A new paper argues that the fix for AI style theft isn't erasing artist concepts from image models - it's purifying them at the moment of generation.
The paper, 'From Concept Erasure to Style Purification: Contrastive Eigenbases for Artist Style Protection' (arXiv:2602.08059), points out that most model-side defenses against style mimicry work like ordinary concept erasure: delete or redirect a target style inside the model, the same way you'd scrub an object or a trademark. The authors ran a causal intervention analysis and found that approach doesn't fit styles well, because artist styles aren't compact, localized units the way objects are. Their alternative, CAPE (Contrastive Artist Style Purification with Eigenbases), skips retraining altogether. It builds contrastive triplets around a generation request, solves for the style-specific directions in the model's attention math, and suppresses those directions - brushstrokes, textures, color handling - on the fly during inference, while leaving the requested content and composition untouched.
That distinction matters because most people worried about AI style mimicry aren't worried about a model that can't draw 'a castle' - they're worried about it drawing a castle that looks unmistakably like a specific living artist made it, cheaply and without consent. A training-free, inference-time filter is also easier for image platforms to deploy than the retrain-and-patch cycle current erasure methods require, since model updates rarely run on the same schedule as legal pressure.
It's a narrower, more surgical answer to a problem that lawsuits and blunt prompt filters have so far mostly failed to solve.