BACKGROUND AND PURPOSE:

• Biro et al. (npj Digital Medicine, 2025) assessed whether primary care physicians could identify and correct errors in AI-generated draft responses to patient portal messages

METHODS:

• Cross-sectional simulation study
  • Hospital system in Baltimore-Washington area
  • 20 primary care physicians from 13 clinical sites
• Participants
  • Primary care physicians
• Study design
  • Primary care physicians were presented with 18 patient portal messages and AI-generated draft responses
    • 4 of the responses contained errors
  • Physician edits to the AI responses were evaluated to see whether these errors were eliminated prior to the messages being “sent”
• Primary outcome
  • Likelihood of physicians addressing errors in AI generated responses
• Secondary outcomes
  • Physician’s perspectives on AI generated responses

RESULTS:

• 20 primary care physicians
  • Mean years of experience post medical school: 14.75 (SD, 8.43)
• Each of the AI drafts containing errors was “missed” by at least 13 participants
• Mean number of erroneous AI drafts that each participant missed: 66.6%
• Only 1 participant successfully addressed all erroneous drafts
• Survey results found that the participants viewed the drafts favorably
  • Agreed drafts were helpful: 95%
  • Agreed drafts reduced cognitive workload: 80%
  • Agreed that the drafts were trustworthy: 90%
  • Agreed the drafts were empathetic: 75%
  • Agreed drafts were accurate: 70%
  • Agreed drafts were safe to use: 75%

CONCLUSION:

• AI generated drafts of responses to patient portal messages contained factual errors or harmful omissions
• Of the drafts that contained errors, most physicians did not revise the drafts to address these errors
• Physicians missed, on average, two thirds of erroneous messages
• The authors state

To better support the use of AI enabled technologies for practice efficiency, additional research is needed to identify the specific types of errors LLMs are likely to make, and the context under which these errors are most prevalent

Learn More – Primary Sources:

Opportunities and risks of artificial intelligence in patient portal messaging in primary care