Adversarial Fragility and Language Vulnerability in Clinical AI: A Systematic Audit of Diagnostic Collapse Under Imperceptible Perturbations and Cross-Lingual Drift in Low-Resource Healthcare Settings

Clinical AI systems fail dramatically under minor, invisible image alterations: The study found that diagnostic accuracy in chest X-ray analysis dropped from 89.3% to 62.0% when images were subjected to imperceptible perturbations—changes so small humans cannot see them. This reveals a critical safety gap: AI systems trained on clean, perfect data may collapse in real-world conditions.

Standard safety measures provide inadequate protection: Common defensive strategies like Gaussian smoothing and ensemble voting failed to restore clinical safety when the AI encountered adversarial perturbations, suggesting that current safeguards are insufficient for high-stakes medical applications.

Language diversity creates dangerous diagnostic failures: When clinical cases were presented in Nigerian Pidgin or Yoruba-inflected English rather than standard English, accuracy dropped significantly—even a model specifically designed for African contexts (NatLAS) fell from 85% to 55% accuracy, with diagnosis consistency dropping to just 50%.

Current AI evaluation practices mask real-world vulnerabilities: Clinical AI systems are typically tested only on clean, standardized English-language data, which does not reflect actual healthcare delivery in low-resource settings where multiple languages are spoken and data quality varies—creating a false sense of safety.

Urgent redesign needed for equitable global healthcare: These findings establish concrete evidence that clinical AI systems are not ready for deployment in primary health centers in low-resource regions without fundamental improvements to adversarial robustness and multilingual capability.