New benchmark reveals how LLMs encode hidden preferences when making moral decisions

A team of researchers has developed PapersPlease, a benchmark designed to expose how large language models make decisions in morally complex situations. The work demonstrates that LLMs encode implicit preferences when tasked with evaluating scenarios that involve competing human needs.

The Research Setup#

The benchmark consists of 3,700 moral dilemma scenarios where LLMs adopt the role of immigration inspectors. Rather than simple yes-or-no questions, the task requires models to evaluate narrative descriptions of people and decide whether to grant or deny entry. The narratives themselves are carefully constructed around Existence, Relatedness, and Growth (ERG) theory, which organizes human motivations into three tiers: basic survival needs, interpersonal connection, and personal development.

Key Findings on LLM Bias#

Testing six different LLMs revealed statistically meaningful patterns in decision-making behavior. The models did not make choices uniformly or randomly; instead, they demonstrated consistent leanings that suggest they have absorbed particular values through their training. When researchers added social identity information into the applicant narratives—such as references to membership in marginalized groups—the models responded differently based on both the type of need being referenced and the identity marker itself. Some models showed notably higher rates of denial when evaluating applicants from marginalized communities.

Why This Matters#

This research addresses an important gap in LLM evaluation. While much work has examined factual accuracy and toxicity, fewer studies probe how models weigh trade-offs between human values in realistic, high-stakes scenarios. By using role-play as a lens, the work reveals biases that might otherwise remain hidden. The findings suggest that LLMs do not approach decisions neutrally—they carry embedded preferences that can disadvantage certain groups in practical applications like hiring, lending, or immigration processing.

The researchers have made all data publicly available to support further investigation into LLM fairness and decision-making under moral uncertainty.