AI ‘scribes’ are changing how doctors take notes, listening in on patient conversations and automatically transforming them into sleek, narrative-style clinical records. The promise? Less paperwork, less burnout, and smoother workflows. But, is this tool improving healthcare, or merely trading accuracy for efficiency?
The High Stakes of Routine Tasks
Like ChatGPT, AI scribes produce polished, plausible, confident-sounding text—even when wrong. That’s a problem because research shows that the better AI sounds, the more likely people are to trust it—even when they shouldn’t.
Doctors are trained to question test results, but will they apply the same skepticism to mundane note-taking work? A single AI error might seem minor, but mistakes could influence diagnoses and treatment plans. The risks of error aren’t distributed equally as research on AI scribes reports that they struggle with accents, hallucinate, and produce racist text.
This isn’t just an “AI bias” glitch that can be fixed with more data. My co-authored article suggests these flaws are built into how AI processes language—it strips away nuance, overlooks the unsaid, and misses crucial context, which can disproportionately affect those who rely on non-standard or indirect communication styles.
Why “Strategic Friction” is Essential
If AI is bound to fail at human communication, perhaps we should design it to fail better and in productive ways. Rather than an AI system that projects a tone of confidence in a glossy final product, we can work with AI’s limitations to insert more human moments.
What might this look like in practice?
A scaled-down, basic version of AI focused on a narrow, well-defined domain (e.g., transcription, structuring notes).
Uncertainty markers in clinical notes—e.g., “Patient reports pain in (⚠️ no clue—review needed).”
Intentionally non-sleek output that feels more like a draft, reminding doctors that AI is still part of the process and not the final authority.
This ‘strategic friction’ can reveal AI’s flaws in ways that make us conscious of our own biases. By disrupting autopilot workflows, imperfect tools encourage doctors to pause, reflect, and engage with each patient’s unique situation, augmenting accuracy. The key is to implement a form of friction that doesn’t slow down busy hospitals or overwhelm doctors. We will need more research to address what form of friction is practical in high-pressure environments.
Moving Forward: AI for Texture Rather than Gloss
AI’s success is often measured in financial efficiency gains, but we must ask whether these tools genuinely enhance patient care and well-being. Strategic friction—making uncertainty explicit—can open space for human moments and reduce the risk of amplifying healthcare disparities. We need more research on keeping AI helpful while ensuring doctors stay engaged. While more research is required, the guiding principle is clear: AI should spark, not replace, the nuanced judgment of medical professionals.
Additional resources
https://news.northeastern.edu/2024/12/05/ai-scribe-healthcare-quality/
