Doctors Are Using a Shortcut for an Age-Old Problem. It Can Be Pretty Risky.

Sign up for the Slatest to get the most insightful analysis, criticism, and advice out there, delivered to your inbox daily.

After one of his patients died, Stephen Hibbs believed A.I. could have saved him—at least, at first. His patient, a young Albanian man, died after a bone marrow transplant in East London. He was discharged with a complex postsurgery medication regimen to keep his immune system in check and prevent infection, one that had to be followed perfectly. But he stopped taking one of the medications too soon and he died before he could explain exactly why he stopped. Hibbs, a blood-cancer specialist, couldn’t stop thinking about it. His patient wasn’t fluent in English. Would discharge 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 raises the odds of things going sideways. Anyone can be vulnerable, including while traveling, or with a cross-lingual 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 cheap and easy way to bridge 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 of their own. And in a setting where a mistranslated word can mean the difference between an accurate diagnosis and a dangerous mistake, the consequences can be severe—or even deadly.

In the U.S., people with limited English proficiency bear the most risk. A 2024 research review in the journal Healthcare found that those with limited proficiency in English are more likely to experience serious harm from medical errors and receive lower-quality healthcare. That’s one reason federal law requires most providers to make healthcare communication meaningfully available in the language patients understand, via a human interpreter for spoken communication and a translator’s review for written information.

Yet finding a qualified human isn’t easy. Some 27 million Americans have low English proficiency, yet the entire U.S. translation and interpretation profession numbers just under 74,000, including plenty of people who work in non-healthcare specialties, according to the Bureau of Labor Statistics. When President Donald Trump proclaimed English the U.S.’s official language in 2025, these issues worsened. The administration couldn’t repeal healthcare translation laws singlehandedly, Lynette Martins, a professor of health law at DePaul University, told me. But it could deprioritize enforcing those laws, and it changed critical government policies that had 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 for people with LEP, unless the letter of the law or “mission-critical” practicality demands it. Some of that is anti-woke virtue signaling, but there have also been serious, practical effects. For years, the Department of Justice’s Civil Rights division maintained a resource-rich LEP.gov website, which offered widely used information on how to support people with limited English proficiency, with prominent translations available to 20 other languages commonly spoken in the U.S. But it has been a husk for over a year. The single page that remains has 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 guidance within 180 days. That would have been Jan. 10, 2026. The DOJ didn’t respond to my emailed request for an updated timeline.

Bondi’s memo also threw shade by implication at the use of human translators, while boosting A.I.: It praised “artificial intelligence and machine translation” as “cost-effective methods for bridging language barriers and reducing inefficiencies with the translation process,” and called for phasing out “unnecessary multilingual offerings.” But many doctors already know what medicine is like without human translators.

A patient arrived at a Denver emergency room one night terrified that he was having a heart attack. In Spanish, he told William Mundo this was his second ER visit in as many days and handed him the discharge instructions he’d received. Though he was early in his residency, Mundo, a doctor who is bilingual, was familiar with this “bounce back.” Patients with LEP are more likely to make an unplanned return to the hospital after an ER discharge, a 2024 University of California, San Francisco review published in BMJ Quality & Safety found.

He also saw what caused the false alarm: The Spanish translation of how to distinguish expected acid reflux from possible heart attack symptoms was clumsy and confusing. To his eye, the telltale errors of off-the-shelf A.I. translation tools were obvious.

He understood why his colleagues used them, even though hospitals generally don’t endorse the use of unvetted tech. Whenever his patients spoke a language he didn’t speak, “I did what everybody else would do: copy, paste, throw it onto Google Translate and hope that it was somewhat accurate,” he told me. According to Mundo, hospital culture puts pressure on doctors to “move the meat.”

“Move the patients around, keep the hospital moving,” he explained. Translation services, when available, take time most clinicians feel they can’t spare.

The American Medical Association recently found that 57 percent of doctors are using, or plan to use, A.I. translation tools. It’s not because they think it’s ideal, or even OK, says Mundo, who noted he is now senior enough to push back against the time pressure. “A lot of my colleagues tell me, ‘Well at least [patients] can read the discharge instructions.’ ” He can see the justification of “a little bit is better than nothing,” but he’s also seen the harm caused by translations. “It’s an ongoing ethical dilemma,” he said. “We try to do the best that we can with what we have.”

So the medical use of unsanctioned A.I. translation tools is something of an open secret. But it’s hard to measure how dangerous it is in practice, since the field has yet to settle on a standard way to measure A.I. translation safety—even for the A.I. tools that hospitals and health systems have approved for translation use, Dipak Patel, CEO of GLOBO Language Solutions, told me. His company is one of a small group of language-services vendors who have long provided translation services in medical settings, via video, phone, and in person. Several, including GLOBO, now offer A.I. translation. Across the small sector, these services draw on a usually unspecified mix of third-party language models and proprietary technology.

“If you ask Hospital A, Hospital B, or any federally qualified health center how they measure the accuracy of interpretation, they’ll all give you the same answer: We let the vendor tell us,” Patel said. With every language-services company using its own accuracy measures, everyone essentially grades their own homework, using their own rules. Without an agreed-upon standard measurement, industrywide accuracy comparisons are impossible.

And existing A.I. translation metrics aren’t really designed for this task. These often measure text similarity or overlap, but they don’t heavily penalize the tiny errors that should fail in a health setting. 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’ll often instruct you to “hold” it, as a synonym for a temporary pause or delay—like asking someone to hold the elevator. But “hold” can have the opposite meaning in a different context: “hold your course” means to maintain your direction. A mistranslation of that “hold” instruction as “adhere to your medication” dangerously reverses the meaning, while being similar enough that it might score decently. (This was an actual A.I. translation error regarding kidney medication from English into Russian that researchers found in a 2025 BMJ Quality & Safety study. The A.I. translated it correctly in Chinese.)

Inspired by his experiences, Mundo recently presented a study of A.I.-translated patient discharge instructions. When doctors reviewed translations with high quality scores, they also found small but frequent mistakes in medication instructions, as well as in warnings about when to urgently return to the ER.  The models also struggled to parse negation correctly, as in “take with food” versus “don’t take with food.”

These problems compound when a model is working in a language that is less represented in A.I. training data. For instance, in a 2025 study of pediatric emergency room scenarios in the journal BMC Health Services Research, a model mistranslated a verb from Urdu, so a description of a child’s liquid intake read like the child had consumed alcohol—chilling, when you consider the mistake could open a child protective services case.

But even in English and Chinese, two of the best-represented languages digitally, changing languages isn’t the same neutral act as changing a phone’s language setting. In a 2025 study in Nature Human Behaviour, MIT professor Jackson G. Lu and his colleagues found that OpenAI and China’s Baidu models responded differently to the same prompts depending on whether they were written in English or Chinese. In English, the models tended to favor more individualistic and analytical responses, while in Chinese, they placed greater emphasis on relationships and broader context. In other words, the language of a prompt influenced the cultural assumptions underlying its response. The study wasn’t focused on healthcare specifically, but it raises important questions about whether patients and doctors using different languages can reliably receive consistent information about diagnoses, treatment options, and medical risks.

“I doubt that we could remove every cultural tendency from a model, because language and culture are deeply intertwined,” Lu told me.

At his East London clinic, Stephen Hibbs—who treated the Albanian patient—found many of his colleagues had similar difficult stories of patients being harmed by language barriers. Translation and interpretation services existed, but not always; take-home materials were invariably in English only.

This was in the early 2020s, the dawn of the smooth-talking chatbot era. Surely such powerful intelligence could translate discharge instructions into a patient’s preferred language? “Great,” Hibbs told me, recalling his reaction at the time. “Here’s an off-the-peg solution that we can scale up rapidly and all over the place.” He brought the idea to Federico M. Federici, director of University College London’s Centre for Translation Studies. Looking back, both chuckle at what Hibbs now calls his “fairly naive hope.”

“It was not unexpected,” Federici told me. The new LLM tools made translation “seem very easy, very quick.” He suspected they’d be neither—his field had been using machine translation since the 1950s. But both wanted to learn whether A.I. translation could be trusted in high-stakes medicine. They launched a study to find out.

Hibbs and Federici published their findings in the journal Supportive Care in Cancer in March 2026. For the study, a simple booklet called Easy Read: How is Myeloma Treated? was translated from English into Bengali twice—once by machine, and once by humans. For the version translated by A.I., a research team member who did not speak Bengali put it through Google Translate. For the human version, the best professional translation practices were applied: One professional translator’s work was checked by another, and a third proofread for readability.

When both were scored for accuracy by a different group of assessors who didn’t know which translation was which, neither proved to be error-free. Out of a possible 100 points on a scale used by the Chartered Institute of Linguists, the human version scored a 73, losing points mainly on points of language and clarity. As for the A.I. translation, most of the booklet was translated accurately, the assessors found. But its score was still a lousy 51.

Although there weren’t many of them numerically, the errors that the A.I. did make were confusing and bewildering enough to be disqualifying. For instance, the line “The medicine will help you to: Kill as many myeloma cells as possible” was mistranslated as instructing the patient to somehow kill as many myeloma cells as possible. There is a specific word for medical “side effects” in Bengali, but the A.I. translated it as “harmful side.” Other errors were pure gibberish. “We can put you in touch with another myeloma patient to talk to” became “We can contact other patients for us to talk to you.” After reading the weird A.I. translation, a Bengali speaker said it would make them question whether they could trust the clinic—or the doctor who handed it to them—at all.

But here’s the thing: When they tested comprehension, the group that read the far superior human translation understood just as little as those who read the A.I. version. “Almost no one understood the key points,” Hibbs said.

The final component of the study is where that changed. All participants had a live conversation with a clinician and a professional interpreter, simulating the informational discussion of the risks and benefits which typically precedes myeloma treatment. Part of this conversation includes walking through the patient consent form. One group reviewed the standard English-only consent form, while the other had a professionally translated “interlinear” form, which placed Bengali and English text near each other. On the test that followed, comprehension improved for both groups, but it was highest for the group whose conversation included the bilingual form.

In other words, more human time and attention boosted understanding.

Still, the takeaway isn’t “A.I. translation bad, human translation good.” Human translators rely on A.I. tools; they have to, just to meet even a fraction of demand, Federici said. And in the LLM era, human involvement remains a necessary failsafe—not just as an error-catcher, but to ensure meaning survives. “It’s the unsupervised use of this technology, scaling it in the way that Stephen [Hibbs] was initially describing, that is fundamentally dangerous,” Federici said.

Hibbs agrees. He started with the hope that A.I. translation could save his patients’ lives. Now that he’s seen the kinds of errors it makes, that hope is gone. Without human review, “I would not use an A.I. tool, even if there was no other alternative,” he said. Unless accuracy improves significantly, he thinks handing a patient a translation that might be wrong is ethically worse than giving them no written information at all—as risky as sending a medical student to perform an operation solo. “In almost every case, it’d be better to avoid the surgery than to expose the patient to the risk of a botched operation.”