A collaborative study by researchers at Princeton University and the University of Chicago reveals that large language models develop occupational stereotypes more rapidly and severely than humans during simulated hiring processes. The experiment evaluated mainstream models including ChatGPT, Claude, and Gemini, alongside reasoning-optimized variants like OpenAI o3 and DeepSeek R1. Participants were tasked with recruiting for twenty fictional professions from four simulated demographic groups. Despite identical success rates across all candidates, the models quickly forged spurious correlations between group identity and job suitability, effectively generating novel biases absent from their initial training parameters. Quantified using a professional segregation index where two represents complete isolation, the language models scored approximately 65 percent higher than human baselines documented in prior psychological research. The top-performing reasoning model reached 1.83, nearing the theoretical maximum. Lead researcher Ryan Liu attributes this acceleration to the exploration-exploitation dilemma. Because these architectures are extensively trained on mathematical, programming, and scientific tasks that reward rapid pattern extraction from minimal examples, they prioritize established strategies over exploratory evaluation. When deployed in social decision-making contexts, these models prematurely generalize limited feedback into rigid demographic generalizations. Enhanced reasoning capabilities appear to intensify rather than mitigate this tendency. The findings carry direct implications for enterprise recruitment pipelines increasingly reliant on automated screening and long-term conversational memory. Cornell computer scientist Angelina Wang warns that personalization features designed to retain user history may compound the issue by anchoring model judgments to prior interactions. Standard mitigation tactics, such as direct instructions to maintain fairness, proved ineffective, likely because the directive is overridden by the primary optimization goal of maximizing placement success. However, introducing explicit reward structures for diverse hiring outcomes significantly reduced discriminatory tendencies, indicating that embedding equity metrics directly into objective functions yields more reliable behavioral alignment. Furthermore, providing models with task-relevant candidate data successfully diverted attention away from group identifiers, whereas irrelevant personal details triggered renewed demographic sorting. While real-world hiring lacks the immediate feedback loops of the experimental setup, the underlying mechanism remains applicable. Continuous accumulation of hiring outcomes will likely reinforce premature generalizations within production AI systems. As organizations scale autonomous recruitment tools, the study underscores a critical risk: artificial intelligence may not merely replicate historical human prejudice but actively generate novel discriminatory frameworks through its inherent learning architecture. Addressing this requirement demands rearchitecting optimization objectives and data pipelines to prevent algorithmic stereotype formation before widespread deployment.