Evaluating LingualAI: a prospective validation of AI-based real-time translation against certified human interpreters

Limited English proficiency affects over 25 million people in the United States and is associated with disparities in healthcare access, safety, and outcomes. We conducted a prospective, within-subject, simulation-based comparison to evaluate whether an in-house AI application (LingualAI) achieves non-inferior translation quality versus certified medical interpreters in English-Spanish otorhinolaryngology encounters. Standardized clinician-patient scripts were translated by LingualAI and by certified interpreters, and bilingual clinicians rated anonymized audio across multidomain quality measures. Using a prespecified non-inferiority margin of 0.30 points (Human - AI) on 5-point scales, LingualAI met non-inferiority for 2 of 3 primary factors (terminology accuracy Δ = 0.07; adequacy of meaning Δ = 0.13) but not clarity (Δ = 0.50). It met non-inferiority for 1 secondary factor (completeness Δ = 0.14), while grammar (Δ = 0.21; upper 95% CI = 0.34), vocabulary (Δ = 0.18; upper 95% CI = 0.32), and cultural appropriateness (Δ = 0.39) exceeded the margin. No voice-related factors met non-inferiority (fluency Δ = 1.13; prosody Δ = 0.59; pacing Δ = 0.40), and conclusive ratings favored interpreters (overall quality Δ = 0.58; clinical confidence Δ = 0.61). These findings suggest LingualAI preserves core clinical meaning and terminology but remains limited by speech naturalness and delivery, supporting use as an adjunct when interpreter access is constrained and favoring interpreter-in-the-loop deployment for higher-stakes communication.