Art Newspaper argues training data removal tests copyright

The Art Newspaper published an analysis arguing that effective machine unlearning could serve as a valid defense against AI copyright infringement claims. The September 4 article posits that if removing a copyrighted image from training data produces no change in generated outputs, the resulting image cannot be accused of infringement. This legal theory suggests courts may adopt technical causation tests rather than relying solely on training data provenance to establish liability. The argument implies that functional independence from specific source material matters more than initial dataset inclusion. Legal scholars and technologists are currently debating whether current unlearning methods meet this proposed evidentiary standard. The analysis highlights growing tension between binary copyright frameworks and probabilistic AI generation processes. This perspective offers AI developers a potential compliance pathway through verifiable technical remediation. However, critics argue that proving negative causation in high-dimensional latent spaces remains scientifically unproven.