Errors are the first concern providers raise about AI medical interpreting and it is the right concern. An omitted allergy question or a misheard dose can change a clinical outcome. Yet the research on interpreting errors points somewhere leaders rarely look first. Errors are documented across medical interpreting, including qualified professional interpreting, which remains the standard of care. The highest clinical risk sits in encounters with no qualified interpreter at all, where a relative, an untrained staff member or no one fills the gap. This article reviews what four peer-reviewed studies measured, where risk concentrates and how AI medical interpreting, used as an additional tool, helps the standard reach more patients.
TL;DR
- Interpreting errors are measured and documented even among qualified professional interpreters, which is why the profession invests in training and certification.
- Clinical risk concentrates in ad hoc and uninterpreted encounters: in one pediatric emergency study, 22% of ad hoc errors and 20% of errors with no interpreter carried potential clinical consequences, against 12% for professionals.
- Professional human interpreting stays the standard. AI medical interpreting is an additional tool that extends that standard across hours, languages and channels.
- Ai medical interpreting is language access at scale
- Leaders can evaluate every interpreting resource with the same discipline the research uses: error types, clinical significance, language-level validation and governance.
Why do health systems worry about AI interpreting errors?
The concern is legitimate and deserves a precise answer. It usually comes from three places:
- experience with general-purpose translation apps,
- uncertainty about who is responsible for an error
- and the lack of a shared way to measure quality.
The consumer app reference point
Many clinicians first met AI translation through a consumer app on a phone. Those tools were built for travel and everyday conversation, not for a medication review or a consent discussion. They are not trained on clinical terminology, not validated for care settings and not designed for HIPAA. When leaders picture AI in the exam room, that app is often the reference point and it is the wrong one.
Who stands behind an error
Responsibility is the second source of concern. An error from a tool can feel harder to accept than an error from a person, because it is less clear who stands behind it. That is a fair question to ask any vendor and the answer should be explicit: responsibility for an interpreting error sits with the interpreter, human or AI and the vendor must stand behind its output.
The missing yardstick
Measurement is the third. Without a shared yardstick, a single AI error reads as proof the technology is not ready, while thousands of clear encounters go unnoticed. Leaders need a consistent way to evaluate quality, the same way the field has evaluated interpreting for more than two decades.
Built for care, held to the same bar
AI medical interpreting built for care answers these concerns by design. It is trained on medical terminology, backed by medical linguists, validated language by language, governed at the organization level, HIPAA compliant and SOC 2 Type II certified. Used as an additional tool, it does not ask anyone to accept a lower bar. It asks for the same bar, applied with the same rigor.
No interpreting resource is error free and none should be judged as if it were. The useful questions are how errors are measured, where they carry clinical weight and how a program reduces them. The research on professional interpreting offers that framework.
Why do errors occur even with qualified interpreters?
Medical interpreting is a demanding cognitive task performed in real time and the research documents errors at every level of training. That finding is not a criticism of interpreters. It is the reason the profession is built on qualification.
Researchers study interpreting errors by recording real encounters, transcribing each utterance and rating every deviation for clinical significance. Using that method, a 2003 study in Pediatrics of hospital outpatient encounters counted an average of 31 interpreter errors per encounter. Omission was the most frequent type at 52%, followed by false fluency, substitution, editorialization and addition.
Omissions deserve particular attention because they are the hardest errors to catch. A clinician who does not speak the patient's language hears a fluent answer and has no way to know that a question about allergies never reached the patient. That is why error research relies on full transcripts rather than impressions from the room.
Later studies refined the picture. In a 2015 analysis in Medical Care of 32 primary care visits, about a quarter of speech segments handled by in-person professionals contained an inaccuracy (25%), with a similar share for professional video interpreting (23%).
Across all visits, 7.1% of errors were rated moderately or highly clinically significant. A 2026 exploratory study by Feinberg and colleagues in the journal Healthcare found that 4.8% of physician speech segments rendered by qualified in-person interpreters carried a clinically significant error.
Training is the variable that moves the numbers. Among professional interpreters in a 2012 study in Annals of Emergency Medicine, those with 100 or more hours of training made a median of 12 errors per encounter, compared with 33 for those with less training. Only 2% of their errors carried potential clinical consequences, against 12% for their less trained peers. Years of experience alone did not show the same effect.
Methods, settings and languages differ across these studies, so the figures indicate direction rather than a precise benchmark. The direction is consistent: interpreting carries measurable error, and qualified, well trained professionals reduce its clinical weight.
Read more about Medical Phone Interpreting here.
Where does clinical risk concentrate?
The same body of research shows that the steepest risk appears when no qualified interpreter is present. That is a coverage problem, not a performance problem.
In the 2012 emergency department study, 12% of professional interpreter errors had potential clinical consequences, compared with 22% for ad hoc interpreters and 20% when no interpreter was used at all. The 2015 primary care study found inaccurate interpreting in 54% of speech segments handled by ad hoc interpreters, roughly double the professional rate. The 2003 study recorded that 77% of ad hoc interpreter errors had potential clinical consequences. Its examples included an omitted question about drug allergies and incorrect instructions for antibiotic dosing.
Ad hoc interpreting rarely reflects a deliberate choice. It reflects a gap: the 2 a.m. admission after the staff interpreter has gone home, the third encounter needing Haitian Creole at the same moment, the call center queue or the less common language a local agency cannot staff on short notice. With 29.6 million people in the United States living with limited English proficiency (GSA, Translation and Interpretation Services SIN 541930 Ordering Guide, Dec 2025), those gaps add up to a large number of encounters.
For leaders weighing error risk, this reframes the question. Concern about a new tool introducing errors is legitimate and every tool deserves scrutiny. The encounter with no qualified interpreter, however, is already the documented high risk scenario, and it happens every day.
How does AI medical interpreting fit alongside the standard?
Professional human interpreting is the standard and stays the standard. AI medical interpreting is an additional tool whose job is scale: bringing medical interpreting into the encounters the standard cannot physically reach.
The role that does not change
As explored in why human interpreting remains the foundation of care, skilled interpreters carry context, emotion and cultural nuance that matter most in sensitive conversations: a serious diagnosis, an end of life decision, a consent discussion with family in the room. That role does not change.
Reach at every touchpoint
What changes is reach. AI medical interpreting is available instantly at every hour, across a wide range of languages and dialects, at every touchpoint where a patient and a care team need to understand each other: the bedside, the phone line, the telehealth visit and the discharge letter. Each of those encounters becomes one that no longer defaults to a relative, a bilingual colleague without training or no interpreter at all.
A layered model
The practical model is layered. In-house and professional interpreters focus on the conversations where their presence changes the outcome. AI medical interpreting absorbs the breadth: overnight volume, simultaneous requests, routine calls and languages the local workforce cannot cover. Each resource is used where it delivers the most value, and fewer encounters fall into the ad hoc gap the research identifies as highest risk.
Protecting the interpreting workforce
The layered approach also protects the interpreting workforce. In-house interpreters are an asset: they know the workflows, the providers and the patients. Asking them to cover every language at every hour stretches that asset thin and adds to the fatigue that any demanding cognitive work is vulnerable to. An additional tool that carries the breadth lets them bring their full skill to the encounters that need it most.
How should leaders evaluate error risk in any interpreting resource?
The research already provides the yardstick. Leaders can apply the same measurement discipline to every interpreting resource in their program, AI medical interpreting included.
- Ask how errors are classified. A credible evaluation uses an established error taxonomy (omission, addition, substitution, editorialization, false fluency) rather than a single accuracy percentage.
- Ask how clinical significance is rated. Clinician review of transcribed or back-translated output, as in the studies above, separates a harmless word choice from a missed dose.
- Ask for results by language. Quality can vary from one language to another, so validation should cover the languages the patient population actually speaks, including less common ones.
- Ask who is responsible for errors. Responsibility for an interpreting error sits with the interpreter, human or AI. A vendor should stand behind its output and show how it monitors and improves quality over time.
- Ask how deployment is governed. Clear rules, set at the organization level, on which encounters use which resource keep sensitive conversations with the right interpreter.
- Ask about security. Interpreted encounters contain protected health information, so HIPAA compliance, a signed BAA and SOC 2 Type II certification are baseline requirements.
- Ask how quality is measured after go-live. A one-time validation is a starting point. Ongoing review of real encounters, with feedback from clinicians and patients, shows whether quality holds across sites, specialties and shifts.
What does a program built around error risk look like?
A program that takes errors seriously combines qualified interpreters, AI medical interpreting and governance, then measures results over time.
No Barrier is built for that layered model: AI-first medical interpreting in the same interface, backed by medical linguists, with 295+ language access options available 24/7, including night shifts. A customer-level governance layer, with tools such as "Point your language" front desk signage, helps each organization define how interpreting is used across its sites, and a dedicated Customer Success team supports every rollout.
Measurement closes the loop. Leaders can track how often encounters still fall to ad hoc interpreting, which languages and hours generate the most requests and where staff interpreters spend their time. Those numbers show whether the program is reducing exposure to the highest risk scenario the research describes.
Deployments span 300 sites in 17 states, from pediatrics to urology to oncology, in clinical and administrative settings. No Barrier is also referenced in a 2026 NEJM Catalyst paper and was selected for the 2026 NACHC/ScaleHealth Accelerator Cohort.
The takeaway
Interpreting errors are not new and they are not unique to any single tool. AI medical interpreting can make errors too, and the right response is not fear but control: knowing how errors are measured, where they carry clinical weight and which safeguards limit their impact on care.
No Barrier has built that framework into every deployment, with medical linguists behind the AI-first interpreting, a customer-level governance layer that defines how interpreting is used across sites and a dedicated Customer Success team that supports quality over time. The result is an additional tool that lets the human standard reach further.
Reach out to the No Barrier team to discover how the framework works in practice.
.svg){kind=small}