Despite the best efforts of developers of large-scale language models (LLMs) to limit racist, sexist and other stereotypes, a new study from Stanford University has found that language models still reproduce extreme racist biases that date back to the antebellum era. Developers are making significant strides to improve their models, but these changes may be superficial - instead of making real improvements, corporations have only gotten better at hiding the problems.

Pratyusha Ria Kalluri, a PhD student at Stanford, notes that these models have not become less racist, but have simply learned to better disguise their prejudices. This is particularly evident in the treatment of African American English (AAE) speakers, who are discriminated against in various areas of life, including employment, education and criminal justice.

In an article for Nature, Calluri and her colleagues Valentin Hofmann, Dan Jurafsky and Shares King demonstrate that LLMs still spread stereotypes about AAE speakers, for example, by calling them ‘lazy’, ‘stupid’ or ‘dirty’. Not only do these biases negatively affect the perception of AAE speakers, but they also have real-world consequences: models are more likely to give less prestigious positions, harsher criminal sentences and even the death penalty to AAE speakers than to Standard American English (SAE) speakers.

The researchers used the method of experimental sociolinguistics to analyse how LLMs react to texts written in AAE and SAE. The results showed that the models were significantly more likely to associate AAE speakers with negative stereotypes from the past that were documented in the Princeton Trilogy, a series of studies conducted in 1933, 1951, and 1969. This indicates that implicit racism is still deeply rooted in modern AI models.

Although the models have become less overt in their racist statements, hidden stereotypes still remain and are even becoming more visible in secret experiments. This poses serious risks when LLMs are used to make life-changing decisions, such as hiring or criminal convictions.

Kalluri calls on companies and policy makers to reconsider the use of LLMs in important areas and do more to combat racial bias. These models are not just technical tools - they have a real impact on people's lives, and ignoring these issues could have serious consequences for society.

Source: https://hai.stanford.edu/news/covert-racism-ai-how-language-models-are-reinforcing-outdated-stereotypes