AI models inherit human biases — recent healthcare studies show clear patterns. What have you noticed?

I start from a simple premise: Humans are not impartial. Not even close. We carry cultural, political, gender, class, and generational biases, and we bring them into everything we do. AI models are trained on data created by humans and then aligned with human preferences. So the idea that “AI is objective because it’s a machine” feels like a pretty dangerous myth to me. What do you think? Do you believe it’s actually possible for a model to be truly impartial? Or, like me, do you think bias is inevitable because it comes from us? And more importantly: Which models have you noticed this most clearly in, and in what kind of topics or situations? (ChatGPT, Claude, Gemini, Grok, Llama, DeepSeek… whatever) Please share concrete experiences. What did you ask, what did it reply, and why did it feel biased (or why were you surprised that it wasn’t)? A few recent studies that illustrate this (especially in healthcare): • Zack et al. (Lancet Digital Health, 2024): GPT-4 systematically stereotyped clinical vignettes and recommendations by race and gender. • Omar et al. (Nature Medicine, 2025): 9 models, 1.7+ million responses on 1,000 ED cases across 32 sociodemographic variations. Cases labeled Black, unhoused or LGBTQIA+ were steered toward urgent care, invasive procedures or mental-health evaluation far more often (sometimes 6–7×). • Same group’s 2026 pain study (Nature Health) and the EQUITRIAGE triage audit found similar patterns, including strong female undertriage in chest pain (ratios of 4.83:1 and 9.10:1 in some models).