AI translation apps have quietly entered the exam room. Clinicians use them to bridge language gaps when a qualified interpreter is not at hand. A recent Slate investigation by Alison J. Stein looked at this shift and at the errors it can produce. The reporting raises fair questions for health systems. Who measures accuracy? Which tools are safe? Where does human expertise still belong? In this post, I build on that reporting with what we see across deployments at No Barrier. I also draw a line the debate often misses: translating text and interpreting speech are two different tasks.
TL;DR: Demand for qualified interpreters far exceeds supply, so clinicians improvise with consumer apps that are not HIPAA compliant. This is not a reproach, far from it. Clinicians do what they can with the time and resources they have. Federal policy now points to AI as the cost-effective answer but the sustainable model is hybrid. The errors described in the Slate reporting mostly concern written translation, not live speech to speech interpreting. Health systems with mature governance now define their own accuracy metrics and vetted medical-grade AI interpreting is showing strong results.
Why are clinicians reaching for consumer translation apps?
Demand for qualified interpreters outpaces supply and clinical teams work under constant time pressure. Consumer apps fill the gap because they sit in every pocket.
A workforce that cannot cover demand
About 29.6 million people in the US have limited English proficiency (LEP), according to the GSA Translation and Interpretation Services Ordering Guide (December 2025). On the other hand, the Bureau of Labor Statistics counts about 73,900 interpreter and translator jobs in the entire country. That number covers every specialty: courts, schools, conferences and publishing. Hospitals employ only around 7 percent of them (Bureau of Labor Statistics). The math does not hold for a health center at 2 a.m.
Volume targets, not tools
Clinical teams are measured on patient volume. They are rarely measured on the language access tools they receive. When a qualified interpreter is hard to reach, staff improvise encounter after encounter. That is a real source of LEP-related burnout. One emergency physician told Slate that, like his colleagues, he would "copy, paste, throw it onto Google Translate and hope." The intent is good. The tool is not built for care. Consumer apps are not HIPAA compliant and they are not tuned for clinical meaning, as we explained in our post on Google Translate in medical settings.
What does the federal policy shift mean for language access?
Federal support for language access has narrowed. The needs of patients and the obligations of providers have not.
A narrower federal stance
A 2025 executive order designated English as the official language of the United States. Federal agencies now provide translation services for people with LEP only when the law requires it or when the work is mission-critical. The Department of Justice resource site that supported LEP access has been reduced to a single page. Civil rights obligations under Title VI and Section 1557 remain in place for health systems that receive federal funds.
Coverage disruption adds pressure
Language access does not end when coverage changes. As I described in our post on Medicaid disruption and language access, health centers and free clinics keep serving patients who lose comprehensive coverage. They still need an interpreter at every visit.
AI named as the cost-effective path
The Attorney General's implementation memo describes AI and machine translation as "cost-effective methods for bridging language barriers." Cost matters. Yet providers and vendors already know the answer is not AI alone. The sustainable model is hybrid. AI covers the encounters it handles well. Professional interpreters cover the cases that call for them. Human interpreters remain the standard and the foundation of the profession. The real question is when AI, when human and with which tools.
Translation or interpreting: which one is the debate about?
Most errors in the public debate concern written translation, which is text to text. Live medical interpreting is speech to speech and raises different quality questions.
Text to text: discharge letters and written instructions
The examples in the Slate reporting are mostly written. Discharge instructions with clumsy wording. A medication instruction where "hold" was rendered with the opposite meaning. A negation that flipped "take with food" into its reverse. These errors matter because a patient reads them at home, alone. Written translation has one advantage: it can be reviewed before it reaches the patient.
Speech to speech: the live encounter
Interpreting happens in real time, in both directions, with turn taking and clarification. The patient can ask again. The clinician can rephrase. The reporting does not really evaluate AI speech to speech interpreting at the point of care. As Tomer Baum explained in our post on how to measure medical translation quality in the AI era, metrics built for documents cannot grade a live clinical conversation.
Who should measure AI accuracy in healthcare?
The field has not yet agreed on a standard way to measure the safety of AI translation. Health systems that are growing their governance now set their own metrics.
When vendors grade their own work
A language services executive told Slate that health systems often rely on vendors to report their own accuracy. Each vendor uses its own method. Without a shared standard, comparisons across the industry are not possible.
What we see in maturing governance programs
Our experience is different. Health systems building AI governance now ask for their own accuracy metrics. They want results for their languages, their specialties and their encounter types. They want to audit, not only to trust. No Barrier supports this with a customer-level governance layer.
Why legacy metrics fall short
Text overlap scores can rate a reversed instruction as acceptable, because most words still match. Clinical quality rests on meaning: translation accuracy, medical-term retention, semantic adequacy and cultural tone. Each needs its own score.
Three questions to ask any vendor
Leaders can open the evaluation with three simple questions. What is measured on every encounter? What triggers a human review? What happens when a reviewer disagrees with the AI? A vendor with clear answers is running a real quality program. A vendor without them is asking for trust instead of earning it.
How far has AI medical interpreting come?
Large language models moved from word-level substitution to context-level understanding. Independent evaluation of medical-grade AI interpreting reflects that change.
Context changed everything
A few years ago, AI in language access mostly converted words. Today, models read the clinical sentence as a whole. They keep qualifiers, timelines and tone. Our post on what changed in LLM capabilities explains this shift in detail.
What Michal Schuster's study found
In the study analyzed by Dr. Michal Schuster, 91 medical sentences in English and Spanish were tested in both directions. Two external certified medical interpreters assessed every output blindly. No Barrier averaged 6.38 on a 7-point accuracy scale. Two phone interpreting agencies scored 5.59 and 4.93. No Barrier also kept nearly the same quality in both language directions. This is not a case against human interpreters. It shows what a vetted, medical-grade tool adds to their work.
Language and culture travel together
Slate also cited a 2025 study in Nature Human Behaviour. Researchers found that the same chatbot gave answers with different cultural assumptions depending on the language of the prompt. General-purpose AI is not neutral. This is why providers need HIPAA compliant, vetted tools built for clinical meaning, with dialects and cultural tone tested by medical linguists. No Barrier is HIPAA compliant and SOC 2 Type II certified.
Is a possibly wrong translation worse than none?
Some clinicians argue it is. The evidence points to a more useful question: how to give every patient accurate information in their language.
The case for caution
A London hematologist and a translation scholar tested a patient booklet translated into Bengali, once by professionals and once by Google Translate. Assessors scored the professional version 73 out of 100 and the machine version 51. He now holds that handing a patient a possibly wrong translation is ethically worse than giving no written information. The view deserves respect. It is also worth debating.
Why "nothing" is not neutral
A 2024 scoping review of 137 studies found that medical errors affecting LEP patients were more likely to cause physical harm. The same review linked interpreter services to better informed consent and fewer readmissions. Silence carries its own risk. In the London study, comprehension rose most when patients had a live conversation and a bilingual consent form.
Human involvement is not a new idea
Translation scholars quoted by Slate say human involvement stays essential in the LLM era. The field already knows this. Hybrid language access has been in place for years. The goal is not to choose between AI and people. It is to give clinicians vetted tools, professional review for written materials and interpreting at every touchpoint.
From every conversation I have had on Care Culture Talks, one lesson stands out: clinicians do want tools they can trust. The answer to risky apps is not less language access. It is better language access, measured, governed and available to every patient.
.svg){kind=small}