Because the dominant contemporary approaches to artificial intelligence take an inductive approach, they represent less of a radical departure in human thought and more of an acceleration or super charging of it. Thus, it should not be surprising that currents of human thought that may be less desirable surface in a variety of AI applications.
One example comes from the health care sector where cautions are being raised about the perpetuation of racist assumptions that could lead to the worsening of health disparities for Black patients. Writing for the Associated Press, Garance Burke and Matt O’Brien report on a recent study from the Stanford School of Medicine that found that four common AI systems (ChatGPT, ChatGPT4, Bard, and Claude) reinforced false beliefs that could lead to poor clinical decisions regarding Black patients.
Burke and O-Brien go on to discuss efforts that are underway to improve AI in the medical sector. One initiative at the Mayo Clinic is testing Google’s medicine-specific model (Med-PaLM) which is trained on medical literature. Additional independent testing and refinement of AI models may lead to next generation systems that reduce or eliminate the bias found in current commonly available systems.
It’s interesting because I had an internship of a Medicare Technology Company as a product manager of their AI medical assistant product in China. What their idea was to use that AI product to evaluate the prescription of doctors in small villages’s clinics, who usually don’t have good medical skills as the doctors in cities. When the doctors input the symptoms, ages, medical history etc. to the medical information system they normally use, and input their decision of what medicine/treatments the patient should take; the system that has been linked to the AI will provide a evaluation of what extra information the doctor have all the necessary information from the patient, or if his/her prescription is correct (evaluation includes not excess dosage, the determination of the disease is correct etc.)Interestingly, when I check the backstage data from hospitals, they did not have races. It probably because China does not have such a diverse races.
Additionally, this company is developing a medical LLM (not a business secret)which I think is not containing races as variable as well, because most of the hospitals I knew in China did not record this information (their training data). Combining what experience I had, with this article, I’m curious if the race is really a critical factor of doctor to decide how to cure people? More clinical advice might be needed of whether races will affect the medical decision due to some potential different biological index of different races, but as a non-medical student who might just took a bold guess, I’m really wondering that if we remove the race’s information when letting the AI to do medical decision will make a better result, especially when the race’s factor will brings to mistake results. I assume if the medical LLM from China’s company will be launched, and when it’s having the accuracy as Bard or ChatGPT, we can use it with the LLM from medical Bard or ChatGPT to the same patients in different races, and compare their results, we might be able to have an answer.