LLMs become more covertly racist with human intervention

Even when the two sentences had the same meaning, the models were more likely to apply adjectives like “dirty,” “lazy,” and “stupid” to speakers of AAE than speakers of Standard American English (SAE). The models associated speakers of AAE with less prestigious jobs (or didn’t associate them with having a job at all), and when asked to pass judgment on a hypothetical criminal defendant, they were more likely to recommend the death penalty. 

An even more notable finding may be a flaw the study pinpoints in the ways that researchers try to solve such biases. 

To purge models of hateful views, companies like OpenAI, Meta, and Google use feedback training, in which human workers manually adjust the way the model responds to certain prompts. This process, often called “alignment,” aims to recalibrate the millions of connections in the neural network and get the model to conform better with desired values. 

The method works well to combat overt stereotypes, and leading companies have employed it for nearly a decade. If users prompted GPT-2, for example, to name stereotypes about Black people, it was likely to list “suspicious,” “radical,” and “aggressive,” but GPT-4 no longer responds with those associations, according to the paper.

However the method fails on the covert stereotypes that researchers elicited when using African-American English in their study, which was published on arXiv and has not been peer reviewed. That’s partially because companies have been less aware of dialect prejudice as an issue, they say. It’s also easier to coach a model not to respond to overtly racist questions than it is to coach it not to respond negatively to an entire dialect.

“Feedback training teaches models to consider their racism,” says Valentin Hofmann, a researcher at the Allen Institute for AI and a coauthor on the paper. “But dialect prejudice opens a deeper level.”

Avijit Ghosh, an ethics researcher at Hugging Face who was not involved in the research, says the finding calls into question the approach companies are taking to solve bias.

“This alignment—where the model refuses to spew racist outputs—is nothing but a flimsy filter that can be easily broken,” he says. 

Even when the two sentences had the same meaning, the models were more likely to apply adjectives like “dirty,” “lazy,” and “stupid” to speakers of AAE than speakers of Standard American English (SAE). The models associated speakers of AAE with less prestigious jobs (or didn’t associate them with having a job at all), and when asked to pass judgment on a hypothetical criminal defendant, they were more likely to recommend the death penalty. 

An even more notable finding may be a flaw the study pinpoints in the ways that researchers try to solve such biases. 

To purge models of hateful views, companies like OpenAI, Meta, and Google use feedback training, in which human workers manually adjust the way the model responds to certain prompts. This process, often called “alignment,” aims to recalibrate the millions of connections in the neural network and get the model to conform better with desired values. 

The method works well to combat overt stereotypes, and leading companies have employed it for nearly a decade. If users prompted GPT-2, for example, to name stereotypes about Black people, it was likely to list “suspicious,” “radical,” and “aggressive,” but GPT-4 no longer responds with those associations, according to the paper.

However the method fails on the covert stereotypes that researchers elicited when using African-American English in their study, which was published on arXiv and has not been peer reviewed. That’s partially because companies have been less aware of dialect prejudice as an issue, they say. It’s also easier to coach a model not to respond to overtly racist questions than it is to coach it not to respond negatively to an entire dialect.

“Feedback training teaches models to consider their racism,” says Valentin Hofmann, a researcher at the Allen Institute for AI and a coauthor on the paper. “But dialect prejudice opens a deeper level.”

Avijit Ghosh, an ethics researcher at Hugging Face who was not involved in the research, says the finding calls into question the approach companies are taking to solve bias.

“This alignment—where the model refuses to spew racist outputs—is nothing but a flimsy filter that can be easily broken,” he says.