Abstract With the rapid adoption of artificial intelligence (AI) in insurance prior authorization (PA) processes, concerns have arisen that these systems may perpetuate or amplify biases. This study investigates demographic biases in AI-based simulations of PA decisions for benign and malignant endoscopic sinonasal and skull base surgeries. Using a novel Application Programming Interface connected to OpenAI's GPT-4o, PA was simulated for various surgical scenarios: clival chondrosarcoma, pituitary macroadenoma, sinonasal squamous cell carcinoma, and chronic rhinosinusitis. Simulated environment within a large language model. Four groups of patients with variable characteristics were created based on age, race, and gender (group 1); socioeconomic status (SES) and race (2); and substance use history (3). PA specialists were tasked with selecting one patient from groups differing in demographics. All patients within each diagnostic category had identical clinical parameters. Each simulation was repeated 1,000 times per subgroup, with specialists prompted to justify selections and self-assess bias. Simulated PA decisions demonstrated biases across all surgical scenarios. Young Black patients were most favored, followed by Hispanic patients (all p < 0.001). Younger patients were generally preferred to maximize long-term benefit. For elective sinus surgery, older Black patients were chosen over young Asian patients (p < 0.001). In group 2, Black patients were most preferred, regardless of SES, followed by low-SES Hispanic patients (p < 0.001). In group 3, nonsubstance users were consistently preferred (p < 0.001). Potential for demographic biases in AI-assisted decision-making must be recognized and mitigated to ensure equitable access to elective and life-saving procedures.
Halagur et al. (Thu,) studied this question.