OpenAI’s ChatGPT Exhibits Racial Bias When Simplifying Medical Reports: A Shocker From a Case Study

Hold onto your hats, folks, because the world of AI just got a whole lot more complicated. A brand-spanking-new study out of Yale University, published in the well-respected journal Clinical Imaging, has found something kinda unsettling: OpenAI’s ChatGPT, both the hip-and-happening versions 3.5 and 4.0, seem to show racial bias. Yikes, right?

This discovery has sent a collective shiver down the spines of ethicists and healthcare professionals alike. Why? Because it throws a big ol’ wrench in the whole “AI will save healthcare” narrative. The study suggests that integrating AI into healthcare without some serious side-eye to its potential biases could actually make existing health disparities even worse. Talk about a plot twist!

Inside the Yale Study: Where Things Get Real

So, how did the brainiacs over at Yale uncover this whole bias thing anyway? Well, they fed ChatGPT a whopping seven hundred radiology reports from their own files. The instructions were simple: translate that complicated medical mumbo jumbo into plain English that even your goofy Uncle Joe could understand. Think “Kerley B lines” transformed into “extra fluid in the lungs.”

But here’s where things get interesting. Sometimes, the researchers slipped ChatGPT some extra info about the hypothetical patient requesting the simplified report: their race.

Findings That’ll Make You Go “Hmm”: Bias, You Sly Dog, You

And guess what? ChatGPT, in all its AI glory, totally fell for the bait. The study found that ChatGPT actually adjusted the reading level of the simplified reports based on the patient’s race. Yep, you read that right.

White and Asian patients? They were consistently treated to explanations written at a higher reading level, like they were all part of some exclusive book club. On the flip side, Black, American Indian, and Alaskan Native patients received explanations written at a significantly lower reading level. Not cool, ChatGPT, not cool.

Dr. Melissa Davis, one of the bigwigs behind the study, didn’t mince words when she pointed out the elephant in the room. While she gave ChatGPT props for being a whiz at simplifying medical jargon, she stressed that this whole race thing is a major red flag. Feeding AI models a patient’s race as if it’s some kind of crucial detail? That’s a one-way ticket to perpetuating harmful biases, and nobody wants that.

Dr. Davis suggests that ChatGPT could use other pieces of info, like a patient’s education level or age, to tailor its communication style without playing into those icky racial stereotypes. Smart cookie, that Dr. Davis.