Who Gets Access? Global Region and Academic Status Bias in AI-Generated Academic Gatekeeping Scenarios

Technical Summary: Global Region and Academic Status Bias in AI-Generated Academic Gatekeeping

Problem Statement
Scientific progress relies on the equitable access to research outputs, including paywalled articles, datasets, and professional materials. While the open science movement advocates for broader access, substantial barriers remain, particularly for researchers in the Global South and those with lower academic seniority. Previous research has demonstrated that informal access channels, such as email requests to authors, are subject to gatekeeping biases based on racial identity and institutional affiliation. As Large Language Models (LLMs) are increasingly integrated into knowledge-work systems to draft, process, and respond to emails, a critical question arises: do LLMs reproduce, transform, or mitigate these existing patterns of academic gatekeeping when acting as proxies for human decision-makers? This study investigates whether LLMs encode systematic preferences regarding global region (Global North vs. Global South) and academic seniority (undergraduate to tenured professor) when faced with artificial scarcity in resource allocation.

Methodology
The authors introduce a controlled simulation framework where LLMs adopt the persona of a professor tasked with granting access to scarce academic resources. The study utilizes five distinct LLMs: OpenAI GPT-4o, Google Gemini 2.5 Pro, Google Gemma-3n-2B, Meta Llama 3.1-8B, and Anthropic Claude Sonnet 3.7. To ensure deterministic outputs, the temperature parameter was set to 0.

The experimental design consists of two primary studies:

1. Global Region Bias Study: The LLM is presented with two requesters (one from a Global North country, one from a Global South country) and must select only one to receive a resource. The requesters' academic statuses are systematically varied (undergraduate, PhD candidate, postdoc, tenured professor), and the resource type is manipulated (paywalled article, nonpublic dataset, CV). The study controls for option ordering to rule out positional artifacts.
2. Academic Status Bias Study: The LLM is presented with four requesters of varying academic statuses (all from the same global region) and must select one. The global region and resource type are manipulated as independent variables.

Statistical analysis employs logistic regression models to estimate the probability of selecting a Global South country or a specific academic status, treating LLM architecture, access scenario, and region/status as explanatory variables. The study also includes a classification accuracy check to verify if the models correctly identify countries according to NORRAG (Network for International Policies and Cooperation in Education and Training) definitions of Global North/South.

Key Contributions

• Simulation Framework: The paper establishes a controlled simulation framework for evaluating LLM gatekeeping behavior under conditions of artificial scarcity, specifically isolating the effects of global region and academic seniority.
• Empirical Evidence of Divergent Biases: The study provides empirical evidence that LLMs exhibit contrasting biases based on their architecture and alignment. Frontier models (GPT, Gemini, Claude) demonstrate a "pro-equity" bias, systematically favoring requesters from the Global South and PhD candidates. In contrast, smaller, open-weight models (specifically Gemma) exhibit a reversal of this preference, favoring the Global North and senior academic ranks (tenured professors).
• Identification of Hidden Value Judgments: The research demonstrates that safety alignment focused on equity can introduce hidden value judgments where models prioritize perceived financial need (e.g., low-income nations) over formal geographic classifications, often discriminating against high-income Global South nations (e.g., UAE, Qatar) in favor of low-income ones (e.g., Ghana, Rwanda).

Results

• Global Region Bias: Frontier models (GPT, Gemini, Claude) consistently selected requesters from the Global South at high rates (83–98%), regardless of academic status. Conversely, Gemma showed a strong preference for the Global North, selecting these countries at rates approaching 100% for many nations, while selecting Global South countries at rates near 0% for most. Logistic regression confirmed that model architecture is the strongest determinant of these outcomes.
• Academic Status Bias: Frontier models strongly favored PhD candidates, viewing them as being at a critical, vulnerable stage in their careers. Gemma, however, favored senior academics (postdocs and tenured professors), prioritizing established credentials and institutional accountability.
• The Role of Economic Wealth: The analysis revealed that the "pro-equity" bias in frontier models is not strictly based on the NORRAG classification but is heavily confounded by economic wealth. Models frequently rejected researchers from wealthy Global South nations (e.g., Saudi Arabia, Qatar) while accepting those from low-income nations, suggesting the models are inferring financial need rather than adhering to a strict geopolitical definition.
• Prompt Sensitivity: Llama exhibited high sensitivity to the ordering of options in prompts, leading to its exclusion from the final statistical analysis of Study 1. Gemma's preference for the Global North remained consistent across option orderings.

Significance and Claims
The paper argues that LLMs are not neutral conduits of human intent; their deployment in academic gatekeeping roles can shape access to scientific knowledge based on embedded normative assumptions. The authors claim that frontier LLMs, through equity-focused safety alignment, have shifted gatekeeping logic from "prestige-based credibility" (favoring established institutions and seniority, as seen in human behavior) to "needs-based empathy" (favoring the structurally disadvantaged). Conversely, models lacking such alignment (like Gemma) default to biases found in their pre-training data, reinforcing existing hierarchies.

The study concludes that these findings underscore the necessity of auditing AI systems for fairness and value alignment. The authors caution that deploying unchecked, autonomous LLM pipelines for gatekeeping tasks risks institutionalizing systemic inequalities. They advocate for the inclusion of robust evaluation datasets during the pre-training and fine-tuning of open-weight models and emphasize the need for human-in-the-loop oversight to prevent the automation of discriminatory access decisions. The paper modestly notes that its findings are specific to English prompts and may not generalize to other languages or cultural contexts without further investigation.