CONCEPT
Hate Content Rate
Abeba Birhane’s empirical metric for measuring the prevalence of hateful material in large-scale AI training datasets—a tool that inverted the industry’s foundational faith by showing that scaling datasets from hundreds of millions to billions of samples does not dilute harmful content but measurably concentrates it.
The Hate Content Rate is the measure
Abeba Birhane and her collaborators developed to quantify what the field had treated as an intuition: that scraping more of the open web for AI training data would eventually average out its toxic fringe. Their investigation of the LAION datasets—LAION-400M and the two-billion-sample LAION-2B subset of LAION-5B, among the most widely used open training corpora for generative AI systems—asked a precise and frightening question: what happens to hateful content as a dataset scales up? The answer inverted the faith. As the dataset grew from 400 million to two billion samples, the Hate Content Rate rose by roughly twelve percent. More data meant more hate, not less. The wider the net cast across the open web, the more of its venom was hauled in. And the consequences were not confined to the alt-text descriptions: models trained on these datasets exhibited racist stereotyping and dehumanizing