causal

Probabilistic Models Can Amplify Under-Representation

Because LLMs are probabilistic and predict the most likely next item, an image model asked for a doctor might default to the majority demographic and thereby produce even more under-representation than actually exists, which is one legitimate reason to deliberately curate training data toward greater diversity.

causalpending

Speaker

Megan McArdle

Evidence Quote

“And you could actually get even more under-representation than exists. Right? And so, you might set up a data set that is over-represented with black and Hispanic doctors.”

Source

What Does "Unbiased" Mean in the Digital World? (with Megan McArdle)— EconTalk
Created: 6/13/2026, 12:19:17 AM

Your notes

Loading notes...