Introduced on day 46
Training-data memorisation
A generative model emitting near-copies of individual training records, which makes generated data a possible leak of the data it was trained on.
In detail
Research has extracted memorised training examples from both language models and diffusion models. Calling generated data "synthetic" therefore does not make it private. Mitigations include deduplicating training data, searching generated outputs for near-duplicates of training rows and applying formal privacy techniques where the risk requires it.