🎭 arxAIv

What happens when you don’t just use AI to help write parts of a research paper, but hand it the whole thing?

arxAIv is a speculative installation of 100 AI-generated computer vision papers, and a parody of how fast generative AI and academic publishing are collapsing into each other. Drafting an abstract or smoothing prose with a language model is mundane by now, so this work skips ahead to the endpoint: titles, figures, affiliations, citations and formatting, all conjured out of prompts.

From a distance the illusion is polished. The titles sound technical. The layouts mimic top-tier conferences. Affiliation lines mix real institutions with subtle fakes. Diagrams look structured and data-driven. At a glance, it’s uncomfortably easy to mistake these for real submissions.

Look closer and the cracks show. The figures unravel into visual noise. Affiliations contradict themselves. References cite ghosts. The structure holds up; the content doesn’t, which says something about how far AI still is from imitating scientific rigor rather than its appearance.

That gap is the point.

arxAIv visualizes the gap with a series of semantic and visual embeddings. Title clusters come from all-MiniLM-L6-v2, animated in 3D with 3d-force-graph into orbiting clouds of thematic similarity. Figure embeddings come from OpenCLIP ViT-B-32, mapped to show how real scientific figures cluster tightly while the AI-generated ones drift apart. The contrast is easy to see: human-made figures have internal logic, and the synthetic ones tend to slide into abstraction or contradiction.

Layering fake papers, hallucinated figures, and surreal affiliations into graphs, stats, and animations does more than parody academic aesthetics. It shows how quickly those aesthetics can be mimicked, and how blurry the line between rigorous work and visual imitation could get.

None of this is a celebration of generative AI. It’s a checkpoint. If we already accept AI-generated fragments in scientific work, how far are we from letting it author the whole thing? And when we get there, how will we know what to trust?