Manovich, L. (2026). A medium that thinks: Generative AI and media cognition. Emerging Media, 4, 225–232. https://doi.org/10.1177/27523543261454458
Summary
Lev Manovich offers a media-theoretical reading of generative AI as an artistic medium, deliberately bracketing the now-dominant ethical and social critiques to ask what kind of medium it actually is. Drawing on McLuhan, Boden, and fifty years of his own practice across painting, photography, and generative code, he proposes six structural properties that characterize AI image generation, the most novel being that it is the first creative medium with built-in cognitive capacity. This “media cognition”—an internalized, encyclopedic knowledge of artistic techniques, conventions, and styles—both empowers the medium and explains its conservatism: because AI has learned how things normally are, it resists the avant-garde gesture of defamiliarization and gravitates toward 19th-century academic and photographic norms. Style and content, Manovich argues, are structurally entangled in these models, complicating the very idea of an “artist’s style.”
Key Contributions
- A six-property framework for analyzing generative AI as a medium rather than as a social or ethical object.
- The concept of media cognition: AI’s internalized practical knowledge of art history, techniques, and conventions.
- Reframing AI as the first artistic medium with intrinsic cognitive capacity, distinct from prior algorithmic/generative art.
- A new axis for categorizing creative practices: whether precise authorial control and editability are essential.
- The claim that style/content entanglement is structural, not incidental, to generative models.
- Positioning AI as the structural opposite of avant-garde defamiliarization.
Methods
Theoretical and historical analysis grounded in media theory and art history. Manovich proceeds through comparative reflection — situating AI against photography, film, generative programming, and avant-garde strategies (Jazz, Cage, Judson Dance Theatre, Surrealism) — and through illustrative probing of mainstream tools, especially Midjourney (including its --sref style reference feature) and Midlibrary’s style catalogues. The focus is deliberately on unmodified, at-scale use rather than experimental critical AI art.
Findings
- Probabilistic generation imposes a non-removable trade-off between variability and precise authorial control.
- AI handles visual surrealism (recombining familiar objects) well but resists logical surrealism — e.g., it struggles to render a cat chasing a dog.
- Outputs cluster around idealized academic and photographic aesthetics: centered subjects, one-point perspective, polished surfaces — an instance of Boden’s “bounded creativity.”
- Naming an artist in a prompt produces “bleed”: Hokusai pulls in waves and Mount Fuji along with style.
- Image-based style references (
--sref) disentangle style from content more cleanly than artist-name prompts. - AI produces homogeneous brushwork and detail across an image, failing to reproduce the hierarchical stylistic distributions characteristic of pre-modern European and East Asian painting.
- Media cognition and aesthetic conservatism are two sides of the same learned distribution over “how things normally are.”
Connections
This paper is an outlier within the topic cluster, which is dominated by empirical studies of AI-generated misinformation, persuasion, and journalistic uptake; Manovich’s media-theoretic register has few direct interlocutors here. The most genuine resonance is with Hepp2026-oi, which similarly takes a theoretical-conceptual approach to generative AI as an infrastructure shaping communication, and tangentially with work on stylistic detection and provenance such as Stanusch2026-ec and Dierickx2026-tw, where the entanglement of style and content has practical stakes for identifying AI outputs.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify