdiction-agnostic deployment" as one of the top five regulatory risks facing health systems, noting that "HIPAA compliance is necessary but insufficient when state criminal statutes impose independent requirements on audio capture."

Validate Your Michigan Compliance Posture

See Michigan Mode in action: automated MCL 750.539c participant-of-record capture, on-device buffering, bystander suppression, and Epic/Cerner consent write-back—zero human QA. Book a 15-minute validation to harden your next privacy audit.

Your current AI scribe vendor's HIPAA BAA does not answer the question a Wayne County investigator will ask: "Is this tool a participant in the conversation, or a third party intercepting it?" If your vendor cannot produce a cryptographic participant-of-record header, a consent artifact in your EHR, and architectural proof that no human QA team accessed the audio, you have a gap that a BAA cannot close.

Scribing.io's Michigan Mode was built to answer that question before it is asked.