How AI Actually Learns: Training Data, Predictions And Hallucinations

What does an AI system learn from examples, and why can a fluent answer still be wrong? This episode explains training data, parameters, loss functions and the difference between doing well on familiar examples and generalising to new ones. We also explore why language models invent plausible claims, how bias can enter a system and what questions to ask of a new AI product. Read the Taylor Tailored companion explainer for a practical guide to evaluating real capabilities. We follow the steps from data and labels through loss functions, validation and deployment. A dog classifier that learns the background rather than the animal shows why a good test score can be misleading. For language models, we explain fabricated citations, retrieval tools and the gap between fluent output and verified knowledge. The Taylor Tailored article offers questions for judging a real product announcement.