Keeping the clinician responsible

About clinicgpt.ai: Educational page only. The SOAP demo is browser-side formatting, not a clinical note system. No production product, BAA, or coding tool is offered here. Not medical, legal, coding, or malpractice advice.

Author-of-record principle

Whoever signs the note owns its clinical and billing meaning. Ambient tools, human scribes, speech-to-text, and template macros change how text is drafted. They do not move legal or professional responsibility onto the software vendor, the model provider, or “the AI.”

That principle is older than large language models. It is the same reason a physician who co-signs a resident note remains accountable for the content they accept. Medicare program documentation rules and medical board expectations consistently treat the billing/signing clinician as responsible for the medical record used to support care and payment. For Medicare physician payment context, use CMS Physician Fee Schedule materials (as of 2026-07-21) and your MAC’s local documentation guidance — not a vendor slide deck.

Review-before-sign (non-negotiable)

A workable AI documentation policy is short:

1. No silent auto-sign. Drafts never become the record without an explicit clinician action.
2. Edit is expected. A non-zero edit rate is normal; a claimed “zero edit” product is a red flag.
3. Time-based services need human time. If you bill by time (for example common psychotherapy code 90837, or E/M levels selected by time such as 99213 / 99214 when time is the basis), documented time must reflect clinician time, not model runtime. (Code numbers only; see CMS E/M materials — we do not reproduce AMA CPT descriptors.)
4. Keep a trail when you can. Many organizations retain draft vs final for audit readiness. Whether your vendor supports that is a buying criterion for real products, not something this demo provides.

Copy-forward and hallucination risks

AI introduces failure modes that look different from tired late-night typing:

Hallucinated positives and negatives. “Patient denies chest pain” appears because the model expects HPI structure, not because the question was asked.

Medication and dose invention. Phonetic ASR errors become confident drug names; the LLM “helps” by completing a regimen.

Problem-list resurrection. Old, resolved, or never-confirmed problems re-enter the assessment because prior context was retrieved carelessly.

Copy-forward pollution. If the model or EHR carries forward yesterday’s plan without reconciliation, the chart drifts from reality faster than manual copy-forward alone.

Mitigations that actually work: force a structured review checklist for high-risk fields (meds, allergies, plan, follow-up); train clinicians to delete confidently wrong text rather than lightly edit around it; measure omission with dual review on a sample of notes.

How / whether to disclose AI assistance

Requirements vary by state board, payer manual, employer policy, and malpractice carrier — and they change. There is no single national checkbox that satisfies every context forever. What is stable:

• Honesty inside the organization. Staff should know when drafts are machine-assisted.
• Patient-facing transparency when policy requires it. Some organizations disclose ambient recording and AI drafting in consent materials.
• Avoid deceptive notes. Presenting machine-invented exam findings as personally performed examinations is a professional integrity problem regardless of AI branding.

A conservative footer some clinics adopt (adapt with counsel):

Portions of this note were drafted with AI assistance, then reviewed, edited, and signed by [Name, credentials]. Time documentation reflects clinician time only.

For a longer discussion aimed at charting practice, see the blog post Documenting AI assistance in chart notes. Treat that post as educational commentary; always verify current board and payer rules for your locations.

Coding and attestation implications (numbers only)

Evaluation and management (E/M) services billed to Medicare (for example office/outpatient codes in the 99202–99215 range) are paid under program rules that depend on medical necessity and documentation supporting the billed level — whether selected by medical decision making or time, as applicable under current CMS policy (see CMS PFS resources, as of 2026-07-21).

Implications for AI drafts:

• The attestation is yours. Signing means you accept the content for clinical and billing purposes.
• Code suggestions from software are not authority. Do not paste AMA CPT descriptor text from anywhere into policies as if it were free to reproduce; use code numbers and official program guidance.
• This site makes no coding-accuracy claim for any tool, including ambient vendors or our demo (which proposes no codes).

Medico-legal basics (plain language)

From a risk perspective, AI-assisted notes fail the same ways human notes fail — plus a few new ways:

• Wrong facts that look fluent are more dangerous than awkward but honest sparse notes.
• Missing decision rationale still weakens defense of care.
• If recording is used, consent and retention policies matter independently of the note.

Carriers and boards increasingly ask whether AI is in the workflow. Having a written policy (review-before-sign, who may use tools, what is logged) is more important than which brand you trial.

What this means on clinicgpt.ai

The demo cannot keep you “compliant.” It cannot attest, code, or retain drafts. It exists to show structure and limits. For HIPAA vendor boundaries, read AI scribes and HIPAA. For category mechanics, read what ambient AI scribes are.

Sources (as of 2026-07-21)

• CMS — Physician Fee Schedule — Medicare physician payment and documentation program context (no CPT descriptor reproduction).
• HHS — HIPAA for Professionals — Privacy/Security context when vendors handle PHI.
• AMA principles of medical record documentation — refer to current AMA professional resources for medical records ethics; do not copy CPT descriptor language from CPT products.