5 Key Takeaways

• 1

  A study analyzed over 1.7 million outputs from large language models, revealing biases in clinical recommendations based on sociodemographic characteristics.

• 2

  The research evaluated nine large language models using 1,000 emergency department cases, comparing responses with a physician-derived baseline.

• 3

  Cases labeled as Black, unhoused, or LGBTQIA+ received more urgent care and invasive intervention recommendations than other groups.

• 4

  Mental health assessment recommendations varied significantly, with LGBTQIA+ cases receiving them six to seven times more than deemed appropriate by physicians.

• 5

  The study emphasized the need for bias evaluation strategies in clinical support tools to prevent unwarranted influences on medical decision-making.