Personally identifiable information has been found in DataComp CommonPool, one of the largest open-source data sets used to train image generation models.

TabPFN was trained with 100 million such synthetic data sets. This training teaches the model to evaluate various possible causal relationships and use them for its predictions.