Windsor Johnston

Tammy Merga reports from San Diego.

Researchers at Oxford and the University of Kentucky tested chat GPT with more than 20 million comparisons between counties, states and cities...

asking, for example, which U.S.

state was lazier.

Initially, the system pushed back against such generalizations.

But when researchers forced it to compare places by rewording the questions, the model tended to reproduce familiar regional stereotypes.

Across categories, it often ranked wealthier Western regions more positively.

Because the system had to pick one place over another, researchers could also track consistent patterns in its responses.

The authors of the study call this the Silicon Gaze.

arguing that large language models reflect inequalities embedded in their training data.

Windsor Johnston, NPR News.

Researchers at Oxford and the University of Kentucky tested chat GPT with more than 20 million comparisons between counties, states and cities.

asking, for example, which U.S.

state was lazier.

Initially, the system pushed back against such generalizations.

But when researchers forced it to compare places by rewording the questions, the model tended to reproduce familiar regional stereotypes.

Across categories, it often ranked wealthier Western regions more positively.

Because the system had to pick one place over another, researchers could also track consistent patterns in its responses.

The authors of the study call this the Silicon Gaze.

arguing that large language models reflect inequalities embedded in their training data.