Latent Diffusion Models Encode a 'Neural Economy,' Paper Argues

Most criticism of generative image models circles the same territory: training data, scraping, copyright, whose art was used without consent. Eryk Salvaggio's paper, submitted to arXiv on June 17, 2026, argues that this focus, however justified, risks missing something embedded deeper in the architecture itself.

Salvaggio describes latent diffusion models as a "neural economy," defined precisely as "a contained symbolic system that abstracts social communication into commensurable vectors as it transfers the social sphere into parcels for sale." The argument works through three distinct forms of exchange value: social exchange value (meaning derived from shared communication), neural exchange value (the reduction of specific things into exchangeable computational units), and commodity exchange value. Each component of the diffusion pipeline is analyzed for how it inscribes ideological positions into generated images: CLIP, the autoencoder, U-Net, and classifier-free guidance.

The core contention is that "the ideology behind the diffusion model facilitates a wholesale transfer of the social sphere into commodity through neural exchange value." Focusing narrowly on copyright, Salvaggio argues, risks perpetuating the model's inherent fetishism, treating images as commodities while leaving the underlying conversion mechanism unexamined.

The honest caveat is that this is a theoretical paper. What it does not give you is an empirical demonstration or a roadmap for what a generative image system that centers social exchange rather than commodity logic would actually look like architecturally. Whether the neural exchange value framework yields testable predictions, or remains primarily a critical lens, is a question the paper leaves open.

For practitioners and researchers, the useful move is widening the frame. If the argument holds, auditing training datasets for copyright violations addresses a symptom while the commodity logic runs in the model itself, in the way it compresses and flattens social communication into exchangeable units. That is a different and harder target for both regulation and redesign.