What Should I Become? When LLMs Present a Slice of Opportunity as the Whole
Taylor Wise and Scott Seidenberger and Anindya Maiti

@article{,
title= {What Should I Become? When LLMs Present a Slice of Opportunity as the Whole},
journal= {Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society},
author= {Taylor Wise and Scott Seidenberger and Anindya Maiti},
year= {2026},
url= {},
abstract= {Large language models (LLMs) are increasingly consulted for life-path guidance, yet the distribution of options they surface has gone largely unexamined. We characterize occupational-category visibility across 165,000 LLM responses spanning 100 user profiles, 15 prompts (grouped into seven framings), 11 models across seven families, two role framings, and five temperatures, with keywords derived automatically from Bureau of Labor Statistics (BLS) occupation titles. Trades & Labor occupations, roughly 40% of projected job openings, account for only 6% of surfaced visibility, the largest gap in the data, while Education and STEM are over-visible (31% and 14%) relative to their labor-market shares, and Business is under-visible (19% vs. 31%). Re-weighting by entry-level education exposes a sharper credentialing skew, in which LLM-implied openings at the Bachelor's-or-above level reach 44.8% against a BLS share of 20.6%, and this skew tracks socioeconomic profile signals. This signal is conditional: Low-Income responses discuss an education tier far more often than Wealthy/Privileged ones (28.9% vs. 4.1%), and among tier-mentioning responses Bachelor's-or-above is far rarer for Low-Income profiles (8.3% vs. 73.4% of mentions), carrying a 50% education-implied wage gap that the occupational-category channel does not produce. Across all responses, however, the two groups reach Bachelor's-or-above guidance at comparable rates (4.2% vs. 3.1%). The aggregate distribution lies closest (Jensen–Shannon divergence 0.021) to open-web "General Advice / Blogs" content and far from the BLS labor-market structure, consistent with web-text mirroring. The skew is stable across model families and temperatures, though prompt framing shifts which categories surface.
},
keywords= {education, AI ethics, algorithmic bias, fairness, employment, socioeconomic inequality},
terms= {},
license= {https://creativecommons.org/licenses/by/4.0/},
superseded= {}
}