A comparative audit of two datasets, LAION-400M and LAION-2B, revealed that as the dataset scale increases, hate content also increases by nearly 12 percent. 

In a recent study titled “Into the LAION’s Den: Investigating Hate in Multimodal Datasets,” researchers examined the impact of scaling datasets on hateful content by comparing two datasets: LAION-400M and LAION-2B. 

The results of the audit revealed that hate content increased by nearly 12 percent as the dataset size grew - an increase measured qualitatively and quantitatively using the Hate Content Rate (HCR) metric.

The finding highlighted the consequences of data scaling in vision-language datasets.