Why Digital Rights Need a New Operating Model in the Age of Generative AI

The machinery of digital rights — licences, territories, windows, residuals — was designed around two assumptions. Copying costs something, and a derivative work is recognisably derived from its source. Generative systems falsify both, and most of the resulting difficulty follows directly from that.

What breaks first

Territory. A licence granted for one market meets a model trained in another and deployed globally, and there is no point in the pipeline where the territorial limit naturally applies.

Derivation. Residual and attribution regimes assume you can tell what a work was derived from. When the derivation is statistical and distributed across millions of inputs, the question stops having a clean answer, and rights frameworks that depend on one stop functioning.

Consent. Performer and contributor consents were written for reproduction and adaptation. They were not written for a system that learns a style and applies it indefinitely without reproducing anything.

What an operating model has to do

Express permissions in a form a machine can act on at inference time, not only a form a lawyer can read at contract time. Carry those permissions through transformation, so a downstream user inherits the constraints rather than discovering them. And produce evidence — a record that a use was permitted, retained by someone with an interest in its accuracy.

Why this is a commercial problem, not only a legal one

Because the alternative to a workable model is not the status quo. It is that rights holders withhold their catalogues, and that AI products serving regulated industries cannot warrant their inputs. Both are already happening, and both are expensive. The operating model is what unlocks supply on one side and defensibility on the other.