Early AI data labeling around late 2022 and early 2023 primarily involved reinforcement learning from human feedback conducted by non-expert annotators