Abstract Generative AI systems like GPT-4 and Med-PaLM 2 are steadily making their entrance into the medical practice in more areas, including documentation and direct contact with patients. Their linguistic fluency and knowledge representation has demonstrated potential but the ethical and clinical implication of implementing such systems in front Thailand of the patient has not been addressed. In this study, the researchers examine how generative AI tools perform on patient queries in a real-world setting, with the researcher looking at five critical dimensions, including clinical accuracy, hallucination frequency, demographic bias, empathy and trustworthiness, and ethical transparency. We have tested 100 standardized queries on GPT-4 and Med-PaLM 2. Responses were evaluated on the basis of expert judgment, demographic bias testing, empathy and readable metrics, human rater judgment. The findings indicate that GPT-4 compared more favorably to Med-PaLM 2 in their clinical accuracy (83% vs. 72%), empathy (3.9 vs. 3.1) and trust (4.2 vs. 3.6), though both generative AI models showed great weaknesses in transparency where they were disclaiming in less than 20 percent of cases, and do not reference credible sources. It is also noticeable that both systems exhibited significant degree of bias in the same when demographical variations were introduced, especially in the case of race and immigration differences. The observations made shed light on the necessity of high ethical standards, human supervision and Model-level auditing until generative AI can be reliably used in the clinical practice. In conclusion, we suggest moderating bias, increasing transparency and co-designing communication systems between AI and patients to be created with the emphasis on safety, empathy and trust.
zobia shabeer (Thu,) studied this question.