arxAIv

arxAIv is a speculative art installation that consists of 100 fully AI-generated computer vision research papers and 156 AI-generated figures, assembled to examine questions of authorship, legitimacy, and automation in academic publishing.

I gathered analytics on the generated corpus — tracking hallucinated affiliations and institutions, common author names and nationalities, recurring features, title similarities, and topic clusters. To map the semantic structure of the collection, I built an embedding-based clustering and visualization pipeline using CLIP embeddings and a 3D force-graph library, producing three distinct force graphs: one using text embeddings, one using figure embeddings, and one comparing AI-generated figures directly against real CVPR paper figures. The result reveals how AI-generated academic content clusters, diverges, and superficially mimics real research.

The full dataset (100 AI-generated papers, 156 figures, and a metadata CSV containing titles, authors, affiliations, detected features, raw text, and GPT-parsed cleaned text) is published on GitHub. The project is presented on an academic-style website with curated paper samples and the interactive embedding graphs.