PROBLEMS WITH AI TRAINING: SOURCES AND CONSEQUENCES OF HALLUCINATION IN NEURAL NETWORK MODELS
Contemporary AI language models are increasingly deployed in safety-critical contexts, yet they remain prone to generating dangerous or factually incorrect outputs. Documented examples include AI-generated recipes suggesting the use of chlorine-producing ingredient combinations [1]. Rather than acknowledging these as model errors, developers have characterized them as “hallucinations”—a term that obscures the underlying technical causes. This paper investigates the structural and methodological sources of hallucination in neural network models, examines three predominant training paradigms, and argues that incomplete training data, flawed train–test partitioning, and the inaccessibility of expert tacit knowledge are root causes of unreliable AI outputs. Recommendations for more rigorous training methodologies are proposed.