ABSTRACT
Some of the most pervasive AI systems in today’s world are deployed in social media platforms. Recommender systems (see e.g., Zhang et al., 2021) are AI systems that deliver a personalized stream of content into each platform user’s daily feed. They do this by observing user behavior on the platform, and learning what users “engage with.” For each platform user, a recommender system observes that user’s behavior, and uses machine learning techniques to form hypotheses about the types of content that particular user typically engages with. It then gives preference to these types of content in that user’s subsequent feed. Content classifiers are AI systems that identify content of specified types on the platform. They can operate on different types of content: a text classifier might identify textual content as “a newspaper article,” or “a positive product review.” An image classifier might identify an image as a picture of a dog, a cat, or a table. Recommender systems make some use of content classifiers. But content classifiers also have a central role in keeping harmful content off the platform (or in flagging problematic content in various ways). Social media platforms all have detailed policies on hate speech, violent content, nudity, and sexual content. The enforcement of these policies partly involves human oversight, by teams of content moderators. But enforcement also involves the large-scale development and deployment of content classifiers. Again, most of these content classifiers acquire their abilities through machine learning methods. They are given large training sets, comprising numerous examples of each type of content they are to identify. From these examples, they construct their own internal definitions of these same content categories, which are sufficiently general to allow them to classify new items they haven’t seen before.
