All AI-mediated artifacts are forgeries of an original.

Every generated token becomes part of the conditioning context for the next. The model's learned distribution determines which continuations are probable, and each continuation changes the conditions under which subsequent probabilities are evaluated.

So the process is inherently path-dependent. The model doesn't simply select from a fixed distribution of completed answers. It progressively constructs the context that determines which subsequent answers become probable. And because its conditional distributions reflect training, each step tends to propagate learned regularities into the next step.

The generator operates within a learned probability landscape, and its successive commitments can reinforce established trajectories. Sampling can introduce variation without necessarily escaping the underlying learned structure.

The consequence is that novelty is not the objective of the generative process; it is an occasional outcome of a process optimized for other criteria.