With LLMs, arguably the code is much less decisive for the specific output than the model, which is data. An interesting question to ask is which human labor contributed to the model. The training data probably would constitute a significant share in that.