Introducing – The AI Scribe

There are many misconceptions about ambient documentation, also known as the AI Scribe. One common misunderstanding is that the AI automatically “stops listening” when the conversation shifts to non-medical topics, like a patient talking about their kids or a recent Disney trip. In reality, the AI hears the entire conversation. It is designed to prioritize medically relevant information, but it does not always perfectly ignore casual or social discussion. This is actually one way AI hallucinations can occur.

For example, if a patient says, “We’ve been exhausted with the kids lately,” the AI might interpret that as clinical information and document it as, “patient reports fatigue and increased stress at home.” While that may sound reasonable, it was never intended as a medical symptom. This is a classic example of an AI hallucination—but we’ll explore that more in a moment. First, let’s take a closer look at what an AI scribe actually is.

1: Patient Intake & Consent

• Patient informed of AI-assisted documentation
• Verbal or written consent obtained (see section below)

2: Ambient Capture

• Audio recorded during visit (or real-time transcription)
• Some tools do NOT store audio (important compliance distinction)

3: AI Draft Generation

• AI generates:
  • HPI
  • ROS
  • Exam (often templated)
  • MDM narrative

4: Provider Review & Editing

• Provider must:
  • Review entire note
  • Edit for accuracy
  • Remove hallucinations or fabricated elements
  • Ensure medical necessity is supported

5: Provider Attestation

Typical compliant statement:

“I personally performed the services described and have reviewed and verified that this documentation accurately reflects the encounter.”

6: Finalization in EHR

• Note locked
• Audit trail preserved

This workflow may seem like a great time-saver and a benefit for both providers and patients. However, it also comes with important risks—similar to those we’ve seen with copy/paste, copy-forward, templates, and cloned notes. These risks arise when documentation is not carefully reviewed and edited to reflect the specific patient details of the individual encounter that occurred that day.

Patient Consent

We can’t talk about AI scribes without addressing the importance of protecting patient information, or Protected Health Information (PHI). Most ambient documentation vendors operate under a Business Associate Agreement (BAA), which means they are required to follow HIPAA privacy and security rules—just like any other documentation tool. In other words, the same expectations that apply to traditional documentation also apply to AI-generated notes.

When using an AI scribe, it’s important to understand how patient information is handled. Key questions to consider include: (1) Is the audio from the visit being stored? (2) Is the data being used to train the AI model? And if so, (3) is that data properly de-identified before it is reused?

There are also some important red flags to watch for.

• “Free AI scribe” tools without BAA
• Data used to train external models without disclosure
• No clear retention/deletion policy
• Copy-forward or hallucinated content introduced into note

AI hallucinations are clinically plausible but incorrect details generated by AI that were never obtained, observed, or performed during the visit. We’ll cover this in more detail in the next article, including how AI hallucinations can differ from real clinical documentation in the medical record.

Remember, providers must:

• Review everything
• Correct errors
• Ensure medical necessity
• Only bill for what they actually did