The Files are in the Computer: Copyright, Memorization, and Generative AI

The model as a copy

Is a trained model a copy of its training works?

A copy where it memorises works

A model that has memorised training data is a copy of that data; not all learning is memorisation

What it says

Drawing on the technical literature, the authors define memorisation as the possibility of reconstructing from the model a near-exact copy of a substantial portion of a training item. Memorisation happens in training and is latent in the model itself, so a model that has memorised training data is a 'copy' of it in the copyright sense; outputs that regurgitate are a symptom, not the cause.

Whether a model is a copy turns on what it has memorised, which depends on training choices.

Sources