The shortage of qualified medical interpreters is severe and long-standing, and machine translation has moved into the gap quickly. At several hospitals we contacted, real-time AI interpretation is now the default for languages where an in-person interpreter is not available within a defined window.

For wayfinding, discharge instructions and routine history-taking, clinicians and patients report the tools work well. The concern raised repeatedly by interpreter services leads is scope creep into conversations where a mistranslation is not recoverable: consent for procedures, goals-of-care discussions, and disclosure of serious diagnoses.

The validation literature is genuinely thin here. Most published evaluations measure translation quality on general or documentation-style text. Studies measuring comprehension and decision quality in consent conversations, with the patient as the unit of analysis, are rare.

Interpreter services associations have proposed a bright-line rule: machine interpretation permitted for informational exchange, human interpreter required for consent and serious news. Several systems have adopted a version of it. Others have declined, citing the practical reality that the alternative to a machine interpreter at 2 a.m. is frequently a family member.