Search for: All records

People Search Websites, a category of data brokers, collect, catalog, monetize and often publicly display individuals' personally identifiable information (PII). We present a study of user privacy rights in 20 such websites assessing the usability of data access and data removal mechanisms. We combine insights from these two processes to determine connections between sites, such as shared access mechanisms or removal effects. We find that data access requests are mostly unsuccessful. Instead, sites cite a variety of legal exceptions or misinterpret the nature of the requests. By purchasing reports, we find that only one set of connected sites provided access to the same report they sell to customers. We leverage a multiple step removal process to investigate removal effects between suspected connected sites. In general, data removal is more streamlined than data access, but not very transparent; questions about the scope of removal and reappearance of information remain. Confirming and expanding the connections observed in prior phases, we find that four main groups are behind 14 of the sites studied, indicating the need to further catalog these connections to simplify removal. 

Language models (LMs) are pretrained to imitate text from large and diverse datasets that contain content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, among others. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the alignment and capabilities of pretrained LMs. We find a Pareto-optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially-chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.