Clinicians' Rationale for Editing Ambient AI-Drafted Clinical Notes: Persistent Challenges and Implications for Improvement

This study of 30 clinicians reveals that edits to ambient AI-drafted clinical notes are primarily driven by the need to correct transcription errors, ensure clinical accuracy, mitigate liability risks, and meet billing standards, highlighting the necessity for improved AI customization, integration, and institutional support to enhance human-AI collaboration.

The Gist

Imagine a busy doctor's office. For years, doctors have been drowning in paperwork, typing notes into computers while trying to listen to their patients. This is like trying to cook a gourmet meal while simultaneously washing dishes and paying bills. To fix this, hospitals started using a new tool: Ambient AI.

Think of this AI as a super-fast, super-attentive scribe that sits in the room, listens to the entire conversation between the doctor and the patient, and instantly writes a draft of the medical note. The idea is that the doctor can just glance at it, say "looks good," and sign off.

But here's the twist: The doctors aren't just signing off. They are editing the AI's work constantly.

This research paper is like a detective story that asks: "Why are the doctors still doing so much editing? What is the AI getting wrong, and how can we fix it?"

Here is the breakdown of their findings, using some everyday analogies:

1. The "Bad Autocorrect" Problem (Accuracy & Attribution)

Imagine you are dictating a text message to a friend, but the phone's autocorrect keeps changing your words.

• The Issue: The AI sometimes hears "less common medication" as a different drug entirely, or it thinks the patient's spouse said the patient has a broken leg.
• The Analogy: It's like a translator who gets the accent wrong and accidentally tells your boss you said you were "angry" when you actually said "energetic."
• The Fix: Doctors have to act as editors, fixing spelling mistakes and making sure the AI knows who said what. If the AI thinks the nurse's comment about a different patient belongs in your chart, the doctor has to delete it to avoid medical errors.

2. The "Generic Chef" Problem (Specialty Precision)

Imagine a chef who is great at making a basic burger but tries to make a complex, high-end sushi platter. They might get the rice right, but they miss the subtle nuances of the fish.

• The Issue: The AI is trained to sound like a "general" doctor. But a heart surgeon or a psychiatrist needs very specific language. A psychiatrist might say, "The patient seems anxious," but the AI writes it as a definitive fact. A rheumatologist needs to know if a disease is "mild" or "severe," but the AI just says "joint pain."
• The Analogy: It's like the AI is a generalist intern who doesn't know the specific rules of the game. The specialist doctor has to rewrite the note to sound like an expert, not a student.
• The Fix: The AI needs to be "customized" for different medical fields, learning the specific jargon and priorities of each specialty.

3. The "Legal Shield" Problem (Billing & Liability)

Think of a medical note as a legal contract and a receipt all rolled into one.

• The Issue: In the real world, doctors have to prove exactly what they did to get paid by insurance companies and to protect themselves if a patient sues. The AI often writes things too vaguely. It might say "we discussed treatment," but the insurance company needs to know exactly which risks were explained.
• The Analogy: It's like the AI writes a summary of a court trial that misses the most important evidence. The doctor has to go back and add the specific details ("We discussed the side effects of Drug X for 15 minutes") just to satisfy the "judge" (the insurance company) and the "lawyer" (liability protection).
• The Fix: The AI needs to be taught to be more "bureaucratically precise," automatically adding the necessary details for billing and safety without the doctor having to hunt for them.

4. The "Cluttered Room" Problem (Formatting & Flow)

Imagine the AI writes a 10-page story about your visit, but it puts the "Plan for next week" in the middle of the "History of your childhood."

• The Issue: The AI often dumps information in the wrong sections of the note. It might mix up what the patient said with what the doctor thinks. It also tends to be too wordy, repeating things over and over.
• The Analogy: It's like a messy room where the laundry is in the kitchen and the dishes are in the bedroom. The doctor has to spend time "cleaning up" the room, moving things to the right place, and throwing away the trash (redundant words) before they can actually use the space.
• The Fix: The AI needs to learn how to organize the "furniture" of the note so it looks like a professional document, not a transcript of a rambling conversation.

5. The "Human Adaptation" (How Doctors are Coping)

Since the AI isn't perfect yet, doctors are changing how they talk to make the AI work better.

• The Analogy: It's like teaching a smart but literal-minded robot. If a patient points to their left arm, the doctor now has to say out loud, "The patient is pointing to their left arm," because the robot can't see gestures.
• The Result: Doctors are "narrating" their exams and summarizing their plans out loud just so the AI can hear and write it down correctly.

The Bottom Line

The paper concludes that Ambient AI is a powerful tool, but it's not a "set it and forget it" solution yet.

To make it truly useful, we need:

1. Better Brains: AI that understands medical nuance and doesn't make up facts.
2. Specialized Training: AI that knows the difference between a heart doctor's note and a skin doctor's note.
3. Better Integration: AI that talks directly to the hospital's computer system so it doesn't have to be copy-pasted.
4. Human Training: Teaching doctors how to "speak to the AI" effectively.

Until these things happen, doctors will keep doing the heavy lifting of editing, acting as the human editors who turn a rough draft into a polished, safe, and legal medical record. The goal is to move from "AI writes a draft, doctor fixes it" to "AI writes a draft, doctor just checks it."

Technical Summary

1. Problem Statement

Ambient Artificial Intelligence (AI) tools are increasingly adopted in US healthcare to automate clinical documentation by transcribing patient-provider conversations and generating draft notes. While these tools promise to reduce clinician burnout and administrative burden, a critical gap exists in understanding why clinicians frequently edit these AI-generated drafts.

• The Gap: Prior research has characterized what changes are made (e.g., correcting errors), but lacks insight into the underlying clinical, legal, and workflow rationales driving these edits.
• The Consequence: Without understanding the drivers of editing, AI vendors cannot optimize models for clinical reality, and healthcare institutions cannot realize the promised efficiency gains or return on investment. Persistent editing burdens may undermine the utility of ambient AI, particularly in complex subspecialty care where documentation requirements are stringent.

2. Methodology

The study employed a qualitative research design to investigate clinician behaviors and rationales.

• Setting & Participants: 30 clinicians (15 primary care, 15 subspecialists) from University of California, Irvine (UCI) Health were recruited. Participants included attending physicians, nurse practitioners, and physician assistants who had used a commercial ambient AI tool (Abridge) for over 50 encounters.
• Data Collection: Semi-structured interviews were conducted remotely (Zoom) between January and February 2026.
  • Protocol: The interview guide was informed by prior content analysis of AI draft–final note pairs. Clinicians were presented with de-identified examples of common edit patterns (factual corrections, specialty refinement, etc.) and asked to explain their decision-making processes.
• Analysis: Thematic analysis was performed using Atlas.ti.
  • Process: Three authors developed a codebook based on an a priori framework (derived from prior studies) and iteratively refined it through double-coding and consensus.
  • Validation: Member checking was conducted with two participants to verify interpretive validity.
• Ethics: The study was approved by the UCI Institutional Review Board (IRB).

3. Key Contributions

This study is among the first to systematically link specific editing behaviors to their underlying clinical, legal, and operational drivers.

• Rationale Mapping: It moves beyond describing that edits occur to explaining why they occur, categorizing them into four primary intents: accuracy, specialty precision, compliance/billing, and professional standardization.
• Multi-Level Recommendations: It provides a structured framework for improvement targeting three distinct stakeholders:
  1. Clinicians: Adaptive communication strategies.
  2. Vendors (AI Developers): Model reliability, customization, and personalization.
  3. Institutions: Training, governance, and EHR integration.
• Workflow Integration Insight: It highlights that ambient AI cannot function as a standalone tool; its efficacy is heavily dependent on integration with Electronic Health Records (EHR) and longitudinal patient context.

4. Key Results

A. Primary Drivers for Editing (Clinician Rationales)

Clinicians edit drafts to satisfy four critical domains:

1. Clinical Accuracy & Attribution:
   • Transcription Errors: Misinterpretation of complex medical terms (e.g., chemotherapy agents) or accents.
   • Speaker Attribution: AI often fails to distinguish between the patient, provider, family members, or staff, leading to incorrect history (e.g., attributing a spouse's transplant history to the patient).
2. Specialty-Specific Precision:
   • Generic Output: AI drafts often sound like general internal medicine notes, lacking the differential reasoning and prioritization required for subspecialties (e.g., Rheumatology, Oncology, Psychiatry).
   • Missing Nuance: Failure to capture "pertinent negatives," specific disease activity levels, or the "reading between the lines" required in mental health.
3. Medico-Legal & Billing Compliance:
   • Certainty Calibration: AI tends to make definitive statements where clinicians use probabilistic language (e.g., "likely" vs. "confirmed"). Clinicians must edit to hedge and align certainty with evidence to avoid liability.
   • Billing Requirements: Drafts often lack the specific detail required for coding (e.g., severity of illness, time spent, counseling details, or documentation of conservative measures tried before surgery).
4. Professional Standardization & Workflow:
   • Tone: Drafts often read like verbatim transcripts rather than professional medical records.
   • Structure: AI struggles to place information in the correct EHR sections (e.g., putting assessment findings in the History of Present Illness).
   • Redundancy: Drafts are often too wordy, requiring condensation for readability.

B. Persistent Challenges Identified

• Context Blindness: AI lacks access to the patient's longitudinal history, medication lists, or prior notes, leading to omissions of context that clinicians must manually add.
• EHR Friction: Poor integration forces clinicians to copy/paste data between the AI tool and the EHR, negating efficiency gains.
• Speaker Confusion: In multi-party encounters (family, interpreters), speaker diarization remains a significant failure point.

C. Recommendations for Improvement

• Clinician Strategies: "Speak-to-document" tactics, such as verbally narrating physical exam findings (e.g., "rash on right arm") and explicitly stating laterality or negative findings to ensure capture.
• Vendor/Technical Targets:
  • Core Reliability: Improve speaker attribution and transcription accuracy for medical jargon.
  • Customization: Enable specialty-specific templates (e.g., Oncology vs. Primary Care) and encounter-type presets (e.g., new visit vs. follow-up).
  • Personalization: Allow clinicians to "re-prompt" or adjust verbosity and include controls for specific documentation styles.
  • EHR Integration: Deep integration to pull longitudinal data (meds, past history) directly into the draft to reduce manual entry.
• Institutional Support: Standardized training on "AI-aware" communication, clear governance on consent/liability, and role-specific onboarding.

5. Significance

• For AI Development: The findings suggest that the next generation of ambient AI must move beyond simple transcription and summarization. To be truly useful, models must understand clinical reasoning, specialty-specific norms, and regulatory constraints (billing/liability).
• For Healthcare Implementation: The study warns that without addressing the "editing burden," ambient AI may fail to deliver on its promise of reducing burnout. Success requires a coordinated effort involving better technology, workflow redesign, and institutional policy.
• For Clinical Practice: It underscores the continued necessity of human oversight. Clinicians act as the final gatekeepers for accuracy, legal safety, and clinical nuance, necessitating a shift toward "human-in-the-loop" collaboration rather than full automation.

Conclusion: The paper concludes that ambient AI is currently a "first draft" generator that requires significant human refinement. To transition from a tool that requires extensive editing to one that enhances efficiency, future iterations must prioritize core model reliability, specialty customization, deep EHR integration, and institutional governance.