AI-Generated Clinical Notes: Promise, Peril, and What We're Getting Wrong

The pitch is seductive: an AI listens to your patient encounter, generates a complete clinical note, and you just review and sign. No more hours of charting after dinner. No more choosing between documentation quality and time with patients.

I've been piloting ambient AI documentation at my hospital for six months. The technology is genuinely impressive — and genuinely concerning. We need to have a more honest conversation about both sides.

What's Working

The time savings are real. Our pilot physicians are spending 30-45 fewer minutes per day on documentation. That's not trivial. For a hospitalist seeing 16-18 patients a day, that's the difference between finishing notes at 5 PM and finishing at 7 PM.

The notes are surprisingly comprehensive. The AI captures details that busy physicians often omit — exact phrases the patient used, the sequence of the review of systems, specific counseling provided. In many cases, the AI-generated note is more thorough than what the physician would have written.

What's Concerning

Here's the problem nobody wants to talk about: the AI sometimes gets things wrong, and the errors are subtle enough that a tired physician doing a quick review might miss them.

In our pilot, we found:

• The AI occasionally attributes symptoms to the wrong body system
• Negatives sometimes become positives ("patient denies chest pain" becomes "patient reports chest pain") — rare but devastating when it happens
• The AI fills in "expected" physical exam findings that weren't actually performed or stated
• Family history and social history can bleed between encounters if the AI model has context from prior visits

Each of these errors, standing alone, seems minor. But a clinical note is a legal document. It drives billing, informs other providers, and can be subpoenaed in litigation. A hallucinated physical exam finding isn't just an inaccuracy — it's a potential malpractice issue.

The Review Problem

The most insidious risk is what I call the "review problem." Signing off on a document you wrote engages different cognitive processes than reviewing a document someone else wrote. When you dictate or type a note, you're actively constructing the narrative. When you review an AI draft, you're doing quality assurance on someone else's work.

Human beings are bad at QA. We skim. We assume. We develop trust in systems that are usually right, which makes us worse at catching the times they're wrong. This is the same automation complacency that causes problems in aviation and manufacturing.

What We Should Be Doing

• Structured review protocols: Don't just tell physicians to "review the note." Give them a checklist. Assessment and plan accurate? Medication list matches? Physical exam findings reflect what was actually done?
• Audit programs: Randomly sample AI-generated notes and compare them against the actual encounter (if recorded). Track error rates by type and feed corrections back to the vendor.
• Clear liability frameworks: Who is responsible when an AI-generated note contains an error that leads to harm? Right now, it's the signing physician — but they didn't write the note. The legal landscape here is unsettled and needs attention.
• Preserve clinical reasoning: The assessment and plan should always be physician-authored. The AI can draft the HPI, ROS, and exam. But the clinical thinking — the "why" behind the decisions — should come from the clinician.

The technology is coming regardless. Our job is to implement it in a way that captures the benefits without creating new categories of harm. That requires the same rigor we apply to any clinical intervention: evidence, monitoring, and a willingness to pump the brakes when the data warrants it.