Less, but Verified: A Rams-Derived Framework for Evaluating Image Provenance in the Age of Generative AI

Generative image models have severed the historical link between a picture and the hand or lens that produced it. In response, industry, regulators, and researchers have assembled a provenance stack of signed metadata, invisible watermarks, statistical detectors, and distributed ledgers. Each component is usually evaluated in isolation and against narrow technical metrics such as robustness or false-positive rate. This article argues that provenance is not only a security problem but a design problem, and that it should be judged by how well it serves the people who make images and the people who look at them. Drawing on Dieter Rams's ten principles of good design, I derive ten evaluation criteria for provenance systems and apply them to four families of mechanisms: C2PA Content Credentials, invisible watermarking, synthetic-image detection, and blockchain registries. The analysis is informed by a model borrowed from defense logistics, where the National Stock Number identifies what an item is but cannot, on its own, establish where a particular part came from. Four harms structure the evaluation: loss of artistic credit, deceptive synthetic media, loss of metadata in distribution, and provenance as surveillance. I find that no single mechanism satisfies the criteria, that robust designs are layered rather than monolithic, and that the artist, not the platform, should serve as the root of trust. The article closes with concrete recommendations for artists, platforms, standards bodies, regulators, and the institutions that exhibit and publish visual work.