Doctors often use Google Translate to talk with non-English speaking patients, but researchers say inaccurate results are common and can lead to serious errors
The patient had just undergone a cesarean section, and now was struggling to put words to her pain in her native Taiwanese. Tweets: @berkeleydatasci , @rickberke , and @amerurological Tweets: @berkeleydatasci : “Obviously, Google Translate wasn't built for health care applications,” said @nikita_mehandru, a Ph.D. student in clinical artificial intelligence at @BerkeleyISchool. “Maybe something should be.” Read more in @statnews on @niloufar_s and Nikita's work: https://www.statnews.com/... Rick Berke / @rickberke : How Google Translate has become a ubiquitous, if imperfect, part of patient care, and why some researchers want a change. By @KatieMPalmer https://www.statnews.com/... via @statnews https://twitter.com/... Amer. Urol. Assn / @amerurological : “There is a growing push by researchers to bring @Google Translate to the surface - both to study its use and risk in the clinic, and to build better versions to backstop traditional language services.” #Telehealth #MedTwitter https://www.statnews.com/...
Context & Ripple Effects
Berkeley clinical-AI researcher Nikita Mehandru's finding lands on an already-documented pattern: a year earlier, Reuters reported that translation tools were mangling legal terms with conflicting meanings, and STAT had separately flagged that hospitals run novel, largely unproven AI decision-support tools without most patients knowing. The cesarean-section patient struggling to describe pain in Taiwanese is the human-scale version of a workflow problem — clinicians reaching for a consumer tool because nothing clinical-grade is in reach.
The follow-on coverage shows where this goes next: by mid-2023, US doctors were using ChatGPT to communicate more empathetically with patients, extending the same habit of improvising with general-purpose AI at the bedside.
First-order effects
- Non-English-speaking patients are exposed to translation errors at the exact moments of highest stakes — post-surgical pain, consent, discharge instructions — while the clinicians using Google Translate bear the liability.
- Mehandru's framing ('maybe something should be built') names the immediate gap: no validated, health-care-specific translation layer exists for the clinicians already relying on the consumer product.
Second-order effects
- Hospitals face pressure to either procure purpose-built clinical translation or formalize interpreter workflows, since ad-hoc consumer-tool use is now documented as a patient-safety issue rather than a convenience.
- Vendors of clinical AI have an opening the research effectively advertises: a translation tool built for care settings would compete against a free incumbent whose flaws are now peer-reviewed.
Third-order effects
- If the pattern holds — consumer AI drifting into clinical use before validation, as with the mammography systems Wired warned could worsen overtesting — regulators and hospital accreditation bodies will eventually be forced to treat informal AI tooling as a governed medical workflow rather than staff improvisation.
- The deeper structural shift is that frontline clinicians, not procurement, are de facto selecting clinical AI — meaning safety evidence trails adoption instead of preceding it.
The trend: General-purpose AI tools are entering clinical workflows through clinician improvisation first and validation second, from Google Translate to ChatGPT at the bedside.