Regulators often play a game of whack-a-mole, banning a specific harmful image or ad only for a new version to appear minutes later. This reactive approach fails because it targets the output rather than the engine. To change this, oversight must move upstream into the development phase of generative models.
First, authorities should mandate transparency in training datasets. If a model is fine-tuned using scraped imagery that prioritizes sexualized poses or specific body types, the harm is baked into the weights of the neural network itself. Regulators can require developers to audit these datasets for inherent biases that facilitate objectification. Instead of judging a single generated image, they should evaluate the mathematical tendency of the model to produce dehumanizing content.
Second, we need standards for
