MGI Distinguishes Real from AI-Generated Data

As generative AI models become increasingly sophisticated, producing outputs indistinguishable from human-created content, a critical challenge emerges: determining whether a given data point originated from a model's training set or was generated by the model itself. This problem, termed Member vs Generated Inference (MGI), is particularly complex when models exhibit memorization and reproduce training data. Existing methods, such as membership inference and attribution techniques, often fail at MGI. Membership inference tends to misclassify generated samples as training members, while attribution methods frequently misclassify true members as generated. This failure stems from both approaches relying on likelihood-related signals that are similarly elevated for both training examples and the model's own outputs, making differentiation difficult. To overcome these limitations, researchers propose the Data Circuit Breaker (DCB), a three-stage method that combines complementary signals from a generative model's autoencoder and latent generator. DCB effectively distinguishes between training members and generated samples across various generative models, including image autoregressive and diffusion models. It remains robust even when models produce near-duplicates of training samples and generalizes well to challenging scenarios where new models are trained on generated data.