Wide range screening of algorithmic bias in word embedding models using large sentiment lexicons reveals underreported bias types

Bias in AI is more extensive than commonly reported: The study identified novel types of algorithmic bias in popular word embedding models that have been largely overlooked in previous research, including negative associations with working-class socioeconomic status, senior citizens, male children, and conservative political views—suggesting the field has significant blind spots in what it measures.

Gender bias is more complex than "anti-female" narratives: Contrary to widely reported findings, the research found that gender bias in embedding models is "multifaceted and often reversed in polarity," indicating that simplistic characterizations of AI bias may miss important nuances about how different demographic groups are actually represented.

Racial bias against African-Americans remains systemic: The analysis confirmed previous findings showing consistent negative bias against names popular among African-Americans across most embedding models examined, demonstrating this particular bias persists as a widespread problem despite increased awareness.

Researcher bias affects what gets studied: The paper suggests that "widely held blind spots when searching for algorithmic bias" and lack of standardized terminology may explain why certain bias types go unreported, indicating that researchers' own assumptions shape which demographic groups get protected—potentially leaving vulnerable populations undefended.

Word embeddings carry multiple loaded associations: Popular AI models simultaneously embed negative associations with numerous demographic characteristics (religious faith, physical appearance, age, class status), meaning a single AI system can perpetuate harm across many dimensions simultaneously.