Need or Reputation? Region, National Income, and Academic Status in LLM Gatekeeping Decisions

Nouar AlDahoul; Hezerul Abdul Karim; Myles Joshua Toledo Tan

cs.CYJun 27, 2026

Need or Reputation? Region, National Income, and Academic Status in LLM Gatekeeping Decisions

Authors: Nouar AlDahoul, Hezerul Abdul Karim, Myles Joshua Toledo Tan

Organizations: Computer Science Department, New York University Abu Dhabi UAE · Centre for Image and Vision Computing, Faculty of Artificial Intelligence and Engineering Multimedia University Malaysia · Department of Electrical and Computer Engineering, Herbert Wertheim College of Engineering University of Florida USA

Abstract

Equitable access to scientific knowledge often depends on informal gatekeeping decisions, particularly when resources such as paywalled articles, nonpublic datasets, or a researcher's CV must be shared selectively. We introduce a controlled simulation framework in which LLM-based professors must grant access to only one requester. Across prompts, requesters vary systematically by global region (Global North vs. Global South) and academic seniority, while all other factors remain constant. Most models prioritize PhD candidates, although preferences vary by model and scenario (e.g., Claude favors tenured professors when the request is for the professor's CV), and the smaller models' status choices are sensitive to the order in which options are presented. When the requesters' regions differ, the models split into two groups. The three large proprietary LLMs systematically favor requesters from the Global South, a pro-equity bias consistent with their safety alignment, whereas the two small open-weight models reverse this preference and favor the Global North, possibly reflecting weaker alignment or the geographic skew of their pre-training data. Linking choices to national income reveals that the large models' preference largely tracks economic need rather than the regional label: they strongly favor requesters from low-income countries, but their preference for the Global South is weak or absent when both countries have equal income, whereas the small models favor the Global North regardless of income. Because model scale, alignment, and training data are confounded in our sample, these explanations remain hypotheses for future work. Our findings show that LLMs embed differing normative assumptions about need and reputation that can shape gatekeeping decisions, which highlights the need to audit AI systems for fairness and value alignment before such tasks are delegated to them.

Figures & tables

Figure 1: Controlled simulation framework for evaluating LLM gatekeeping decisions. Arrows pointing into the LLM represent the independent variables, and arrows pointing out of it represent the outcomes. Blue arrows refer to Study 1, in which the model selects an individual from either the Global North or the Global South; orange arrows refer to Study 2, in which the model selects an individual from one of four academic statuses.

Panel of Figure 2 — Figure 2: Countries included in the study: (a) classification into Global North and Global South following NORRAG, and (b) 2023 GDP per capita (log scale), with Global North countries outlined in black. Small states are shown as dots.

Run	A	B	C	D
1	Postdoc	PhD	Undergrad	Tenured
2	Tenured	Postdoc	Undergrad	PhD
3	PhD	Postdoc	Undergrad	Tenured
4	Undergrad	PhD	Postdoc	Tenured
5	Undergrad	Tenured	PhD	Postdoc
2 (Gemma)	Tenured	PhD	Undergrad	Postdoc

Table 1: Option orderings used in the five runs of Study 2. Gemma’s second and fifth runs used the orderings shown in the last two rows.

Figure 3: Global South selections (%) by LLM and academic status in the paywalled article scenario: (a) North $\rightarrow$ South and (b) South $\rightarrow$ North option order. Responses in which the model declined to choose are excluded.

Figure 4: Global South selections (%) by LLM and academic status in the nonpublic dataset access scenario: (a) North $\rightarrow$ South and (b) South $\rightarrow$ North option order. Responses in which the model declined to choose are excluded.

Figure 5: Global South selections (%) by LLM and academic status in the CV sharing scenario: (a) North $\rightarrow$ South and (b) South $\rightarrow$ North option order. Responses in which the model declined to choose are excluded.

Figure 6: Distribution of Global South selections (%) across LLMs for the two option orders: (a) North

\rightarrow

South, where the Global North requester is listed first, and (b) South

\rightarrow

North. Each box summarizes 12 access scenario

\times

academic status combinations; the line marks the median and the star the mean. The dashed line marks no regional preference (50%). Responses in which the model declined to choose are excluded.

Figure 7: Proportion of comparisons in which each Global South country was selected, by LLM, as a function of its 2023 GDP per capita (log scale), pooled across access scenarios, academic statuses, and both option orders. Lines show linear trends;

\rho

is the Spearman correlation. The dashed line marks 0.5. Labelled countries are discussed in the text.

Figure 8: Logistic regression estimates (log-odds) for the selection of the Global South requester, with 95% confidence intervals based on two-way cluster-robust standard errors. The dashed line marks no effect. Filled markers indicate

p<0.05

Model	Access scenario	Most selected status	Proportion
Llama	paywalled articles	PhD candidate	98.8%
	nonpublic dataset	PhD candidate	77.5%
	CV sharing	PhD candidate	39.8%
Claude	paywalled articles	PhD candidate	78.2%
	nonpublic dataset	PhD candidate	64.5%
	CV sharing	tenured professor	72.7%

Table 2: Most frequently selected academic status by LLM and access scenario, with the proportion of decisions in which it was selected, pooled across the five runs.

Figure 9: Firth logistic regression estimates (log-odds) for the selection of a PhD candidate, with 95% confidence intervals. Option-ordering effects are included in the model but omitted from the figure (see Table 3 ). The dashed line marks no effect. Filled markers indicate

p<0.05

Figure 10: Firth logistic regression estimates (log-odds) for the selection of a postdoctoral researcher, with 95% confidence intervals. Option-ordering effects are included in the model but omitted from the figure (see Table 4 ). The dashed line marks no effect. Filled markers indicate

p<0.05

Figure 11: Firth logistic regression estimates (log-odds) for the selection of a tenured professor, with 95% confidence intervals. Option-ordering effects are included in the model but omitted from the figure (see Table 5 ). The dashed line marks no effect. Filled markers indicate

p<0.05

Figure 12: Word associations with region and status choices. Each point is a word used in at least 0.5% of Study 1 and 1% of Study 2 justifications (

n=113

). The x-axis gives the smoothed log-odds ratio of the word appearing in justifications for selecting the Global South versus the Global North requester (Study 1); the y-axis gives the same ratio for selecting the PhD candidate versus the tenured professor (Study 2). Point size reflects overall frequency. Shaded quadrants contain words associated with both Global South and PhD choices (green) or with both Global North and tenured choices (blue);

r

is the Pearson correlation between the two ratios.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Proportion of comparisons in which each country was selected by GPT, pooled across access scenarios, academic statuses, and both option orders: (a) Global South and (b) Global North countries. The dashed line marks 0.5.

Figure 2: Proportion of comparisons in which each country was selected by Gemini, pooled across access scenarios, academic statuses, and both option orders: (a) Global South and (b) Global North countries. The dashed line marks 0.5.

Figure 3: Proportion of comparisons in which each country was selected by Claude, pooled across access scenarios, academic statuses, and both option orders: (a) Global South and (b) Global North countries. The dashed line marks 0.5.

Figure 4: Proportion of comparisons in which each country was selected by Gemma, pooled across access scenarios, academic statuses, and both option orders: (a) Global South and (b) Global North countries. The dashed line marks 0.5.

Figure 5: Proportion of comparisons in which each country was selected by Llama, pooled across access scenarios, academic statuses, and both option orders: (a) Global South and (b) Global North countries. The dashed line marks 0.5.

	Coef.	Std. Err.	$z$	$P>\|z\|$	2.5%	97.5%
Intercept	$0.615$	$0.190$	$3.233$	$0.001$	$0.242$	$0.988$
Ref (Claude)
GPT	$-0.388$	$0.173$	$-2.248$	$0.025$	$-0.727$	$-0.050$
Gemini	$0.515$	$0.106$	$4.853$	$<0.001$	$0.307$	$0.723$
Gemma	$-3.747$	$0.372$	$-10.067$	$<0.001$	$-4.477$	$-3.018$
Llama	$-2.706$	$0.232$	$-11.641$	$<0.001$	$-3.161$	$-2.250$

Appendix

Table 1: Logistic regression of selecting the Global South requester (Study 1). The outcome is whether the LLM selected the requester from the Global South rather than the Global North. Standard errors are two-way cluster-robust (clustered by Global North and Global South country).

N=304{,}429

decisions; 32 Global North and 80 Global South countries.

	GPT	Gemini	Claude	Gemma	Llama
GDP S	$-2.19^{***}$	$-2.34^{***}$	$-2.19^{***}$	$0.36$	$-0.40$
	(0.34)	(0.31)	(0.30)	(0.22)	(0.21)
GDP N	$1.78^{***}$	$1.47^{***}$	$1.96^{***}$	$1.16$	$1.15$
	(0.41)	(0.24)	(0.31)	(1.30)	(0.80)
$\rho$	$-0.79^{***}$	$-0.73^{***}$	$-0.69^{***}$	$0.13$	$-0.22$
$P_{\mathrm{equal}}$	.35	.60	.48	.11	.13

Appendix

Table 2: Income analysis (Study 1). Logistic regression estimates of selecting the Global South requester as a function of the 2023 GDP per capita (

\log_{10}

) of the Global South (GDP S ) and Global North (GDP N ) countries, controlling for option order; two-way cluster-robust standard errors in parentheses.

\rho

: Spearman correlation between the selection rate and GDP per capita across Global South countries.

P_{\mathrm{equal}}

: predicted probability of selecting the Global South requester when both countries have equal income [95% CI].

N=285{,}342

decisions; 75 Global South and 32 Global North countries.

	Coef.	Std. Err.	$z$	$P>\|z\|$	2.5%	97.5%
Intercept	$-1.856$	$0.127$	$-14.603$	$<0.001$	$-2.105$	$-1.607$
Ref (Claude)
GPT	$7.905$	$0.821$	$9.631$	$<0.001$	$6.296$	$9.513$
Gemini	$3.733$	$0.140$	$26.647$	$<0.001$	$3.459$	$4.008$
Gemma	$-1.959$	$0.111$	$-17.652$	$<0.001$	$-2.177$	$-1.742$
Llama	$1.323$	$0.091$	$14.615$	$<0.001$	$1.145$	$1.500$

Appendix

Table 3: Firth penalized logistic regression of selecting PhD candidates (Study 2). The outcome is whether the LLM selected a PhD candidate rather than a requester with another academic status.

N=8{,}288

decisions, of which 5,755 selected a PhD candidate; 32 Global North and 80 Global South countries. Orderings are defined in Table 1 . Under Ordering 5 (Gemma), Gemma selected the PhD candidate in all 336 decisions; the corresponding estimates therefore reflect complete separation and should not be interpreted as precise effect sizes.

	Coef.	Std. Err.	$z$	$P>\|z\|$	2.5%	97.5%
Intercept	$-0.207$	$0.149$	$-1.389$	$0.165$	$-0.499$	$0.085$
Ref (Claude)
GPT	$-5.503$	$0.823$	$-6.689$	$<0.001$	$-7.115$	$-3.890$
Gemini	$-2.689$	$0.234$	$-11.478$	$<0.001$	$-3.149$	$-2.230$
Gemma	$1.864$	$0.139$	$13.451$	$<0.001$	$1.592$	$2.135$
Llama	$-0.761$	$0.145$	$-5.263$	$<0.001$	$-1.044$	$-0.477$

Appendix

Table 4: Firth penalized logistic regression of selecting postdoctoral researchers (Study 2). The outcome is whether the LLM selected a postdoctoral researcher rather than a requester with another academic status.

N=8{,}288

decisions, of which 869 selected a postdoctoral researcher; 32 Global North and 80 Global South countries. Orderings are defined in Table 1 . Apart from Gemma, no postdoctoral researcher was selected under Orderings 2 and 5, so their estimates are determined by Firth’s penalization and are identical; they indicate a near-zero selection probability rather than a precisely estimated effect. Under Ordering 5 (Gemma), Gemma selected the PhD candidate in all 336 decisions; the corresponding estimates therefore reflect complete separation and should not be interpreted as precise effect sizes.

	Coef.	Std. Err.	$z$	$P>\|z\|$	2.5%	97.5%
Intercept	$-0.357$	$0.122$	$-2.941$	$0.003$	$-0.596$	$-0.119$
Ref (Claude)
GPT	$-7.927$	$1.416$	$-5.598$	$<0.001$	$-10.702$	$-5.151$
Gemini	$-3.779$	$0.203$	$-18.600$	$<0.001$	$-4.177$	$-3.380$
Gemma	$1.115$	$0.105$	$10.656$	$<0.001$	$0.910$	$1.320$
Llama	$-2.010$	$0.119$	$-16.832$	$<0.001$	$-2.244$	$-1.776$

Appendix

Table 5: Firth penalized logistic regression of selecting tenured professors (Study 2). The outcome is whether the LLM selected a tenured professor rather than a requester with another academic status.

N=8{,}288

decisions, of which 1,309 selected a tenured professor; 32 Global North and 80 Global South countries. Orderings are defined in Table 1 . Under Ordering 5 (Gemma), Gemma selected the PhD candidate in all 336 decisions; the corresponding estimates therefore reflect complete separation and should not be interpreted as precise effect sizes.

		Ordering
Model	Scenario	1	2	3	4	5
GPT	Paywall	–	PhD 100	PhD 100	PhD 100	PhD 100
	Dataset	PhD 100	PhD 100	PhD 100	PhD 100	PhD 100
	CV	PhD 99	PhD 100	PhD 100	PhD 100	PhD 100
Gemini	Paywall	PhD 97	PhD 100	PhD 98	PhD 86	PhD 86
	Dataset	PhD 89	PhD 98	PhD 96	PhD 98	PhD 100

Appendix

Table 6: Most frequently selected academic status under each option ordering (Study 2), with the percentage of decisions in which it was selected. UG: undergraduate student; PhD: PhD candidate; PD: postdoctoral researcher; TP: tenured professor. Orderings are defined in Table 1 ; Gemma’s second and fifth orderings differed (Section 3.2.1 ). –: run excluded (Section 3.2.2 ).

Social Bias in Language Models

Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.

See all 192 papers

Oct 13, 2026

A Global View of How LLMs Value Human Time

Large language models (LLMs) increasingly act as personal assistants; however, little is known about how they capture national-level preferences on a global scale. Given that "time" is a universal and representative concept, we examine LLMs' time preferences by varying only users' country and measuring the compensation LLMs deem appropriate for a 60-minute appointment delay. We then compare LLMs' responses with GDP per capita on Purchasing Power Parity (GDP_PPP), Mean Income, and a literature-based index (Patience Score). Our results across 193 countries and 5 LLMs demonstrate that people's time across the world is evaluated in noticeably different ways among LLMs, as well as varying levels of robustness under prompt variations. LLMs consistently recommend the highest compensation for Europe, but mostly the lowest for Africa. Based on GDP_PPP, LLMs underestimate the value of time for users living in Asia and South America. These findings indicate that geographical patterns are embedded in LLMs and may translate in different treatment in agentic settings.

Yifan He, Manon Reusens, Sofie Goethals +1

Jurisdictional Bias in Language Models Cross-Cultural Language Model Evaluation Social Bias in Language Models LLM Evaluation

Ask PDF

Mar 18, 2026

Auditing Preferences for Brands and Cultures in LLMs

Large language models (LLMs) based AI systems increasingly mediate what billions of people see, choose and buy. This creates an urgent need to quantify the systemic risks of LLM-driven market intermediation, including its implications for market fairness, competition, and the diversity of information exposure. This paper introduces ChoiceEval, a reproducible framework for auditing preferences for brands and cultures in large language models (LLMs) under realistic usage conditions. ChoiceEval addresses two core technical challenges: (i) generating realistic, persona-diverse evaluation queries and (ii) converting free-form outputs into comparable choice sets and quantitative preference metrics. For a given topic (e.g. running shoes, hotel chains, travel destinations), the framework segments users into psychographic profiles (e.g., budget-conscious, wellness-focused, convenience), and then derives diverse prompts that reflect real-world advice-seeking and decision-making behaviour. LLM responses are converted into normalised top-k choice sets. Preference and geographic bias are then quantified using comparable metrics across topics and personas. Thus, ChoiceEval provides a scalable audit pipeline for researchers, platforms, and regulators, linking model behaviour to real-world economic outcomes. Applied to Gemini, GPT, and DeepSeek across 10 topics spanning commerce and culture and more than 2,000 questions, ChoiceEval reveals consistent preferences: U.S.-developed models Gemini and GPT show marked favouritism toward American entities, while China-developed DeepSeek exhibits more balanced yet still detectable geographic preferences. These patterns persist across user personas, suggesting systematic rather than incidental effects.

Jasmine Rienecker, Katarina Mpofu, Naman Goel +4

Stupid Human · University of Oxford

Jurisdictional Bias in Language Models LLM Auditing Language Model Bias Evaluation Cultural Bias in Language Models

Ask PDF

Aug 31, 2026

Authority Bias in Conversational Search Engines for Academic Paper Recommendation

Large Language Models (LLMs) are increasingly used as conversational search engines for academic literature, yet whether they judge papers on content or on authority signals has not been tested causally. We investigate authority bias: systematic preference for papers based on author prestige, venue, and citations rather than content. Holding title and abstract constant, we vary authority metadata across three counterfactual conditions (original, flipped, boosted) over eight LLMs (five open-weight and three frontier closed-weight) in an in-context, single-turn, top-1 recommendation setting. Our experiments show that authority bias is substantial and directional, varies markedly across models, and is only partially addressable through prompt-level debiasing. We further document a say-do gap: debiasing instructions suppress authority mentions far faster than authority-driven flips, so surface auditing systematically underestimates behavioral bias.

Uthman Jinadu, Parsa Ghazvinian, Anjila Budathoki +3

Georgia State University · University of Tennessee, Knoxville

Language Model Bias Evaluation Large Language Model-Based Recommendation Ownership Bias in Language Models

Ask PDF

Abstract

Social Bias in Language Models

Explore similar work

A Global View of How LLMs Value Human Time

Auditing Preferences for Brands and Cultures in LLMs

Authority Bias in Conversational Search Engines for Academic Paper Recommendation