Researchers build an AI detector that works across completely different AI architectures

A new method learns to detect when language models have memorized training data by training on synthetic examples, rather than using hand-coded rules. This detector works on transformer models, state-space models, recurrent models, and other architectures that share no underlying design — suggesting there's a universal signature of memorization that any neural network leaves behind when it overfits.

Until now, detecting memorization required separate custom detectors for each model type, built by hand by someone guessing at what matters. This paper shows the signature is learnable and transfers: train once on transformers, deploy on Mamba or RecurrentGemma with no retuning and better accuracy than baselines. The practical implication is immediate — companies training or deploying large language models can now use a single detector to audit whether their models have leaked training data, regardless of architecture. This matters because memorization is a real privacy risk: if your model memorized your proprietary training data, a user can extract it.