Study on medical data finds AI models can easily spread misinformation, even with minimal false input