Authors Bender and Muldowney argue patients should refuse AI ambient scribing tools in clinical visits, citing nine specific harms across privacy, consent, accuracy, and care quality.
Key Takeaways
AI scribing systems route audio recordings to third-party vendors; HIPAA compliance guarantees are not equivalent to strong security protocols.
Automation bias cuts two ways: providers may accept incorrect notes, but more critically, may fail to notice what the AI omitted entirely.
Providers exposed to scribing systems shift into a technical “doctor-to-doctor” register mid-visit, confusing medical interpreters and patients alike.
Speech recognition accuracy is unequal across dialects, second-language speakers, and patients with dysarthria, creating disproportionate correction burden on already-strained providers.
The authors frame mass patient refusal as a systemic lever: low consent rates make institutional “efficiency” claims harder to sustain and slow provider-load increases.
Hacker News Comment Review
Deployed practitioners directly contradict the “false promise of efficiency” framing – a healthcare CIO and a physician’s spouse each report immediate, measurable time savings and improved note quality after rollout.
Technical commenters push back on the privacy framing: existing EHR data is already far more sensitive and widely shared (including via fax), so singling out audio recordings as uniquely risky is seen as inconsistent risk accounting.
Commenters acknowledge the charting-as-care argument has merit but note that manual documentation has always produced errors; ambient audio with local speech recognition is not categorically worse.
Notable Comments
@burnte: 12-year healthcare CIO who deployed two scribing tools reports immediate gains in patient NPS, provider satisfaction, and note verbosity across every deployment.
@ivraatiems: Physician spouse cut after-hours charting from 8-16 extra hours per week to under half; characterizes the efficiency claim as real, not false.
@16bytes: Notes many e-scribes run speech recognition entirely locally, not sending audio to cloud; calls blanket refusal advice “absolutely horrible” for long-term care quality.
