Can AI help hospital translators finish discharge instructions sooner?
There is a hopeful result in a 2025 hospital translation study: professional translators working from AI drafts finished sooner on average, with overall quality ratings similar to or better than translations written from scratch. The useful version of this finding keeps the translator in the job. Researchers used 20 short pediatric discharge texts across six languages, comparing ChatGPT-4o alone, professional translation, and AI drafts edited by separate professional linguists. Mean translation time was 7.1 minutes for the edited drafts versus 16.8 minutes from scratch. But those clocks excluded administrative processing and production queues. They don't tell us when a family actually received the instructions. Armenian also did not show a statistically significant time saving. The unedited AI translations performed unevenly, with particularly poor ratings for Armenian and Somali. This was a specific older model and a small set of texts, not a verdict on every translation tool available today. Nor were the quality ratings evidence of fewer medication mistakes at home. If a hospital is considering this approach now, I'd ask it to measure the time from requesting the translation to handing the checked version to the family, separately for each language. That's a follow-up I'd propose, not an outcome this study measured. A faster draft would be welcome if it helps a parent leave with instructions they can use, rather than another promise that the translated copy will arrive later.
Comments
I'd ask what happens to the translation budget before celebrating the saved minutes. If a hospital uses faster drafts in some languages to cut staffing across the service, families who still need the slower route could wait longer. That's a risk to guard against, not a finding from this study. Keep the professional translation route funded where the AI draft doesn't help, and check waiting times by language before reducing cover. A parent's access to checked instructions shouldn't depend on whether their language makes the software look efficient.
Before calling the instructions delivered, I'd ask the parent to open the checked copy in the form they'll use at home. A portal notification isn't much help if they can't sign in while holding a tired child. Offer a printed copy if they want one, and check that the text is readable rather than assuming a translated PDF settles access. That's a handover check I'd add to a hospital pilot, not an outcome this study measured. I wouldn't want a family reaching for a chatbot to translate the English sheet again because the professionally checked version is stuck behind a password.
I'd settle who is qualified and scheduled to do that review before putting AI drafts into the discharge process. ‘Ask someone bilingual to check it’ can turn into an extra job for whichever nurse or receptionist happens to be nearby. The study used professional linguists; it doesn't establish that an informal colleague check gives the same result. I'd have the translation-service lead assign the review and agree what happens when no reviewer is available, rather than leave ward staff to improvise under discharge pressure. That's a staffing condition I'd put on a pilot, not a result the study measured. Speaking two languages shouldn't quietly make someone the hospital's translation service.
The reviewer also needs to know which source copy is final. Suppose the English discharge sheet gets a corrected follow-up date after the translation was checked. I'd want that change to reopen the translation job, with someone responsible for replacing any earlier patient copy. A perfectly accurate translation of yesterday's instructions is still the wrong handout. That's a version-handling rule I'd ask a hospital to include in a pilot, not a result from this study. Don't leave a family comparing two sheets in languages they may not both read.