AI Can Translate for Hospital Patients. Should It?

AI has moved quickly into hospital language services.

A 2026 paper in BMJ Health Care Informatics reported that 57% of U.S. physicians are already using or planning to adopt AI translation tools within the next year, a 30% increase from 2023. Translation now ranks among the most familiar clinical uses of generative AI.

The shift affects a large patient population. An estimated 25.7 million Americans have a non-English language preference, the same BMJ paper found.

Hospitals have long relied on professional interpreters and translators to bridge that gap. Written discharge instructions can take one to five days to translate at large academic centers and are sometimes deferred because of cost constraints.

Now providers like Children's Hospital Los Angeles are testing AI tools to produce discharge summaries and other patient materials more quickly. AI vendors say their systems can produce translations in minutes.

Written Translation Moves First

Spence Green, CEO of LILT, an AI translation platform, focuses the company on "high assurance" environments where errors carry regulatory or factual risk. Its platform offers a "contractual quality guarantee," Green told AIM Media House.

LILT is deployed at one of the top hospital systems in the United States for internal patient communications. The system is configured to meet strict information security standards and can run within a hospital's existing cloud environment.

Hospitals choose such systems, Green explained, because they need to know which models are used and how data is governed. In health care, security and compliance reviews can take months before a tool goes live.

That reality is visible on the ground. Dale Lundstrom, a Spanish medical interpreter and language services education coordinator at Intermountain Health, told AIM Media House that his organization is still working through the vetting process. "Our company is still in the process of vetting AI tools for HIPAA compliance, security, quality and effectiveness," he said. A timeline that echoes what vendors like Green describe from the hospital side.

Speed and cost are part of the appeal. AI tools are "absolutely" less costly than relying solely on human translators, Green noted. For some hospital workflows, translated materials are required in under an hour.

Other startups are targeting similar needs.

Jaide Health, founded by physician Joe Corkery, is working with Children's Hospital Los Angeles to translate after-visit summaries so patients can leave with instructions in their preferred language, Corkery explained in an interview with Slator. The goal is to ensure patients understand care plans at the moment they leave the hospital. No Barrier, another AI translation startup, says its platform is active in more than 100 health care sites across 12 states and supports more than 40 languages.

These deployments focus largely on written material. Under 2024 federal regulations implementing Section 1557 of the Affordable Care Act, critical documents such as consent forms and discharge instructions must be reviewed by a qualified human translator before being given to a patient. That structure makes written translation easier to automate. Output can be generated by AI, checked by a human reviewer and corrected before release.

Large language models can produce fluent translations, Green noted, but organizational context still matters. In one example he described, a document translated into German appeared flawless to a general reader but contained errors in product names and measurements that factory staff would immediately recognize.

Interpretation Under Pressure

Live medical encounters present a different set of challenges.

Access to interpreters has long been uneven. A 2024 study in the Health Promotion Journal of Australia found hospital staff faced long wait times for telephone interpreters, inflexible booking systems and limited availability of in-person interpreters during the COVID-19 pandemic. Clinicians reported that these barriers contributed to underuse of interpreter services.

Those access gaps, researchers and practitioners say, push clinicians toward whatever tool is nearest.

Anne Cronin, who investigates language access for migrants and refugees at Ireland's Health Service Executive and the University of Limerick, observed the same pattern. "GPs and healthcare workers rely on AI-powered apps like Google Translate when they don't have ease of access to interpreting services of varying modalities," she told AIM Media House. The consequences, she added, are significant: "When healthcare professionals rely on AI instead of human interpreters they are likely to encounter inaccurate information transfer, leading to errors, misdiagnosis, possible inappropriate treatment, hallucinations and uninformed patient consent, all of which introduce risk to the patient."

Cronin also flagged a legal dimension that remains largely unexplored. "In terms of litigation, the area is underexamined, but it certainly raises the question about healthcare workers at risk of litigation when using general purpose apps that are not designed for clinical practice. In what other domain of health are we using untested tools?"

Dr. Pastora Hernández Barbee, an interpreter and licensed mental health counselor, offered a pointed assessment in an exchange with AIM Media House, saying AI "has not changed anything regarding interpretation services" in her experience. Human interpreters remain essential even where AI tools are available.

Policy debates have intensified as some hospitals shift toward digital tools. In New York, State Sen. Monica R. Martinez criticized a decision by Northwell Health to eliminate in-person interpreter services at one hospital and replace them with digital and phone-based options. Replacing trained interpreters with digital services "compromises patients' access to proper treatment," according to an October 2025 press release.

At the same time, companies Jaide Health and No Barrier are building speech-to-speech systems aimed at real-time interpretation. No Barrier CEO Eyal Heldenberg described his system as a real-time audio-to-audio pipeline designed to mirror how human interpreters pause, clarify and confirm details.

Performance Gaps and Policy Limits

Federal law continues to require meaningful language access under Title VI of the Civil Rights Act and Section 1557 of the Affordable Care Act. Detailed national standards for validating AI translation quality across languages remain in development.

Meanwhile, AI translation does not perform evenly across languages. Cultural nuance presents one layer of risk. In British Columbia, researchers developing a Punjabi Translation Framework found that some English-to-Punjabi medical materials appeared to rely on literal AI output, leading to inappropriate or stigmatizing word choices. Punjabi is a gendered language, and default masculine translations can misrepresent meaning, the researchers found.

Performance differences also appear in clinical testing. Research summarized in BMJ Health Care Informatics found that GPT-4 translations of pediatric discharge instructions into Spanish were often preferred over professional human translations for fluency and clarity. For Haitian Creole, the same review reported clinically significant error rates of 8.3% for professional human translations, compared with 23.3% for Google Translate and 33.3% for ChatGPT.

Interpreter shortages compound the issue. In California, 177 certified Cantonese and 388 certified Mandarin medical interpreters serve more than 900,000 Chinese speakers with limited English proficiency, according to hospital data. Hospitals are deploying AI tools across dozens of languages while performance data exist for only a subset of them.

Underlying all of it, researchers say, is a governance problem that the industry has yet to confront.

Stephen Ma, medical informatics director of analytics and evaluation at Stanford Medicine and a co-author of the BMJ paper, puts it in an exchange with AIM Media House: "The current situation is problematic because people are using it despite the fact that they're not supposed to and we have no idea what the actual impact is on downstream care. There needs to be a regulatory path forward so that the need can be met in a way where we understand the impact, the potential failures, and can set up guardrails, iterate and improve."

As written translation expands and real-time interpretation remains under human supervision in many settings, the pace of AI adoption is outstripping the evidence base that governs how it performs across languages. Without clearer standards, the quality of machine-assisted care may continue to vary, not just by language, but by whatever tool a clinician happens to reach for when no interpreter is available.