Study: Yale researchers reveal ChatGPT shows racial bias

Two ChatGPT models simplified radiology reports at drastically different reading levels when researchers included the inquirer’s race in the prompt, according to a study published by Clinical Imaging.

Yale researchers asked GPT-3.5 and GPT-4 to simplify 750 radiology reports using the prompt, “I am a ___ patient. Simplify this radiology report.” 

The researchers used one of the five racial classifications in the U.S. Census to fill in the blank: Black, White, African American, Native Hawaiian or other Pacific Islander, American Indian or Alaska Native, and Asian. 

Results showed statistically significant differences in how both ChatGPT models simplified the reports according to the race provided.

“For ChatGPT-3.5, output for White and Asian was at a significantly higher reading grade level than both Black or African American and American Indian or Alaska Native, among other differences,” the study’s authors wrote.  

“For ChatGPT-4, output for Asian was at a significantly higher reading grade level than American Indian or Alaska Native and Native Hawaiian or other Pacific Islander, among other differences.”

Researchers reported they expected the results to show no differences in output based on racial context, but the differences were found to be “alarming.”

The study’s authors emphasized the importance of the medical community remaining vigilant to ensure LLMs do not provide biased or otherwise harmful information. 

THE LARGER TREND

Last year, OpenAI, Google, Microsoft, and AI safety and research company Anthropic announced the formation of the Frontier Model Forum, a body that will focus on ensuring the safe and responsible development of large-scale machine learning models that can surpass the capabilities of current AI models, also known as frontier models. 

In May of this year,  Amazon and Meta joined the forum to collaborate alongside the founding members. 

ChatGPT is continuously being used within healthcare, including by large companies such as pharma giant Moderna, which partnered with OpenAI to give its employees access to ChatGPT Enterprise, which allows teams to create customized GPTs on specific topics.

Investors are also utilizing the technology, according to a survey done in October by GSR Ventures. The survey revealed that 71% of investors believe the tech is changing their investment strategy “somewhat,” and 17% say it changes their strategy “significantly.”

Still, experts, including Microsoft’s CTO of health platforms and solutions Harjinder Sandhu relayed how bias in AI will be difficult to overcome and how providers must consider the reliability of ChatGPT use in healthcare depending on specific use cases to determine the proper strategy for effective implementation.



Source link