Large language models (LLMs) are increasingly integrated into our daily lives, but do they carry our biases with them? A new study reveals how LLMs perceive gender roles in different occupations, comparing their responses to human perceptions and real-world statistics. Researchers created unique test scenarios to avoid the problem of LLMs simply regurgitating information they’ve already seen in their training data. They then asked several LLMs to guess the gender associated with various jobs and compared these answers to how humans perceive those same roles and to actual workforce demographics. The results are fascinating and a little unsettling. The study found that LLMs often deviate significantly from a gender-neutral baseline. While the newer, larger models like GPT-4o showed better alignment with human perceptions, all LLMs tested leaned more toward reflecting statistical workforce data, even when those statistics revealed existing gender imbalances. This raises important questions about how AI systems might perpetuate biases, even if unintentionally. Are we training AI to mirror our world, warts and all? Or should we strive for AI that challenges our biases and helps us build a more equitable future? Further research is needed to understand how these biases evolve as LLMs develop and to explore strategies for mitigating these potential harms. This study provides a valuable framework for future research, emphasizing the importance of careful evaluation and ethical considerations in AI development.