Pinterest uses an AI model to automatically screen comments on Idea Pins for unsafe content, spam, and negative or harmful language before those comments become widely visible. The system works in near-real time as comments are posted. Since the system was introduced, Pinterest has reported a 53% decline in policy-violating comment reports.

he system uses a multi-task DistilBERT model — a compact, efficient version of a transformer-based language model — trained to simultaneously classify comments into three categories: policy-unsafe, spam, and negative sentiment. Comments flow from a real-time message queue (Apache Kafka) through a processing layer (Apache Flink) to the model, which scores each comment and routes it for action. Multi-task learning allows a single model to handle all three classification jobs at once, which is more efficient than running three separate models.