The article addresses the puzzle of how large language models acquire causal structure despite the limitations of standard formalisms like Judea Pearl's interventionist approach and the Neyman-Rubin framework. It argues that LLMs utilize a specific inductive method known as variational induction, which relies on difference-making logic. During training, models process vast amounts of text from diverse contexts to identify what constitutes a difference-maker or an indifference-maker within word sequences. The analysis examines how architectural components, specifically token embeddings and self-attention mechanisms, facilitate this variational induction process. This logical framework fundamentally parallels the experimental method used in science. In both cases, causal relations are derived by systematically varying individual circumstances to observe their influence on a phenomenon.