Building reliable AI models requires understanding the people behind the datasets

Correlation with the original Ruddit offensiveness score by race. Annotations by White participants have the highest correlation with the Ruddit score, while annotations by Asian and Black participants are significantly less correlated. Credit: arXiv (2023). DOI: 10.48550/arxiv.2306.06826

Social media companies are increasingly using complex algorithms and artificial intelligence to detect offensive behavior online.

These algorithms and AI systems all rely on data to learn what is offensive. But who’s behind the data, and how do their backgrounds influence their decisions?
In a new study, University of Michigan School of Information assistant professor David Jurgens and doctoral candidate Jiaxin Pei found that …
Read more…….