6.9 Artifacts, process, and evaluation

Computational-creativity research asks more than whether an output differs from its training examples. Ritchie proposes empirical criteria relating a program, its inspiring set, and its generated artifacts [ Ritchie , 2007 ] . Colton’s “creative tripod” asks a system to exhibit skill, appreciation, and imagination, thereby treating generation and self-evaluation as distinct capacities [ Colton , 2008 ] . Jordanous’s SPECS methodology begins by declaring what creativity means for the system under study and then deriving tests from that declaration [ Jordanous , 2012 ] .

Miller’s cross-domain survey provides a complementary historical snapshot of AI-powered practice in visual art, music, literature, poetry, and performance before the present generation of frontier models [ Miller , 2019 ] . Its breadth is useful here because it shows how many different artifacts had already been called creative by 2019. It is a map of systems and practices, rather than a single computational theory by which their creative mechanisms can be identified.

These approaches anticipate two principles of this book. Evaluation must be declared rather than improvised after seeing the output, and the process can matter as much as the artifact. A system that accidentally emits a valuable formula is not equivalent to one that represents its assumptions, diagnoses a failure, proposes a change, and records why the change survived testing.

The status-bearing extension dossier is consequently both an artifact and a process trace. It records the prior theory, obstruction, proposal, transport, predictions, and epistemic status. The package does not settle the philosophical question of whether the system is “really” creative. It supports an empirical judgment about a specified capacity while preserving the evidence needed for another evaluator to disagree.