Holly Herndon and Mat Dryhurst told The Atlantic that the future of AI art does not have to belong to slop. It is a provocation worth taking seriously from two artists who have been building inside AI tooling longer than most critics have been writing about it. The same week, a 2026 arXiv paper by Keluskar, Bhattacharjee, and Liu titled When Does Personality Composition Matter for Multi-Agent LLM Teams? found that personality prompting, the practice of telling an AI to behave like a creative or a critic or a collaborator, produces surprisingly inconsistent behavioral shifts. The model's underlying character is sticky. Which raises the question Herndon and Dryhurst are actually asking: if you cannot reliably reshape the AI's personality, can you still use it to make something non-generic?
The Slop Default and How to Override It
Slop is the term of art for AI-generated content that is technically competent and aesthetically dead. It is not a bug in the model. It is the statistical center of gravity of the training data, the most average version of everything the model has ever seen. Herndon and Dryhurst's practice is explicitly about working at the edges, using AI as a collaborator with intentional constraint, not as an image or text dispenser. The Atlantic piece positions this as optimistic, and it is, but it is also a description of significant labor. Most people using generative tools are not building bespoke workflows. They are clicking generate. The slop is structural, not accidental.
What Taste Does in an AI World
A 2026 arXiv paper by Mahadevan, Odyssey: Constructing Verifiable Local Truth-Preserving Foundation Models, proposes frameworks for keeping AI outputs locally coherent and verifiable. It is a technical paper, but its cultural resonance is direct: the problem of AI output is not power but undifferentiation. Kyle Chayka's Filterworld argument that algorithmic systems produce homogenization applies at the generation layer too. Herndon and Dryhurst are arguing for taste as the override mechanism. The arXiv paper is arguing for architecture as the override mechanism. Both are bets against the mean.