Medical Diagnosis AIs Can Reveal the Origin of Their Training Data When Prompted

A recent study demonstrates that medical diagnosis AI systems can be tricked into exposing their training data sources. Researchers crafted prompts that caused the models to reveal dataset identifiers. The vulnerability highlights privacy concerns for patient‑level information. The findings apply to several widely used diagnostic AI platforms. Developers are urged to implement safeguards against unintended data leakage. The paper discusses potential regulatory implications for AI in healthcare. Critics argue that the exposure could undermine trust in AI‑driven diagnostics. Future work will focus on robust methods to hide training data provenance.