The sociological applications of supervised machine learning, already well proven in industrial/
business applications, raise specific questions. The reason for this specificity is that in these applications, the algorithm is tasked with learning complex concepts (e.g. whether a tweet contains hate speech). Supervised learning consists of learning to classify previously annotated (hate
speech/non-hate speech) texts by the algorithm, looking for characteristic text patterns. The
questions that arise are: how to prepare annotation? How can a hermeneutic challenge such as
hate speech recognition be performed by annotators? Are routinely applied, detailed annotation
guidelines helpful? The article also discusses how large companies perform coding on crowdsourcing platforms, and describes AI bias, which in this case means that annotators themselves
introduce bias into the data. I illustrate these issues with our own research experiences.