To really understand the impact on the ecosystem, the best metric to track going forward will be what's the overall prevalence of harms on Meta's platforms (i.e., what % of views are found to be violating), which for the moment CSER provides. Here's an example of hate speech prevalence on FB. 🧵 7/12

Here's a typical precision/recall curve for an ML classifier, the tech that underlies automated content moderation systems today. You can see, for example, that increasing precision from just 80% -> 90% means that you start missing more than twice as much harm (35% vs 15%). No free lunch!🧵 6/12

One of the most relevant stats from CSER is the "proactive rate", which is the % of violating posts Meta took action on prior to a user report. For example, in Q3 2024, of all the posts Instagram found to be bullying & harassment, a whopping 96% were taken down first by automated systems. 🧵 3/12