Researchers demonstrate that artificial intelligence systems' tendency to generate false information can be systematically controlled by adjusting how frequently different facts appear during training. Heavy-tailed data distributions that naturally minimize rare information reduce hallucination rates, while deliberate oversampling of uncommon facts increases model accuracy on edge cases. This finding challenges the prevailing assumption that AI hallucination is primarily a technical architecture problem rather than a data curation issue. The implications extend beyond computer science into health information systems, where AI increasingly assists with medical decision-making and patient education. Healthcare applications using large language models could potentially be made more reliable through strategic training data management, reducing the risk of AI-generated medical misinformation. However, the trade-off between reducing hallucinations and maintaining comprehensive medical knowledge coverage remains unclear. The research also raises questions about whether current AI safety measures in healthcare adequately account for training data biases. As medical AI systems become more prevalent in clinical settings, understanding these controllable factors in AI reliability becomes crucial for patient safety and trust in automated health information systems.