Mislabeling and related categorization errors are part of broad problem in data science and create particular problems for AI training sets:
Quach, Katyanna (1 April 2021). Turns out humans are leading AI systems astray because we can’t agree on labeling. The Register. London, United Kingdom.
Northcutt, Curtis G, Lu Jiang, and Isaac L Chuang (2021). Confident learning: estimating uncertainty in dataset labels — Manuscript. arXiv 1911.00068. Version 4.
Quach (2021) also suggests online crowdworkers brought in to check on label accuracy have an incentive to stay with the status quo. Northcutt et al (2021) indicate that the same effort directed to establishing confidence in model predictions should also be directed to establishing confidence in label quality.