Abeba Birhane's analysis of AI data sets revealed pervasive derogatory terms, leading to MIT taking the data offline, highlighting the urgent need for AI accountability.
Birhane emphasizes that AI models trained on unfiltered internet data often replicate hate and biases; her work aims to uncover these damaging implications.
"We are not evaluating systems for some hypothetical, potential risks in the future," she asserts, noting these audits reveal real problems like racism and sexism in AI.
The findings underscore a troubling trend where larger data sets increase the likelihood of biased AI outputs, particularly mislabeling marginalized communities.
Collection
[
|
...
]