Chen Liying reviews Vector Media

Leonardo Impett and Fabian Offert. Vector Media. Minneapolis: University of Minnesota Press, 2026. 176 pp.

8 October 2026

Much criticism of artificial intelligence begins where the machine ends: with the image generated, the sentence produced, the stereotype reproduced, or the artwork appropriated. Leonardo Impett and Fabian Offert’s Vector Media asks us to reverse the direction of inquiry. Their “radical recalibration of our critical focus” moves “from machine images to image machines, from external to internal representations, or, more precisely, from pixel space to vector space” (p. 17). Pixels record an image as a grid of color values; a learned vector represents it through numerical coordinates that allow a model to compare it with other images or words. The shift matters because the model’s internal organization establishes which resemblances and differences count, shaping what it can recognize or generate. The result is an unusually ambitious intervention in critical AI studies: a media theory not simply of artificial images but of the representational machinery through which culture becomes computationally intelligible.

Machine vision, Johanna Drucker emphasizes in her introduction, should not be mistaken for vision in any phenomenologically substantial sense: what automated systems largely achieve is “picture processing” grounded in historically specific pictorial and photographic conventions (p. 5). Impett and Offert join this insight to Philip Agre’s critical technical practice, treating technical architectures as philosophical propositions rendered operational (p. 30).[1]

This wager is most productive when the authors move beyond datasets toward models. Ideology, they argue, inhabits not only representational content: bias is “embedded in the logic of representation itself” (p. 38). Here the ideology is a universalism that treats culturally situated judgments as neutral measures of the world. A model trained to identify an image’s “correct” orientation, for example, makes the photographic convention of ground below sky into a general standard, although it would make little sense for a microscope slide (p. 37). OpenAI’s CLIP (Contrastive Language–Image Pre-training), a model that compares images and words, likewise identifies Paris most strongly with its center and Johannesburg with its poorest areas, reproducing social stratification as computational resemblance (p. 147). Such assumptions shape how the system represents an image and what it can recognize. Their machine Bildwissenschaft—a critical study of the image theories embedded in machine vision—thus treats computational architectures as cultural artifacts, asking what assumptions about vision, similarity, and knowledge make classification possible.

The book’s conceptual center is “epistemic compression.” Its strongest formulation is concise: “embedding is knowledge, and compression is the route to embedding” (p. 82). This makes embedding, projection, and dimensional reduction objects of humanistic criticism because they decide what counts as proximity, difference, information, and intelligibility.

The argument culminates in the concept of “neural exchange value.” Neural networks, the authors contend, construct spaces in which heterogeneous media become commensurable; their aim is “precisely to produce neural exchange value” (p. 123). The consequence is “media collapse”: vector media become “a kind of violent meta-medium,” translating text, images, sound, and other modalities into a common computational space (p. 145). Lev Manovich has already described how digital software made common operations applicable across different media.[2] The further change lies in models such as CLIP, introduced in 2021, that learn a shared measure of similarity between images and words: a verbal query retrieves pictures by their calculated proximity. What is recent is the use of such learned comparisons at scale to organize recognition and retrieval across media. Collapse names the subordination of media differences to that measure, not their disappearance from cultural experience.

Here the book is most illuminating—and most vulnerable. If embeddings make cultural objects commensurable, when does geometric comparability acquire the social force of exchange value? The authors themselves acknowledge the epistemic incommensurability of media forms and the incompatibility of competing vector spaces. Yet these limits leave the question unresolved: neural exchange value may name less an achieved universal equivalent than an attempt to make computational commensurability operate as exchangeability.

These tensions sharpen rather than diminish the book’s significance. Its larger achievement is methodological: criticism must become technically literate without surrendering historical, philosophical, or political judgment. The immediate beneficiaries, on this reading, are companies that develop and supply the models—OpenAI and Meta figure in the book’s examples—and institutions that adopt their classifications. The companies set the measures of relevance and resemblance through which cultural materials become comparable; institutions turn those measures into decisions about which materials to retrieve and circulate. Control over a space of exchange, as the authors suggest, thus exceeds ownership of individual commodities. After Vector Media, the decisive question is not simply what machines can create, but who gains power when culture is organized as though everything could be exchanged with everything else.

[1] See Philip E. Agre, Computation and Human Experience (New York, 1997).

[2] See Lev Manovich, The Language of New Media (Cambridge, Mass., 2001)