Emily Bender and Decca Muldowney (Mystery AI Hype Theater 3000) lay out nine reasons patients should decline AI scribing tools now infiltrating clinics from small offices to Kaiser.
Key Takeaways
AI scribing systems send recordings to third-party vendors; HIPAA compliance assurances do not equal strong security protocols for transcripts.
Automation bias cuts two ways: providers over-trust draft notes for what is included and systematically miss what was never captured.
Speech recognition accuracy degrades for non-standard dialects, second-language speakers, and patients with dysarthria, creating disproportionate correction burden on already-stretched providers.
Providers accustomed to scribing tools shift into a technical “doctor-to-doctor” register mid-visit, confusing medical interpreters and potentially patients.
The article argues charting itself is cognitive care: reflection time on symptoms and progress degrades if offloaded to a draft-and-approve loop.
Hacker News Comment Review
Practitioners with actual deployment experience pushed back hardest: a healthcare CIO who deployed two such tools reported immediate gains in patient NPS, provider satisfaction, and note verbosity, directly contradicting the “false promise of efficiency” framing.
Several commenters argued the privacy objection is weak because all patient data already flows through third parties, and fax-based record transfer is a worse baseline; the “more patients” outcome was reframed by one commenter as a genuine efficiency win, not a harm.
A sharp practical alternative surfaced: local on-premises inference on an NVIDIA workstation eliminates the third-party data transfer entirely, and transcription workloads are not GPU-demanding enough to make this impractical for a clinic.
Notable Comments
@dlcarrier: Notes the real risk is the existence of the record itself – a depression diagnosis obtained for insurance coverage can affect pilot certification and firearm ownership.
@cromka: Argues on-prem AI on a single NVIDIA workstation solves the privacy problem; questions why cloud became the default assumption for transcription.
