Why AI translation apps can fail doctors and patients.

After one of his patients died, Stephen Hibbs thought AI could have saved him, at least at first. His patient, a young Albanian, died after a bone marrow transplant in east London. He was discharged with a complex post-operative medication regimen aimed at controlling his immune system and preventing infections, a regimen that had to be followed perfectly. But he stopped taking one of the medications too soon and died before he could explain exactly why he stopped. Hibbs, a blood cancer specialist, couldn’t stop thinking about it. His patient did not speak English fluently. Would the exit instructions in his native Albanian have been easier to follow? Could his death have been avoided? Could others?

Even when everyone speaks the same language, patients and doctors struggle to communicate about symptoms, treatment options and risks. Adding language differences increases the chances of things going wrong. Anyone can be vulnerable, including when traveling or with a multilingual medical team. With the rise of generative artificial intelligence tools like ChatGPT and translation apps like Google Translate, healthcare providers have what seems like a simple, inexpensive way to fill the gap in their lab coat pockets. But while these technologies promise to make medical communication easier and more affordable, they also introduce new risks. And in a context where a mistranslated word can mean the difference between an accurate diagnosis and a dangerous error, the consequences can be serious, even fatal.

In the United States, people with limited English proficiency are at greatest risk. A 2024 Research Review in Healthcare Journal fIt appears that people with limited English proficiency are more likely to experience serious harm from medical errors and receive lower quality health care. This is one reason why federal law requires most providers to make health care communication meaningfully available in the language patients understand, through a human interpreter for oral communication and translator review for written information.

However, finding a qualified person is not easy. Some 27 million Americans have poor command of English. entire In the United States, the translator and interpreter profession numbers just under 74,000 people, including many people working in non-health care specialties, according to the Bureau of Labor Statistics. When President Donald Trump proclaimed English the official language of the United States in 2025, these problems worsened. The administration alone couldn’t repeal health care translation laws, Lynette Martins, a health law professor at DePaul University, told me. But it could de-prioritize enforcement of these laws and change critical government policies that previously supported language access. Martins describes it as an “exacerbation of inequality”.

It wasn’t exactly subtle. The federal government’s new policy is to avoid providing translation services to people with LEP unless the letter of the law or “critical” practicalities require it. Some of this is anti-woke virtue signaling, but there have also been serious and practical effects. For years, the Justice Department’s Civil Rights Division maintained a resource-rich website, LEP.gov, that offered widely used information on how to help people with limited English proficiency, with major translations available in 20 other languages ​​commonly spoken in the United States. But it remained in ruins for more than a year. The only page that remains contains a link to Trump’s “English only” executive order. The only other link on the site is to former Attorney General Pam Bondi’s implementation memo, which promised new LEP guidelines within 180 days. This would have been January 10, 2026. The DOJ did not respond to my email request for an updated timeline.

Bondi’s memo also casts an implicit shadow over the use of human translators, while boosting AI: it hails “artificial intelligence and machine translation” as “cost-effective methods for overcoming language barriers and reducing inefficiencies in the translation process”, and calls for the phasing out of “unnecessary multilingual offerings”. But many doctors already know what medicine looks like without human translators.

A patient arrived at the Denver emergency room one night, terrified of having a heart attack. In Spanish, he told William Mundo that this was his second visit to the emergency room in as many days and handed him the discharge instructions he had received. Even though he was early in his residency, Mundo, a bilingual doctor, knew about this “rebound.” Patients with LEP are more likely to make an unplanned return to the hospital after discharge from the emergency room, according to a 2024 study from the University of California, San Francisco published in BMJ Quality & Safety.

He also saw what caused the false alarm: The Spanish translation explaining how to distinguish expected acid reflux from possible symptoms of a heart attack was clunky and confusing. To him, the telltale errors in commercially available AI translation tools were obvious.

He understood why his colleagues were using them, even though hospitals generally don’t approve the use of unapproved technologies. Whenever his patients spoke a language he didn’t speak, “I did what everyone would do: copy, paste, throw it on Google Translate and hope it was accurate enough,” he told me. According to Mundo, hospital culture puts pressure on doctors to “move the meat.”

“Move the patients, move the hospital,” he explained. Translation services, when available, take time that most clinicians feel they cannot devote.

The American Medical Association recently found that 57% of doctors are using or considering using AI-based translation tools. It’s not because they think it’s ideal, or even acceptable, said Mundo, who emphasized that he is now senior enough to withstand the pressure of time. “A lot of my colleagues tell me, ‘Well, at least (patients) can read the discharge instructions.’ » He understands the rationale for “a little is better than nothing,” but he has also seen the harm caused by translations. “It’s an ongoing ethical dilemma,” he said. “We try to do the best we can with what we have.”

The medical use of unauthorized AI translation tools is therefore an open secret. But measuring how dangerous this is in practice is difficult, because the field has yet to find a standard way to measure the safety of AI translation, even for AI tools used by hospitals and health systems. to have approved for use in translation, Dipak Patel, CEO of GLOBO Language Solutions, told me. His company is one of a small group of language service providers that have long offered translation services in medical settings, via video, telephone and in person. Several, including GLOBO, now offer AI translation. Across the small industry, these services rely on a generally unspecified mix of third-party language models and proprietary technologies.

“If you ask Hospital A, Hospital B or any Federally Qualified Health Center how they measure interpretation accuracy, they will all give you the same answer: We let the provider tell us,” Patel said. As each language services company uses its own measures of accuracy, each essentially grades its own assignments, using its own rules. Without an agreed standard measurement, industry-wide accuracy comparisons are impossible.

And existing AI translation metrics aren’t really designed for this task. These often measure text similarity or overlap, but they do not heavily penalize small errors that should fail in a healthcare context. It doesn’t help that doctors tend to use language in their own way. For example, when a doctor wants you to stop taking a medication, they will often ask you to “hold” it, meaning a temporary break or delay, like asking someone to hold the elevator. But “hold” can have the opposite meaning in a different context: “stay the course” means maintain one’s direction. Mistranslating this instruction “hold” as “adher to your medicine” dangerously reverses the meaning, while still being similar enough that it could get a decent score. (This was a real AI translation error regarding kidney drugs from English to Russian that researchers found in a 2025 BMJ quality and safety study. The AI ​​translated it correctly into Chinese.)

Inspired by his experiences, Mundo recently presented a study on AI-translated patient discharge instructions. When doctors reviewed translations with high quality scores, they also found small but common errors in medication instructions, as well as warnings about when to urgently return to the emergency room. The models also struggled to correctly parse negation, such as in the comparison between “take with food” and “not take with food.”

These issues become worse when a model operates in a language that is less represented in the AI ​​training data. For example, in a 2025 study of pediatric emergency room scenarios in the journal BMC Health Services Research, a model mistranslated a verb from Urdu so that a description of a child’s fluid intake made it sound like the child had consumed alcohol — which is chilling, considering the error could open a child protective services case.

But even in English and Chinese, two of the most digitally represented languages, changing languages ​​is not the same neutral act as changing a phone’s language settings. In a 2025 study in Nature Human Behavior, MIT Professor Jackson G. Lu and colleagues found that OpenAI and Chinese Baidu models responded differently to the same prompts depending on whether they were written in English or Chinese. In English, the models tended to…