Even massive, comprehensive text datasets reflect biases present in the real world

Evidence

Levendowski’s pioneering article (and subsequent book-length scholarship) on the relationship between copyright law and the implicit-bias problem in AI training data is foundational for understanding the equity dimensions of contemporary AI-IP debate. Levendowski argued that the chilling effects of copyright on AI training datasets contributes to bias, since developers default to public-domain corpora skewed toward older, whiter, more male material. — , The Braided Reading EPUB, 2242 (#expands, via linked claim)

Couldry and Mejias theorize “data colonialism” as a continuation of historical colonialism’s logic of appropriation, now extracted from human experience itself. The book centers Global South perspectives and demands that platform critique attend to long histories of dispossession. — , The Braided Reading EPUB, 398 (#expands, via linked claim)

Categorization

• corroborated: 2 independent passages support this claim
• XREF: connects to: Machine learning systems are not inherently fair or just by default; Machine-learning bias often stems from unrepresentative training data; Word embeddings serve as novel objects for sociological analysis of bias (+6 more)

Related

• Machine learning systems are not inherently fair or just by default (similarity 0.91)
• Copyright’s chilling effect on AI training data worsens algorithmic bias (similarity 0.90)
• Platform data extraction continues historical colonialism’s logic of appropriation (similarity 0.80)

Broader Connections

• Copyright’s chilling effect on AI training data worsens algorithmic bias #expands (0.50) — , The Braided Reading EPUB, 2242
• Platform data extraction continues historical colonialism’s logic of appropriation #expands (0.45) — , The Braided Reading EPUB, 398