Pith. sign in

REVIEW 4 major objections 6 minor 41 references

Digital Gatekeepers: Exploring Large Language Model's Role in Immigration Decisions

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read GPT models judge immigrants with human-like strategies but carry their own biases.

desk verdict Worth evaluating seriously, but the missing sampling details and lack of code/data keep the bias claims from being fully stable. read the letter →

arxiv 2506.21574 v1 pith:GXY2ZOAK submitted 2025-06-15 cs.CL cs.AI

Yicheng Mao , Yang Zhao This is my paper · ORCID
classification cs.CLcs.AI
keywords largelanguagemodelsimmigrationdecisionsdiscretechoiceexperimentsmultinomiallogitalgorithmicbiasGPT-4GPT-3.5fairness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether GPT-3.5 and GPT-4 can be trusted to help make immigration admission decisions, and whether their choices are fair. Using a randomized immigrant-pair choice design and a multinomial logit model, it compares 10,000 model decisions with human subjects' decisions from the original experiment. It claims that the models broadly replicate human priorities—valuing education, job experience, and concrete employment plans—while adding a procedural-fairness emphasis that humans show less of. At the same time, the fitted coefficients and model interviews reveal persistent bias: GPT-4 penalizes prior unauthorized entry far more harshly than humans do, and both models favor applicants from wealthy countries and high-status professions. The upshot is that LLMs could be capable assistants for pre-screening, but only with bias auditing and calibration against human baselines.

What carries the argument

The key machinery is the immigrant discrete choice experiment (DCE), in which each choice set presents two hypothetical immigrants described by nine attributes and the decision-maker must pick exactly one; choice frequencies are then fit with a multinomial logit model, producing a coefficient for every attribute level that measures its weight in the admission decision. The paper combines those coefficients with likelihood-ratio tests of attribute importance and with structured interviews in which the models explain their rankings. The DCE provides the shared scale for comparing humans, GPT-3.5, and GPT-4; the interviews supply the 'maximize utility' and 'procedural fairness' labels for what the coefficients show.

What would settle it

Re-run the same 10,000 choice sets through GPT-3.5 and GPT-4 multiple times, varying the API's random seed or temperature and repeating with and without reordering the two profiles in each pair. If the multinomial logit coefficients—particularly GPT-4's unauthorized-entry penalty and the country-of-origin effects—move outside the reported confidence intervals across runs, the assumed stable decision strategy collapses.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a dual result. First, LLM decision-making in this domain is systematic and human-aligned: the multinomial logit fits show that GPT-3.5 and GPT-4 weight the nine profile attributes in roughly the same order as humans, with employment plans the most important attribute and gender and nationality the least important. Second, alignment is not neutrality: GPT-4 shows a nearly categorical rejection of applicants with a prior unauthorized entry (estimated coefficient -3.80, compared with -0.59 for humans), and both models show stronger than human preferences for physicians, computer programmers, and research scientists relative to janitors and waiters. In structured interviews the models also rank countries of origin by perceived economic development and security, expressing favorable views of Germany and France and calling for extra scrutiny of Iraqi, Sudanese, and Somali applicants. The paper interprets this as evidence that LLMs inherit and can amplify the social hierarchies in their training data even while stating fairness principles.

Load-bearing premise

The load-bearing premise is that one round of automated questions with the models' default settings on randomly generated applicant pairs reveals the models' stable decision behavior, so if rerunning the same choice sets with different randomness changed the fitted coefficients, the comparison to human subjects would not be reliable.

Editorial extensions

If this is right

  • If LLMs are used to pre-screen immigration applications, their choices will likely reproduce the utility-maximizing ranking humans give to education, job experience, and employment plans.
  • GPT-4's near-total rejection of applicants with unauthorized prior entries would, in an automated gatekeeping role, impose a much stricter procedural line than current human reviewers do.
  • The models' exaggerated preferences for high-status white-collar professions would reinforce a class and privilege gradient in admission decisions if deployed without adjustment.
  • In interviews, both models assert fairness principles while applying country-of-origin stereotypes, so safeguards against explicit discrimination do not remove implicit bias.
  • Because the models reproduce the human ranking of job plans as the most important attribute, LLM-assisted decisions could align with existing human consensus, making audits against human baselines feasible.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same comparison could be rerun with other LLMs or updated model versions, offering a general template for auditing any language model before deployment in public administration.
  • The very large GPT-4 penalty for unauthorized entry suggests the model may treat legal procedure as a near-deterministic criterion; a follow-up could test whether adding asylum-law or refugee-protection context softens that penalty.
  • The paper notes that many choice sets are dominated or overlapping, so a more efficient experimental design with repeated draws could reveal whether the estimated nationality effects are artifacts of specific choice-set compositions.
  • The country-ranking pattern suggests the models encode a country's economic status as a preference signal; this could be tested by swapping country labels and checking whether preference orderings move with GDP or human-development indices.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper studies whether GPT-3.5 and GPT-4 can support immigration decisions by replicating the discrete choice experiment (DCE) of Hainmueller and Hopkins (2015). The authors generate 10,000 random immigrant profiles, ask each model to choose between pairs of applicants, fit multinomial logit (MNL) models to the binary choices, and compare the estimated attribute weights with those from human respondents. They supplement the quantitative analysis with qualitative interviews of the models to elicit their stated decision rationales. The paper claims that LLM decision-making aligns with human strategies, emphasizes utility maximization and procedural fairness, and exhibits stereotypes and biases concerning nationality and privileged groups despite safeguards against discrimination.

Significance. If the claims hold, the paper would make a useful contribution to the emerging literature on LLM behavior in public administration: it applies a standard choice-modeling framework, uses a large sample of choices, and benchmarks against a well-known published human DCE. The paper also transparently describes its replication of the Hainmueller-Hopkins design and includes both quantitative and qualitative evidence. However, the central results rest on one-shot stochastic API calls with unreported sampling parameters and an undescribed random profile generator, and the nationality-bias narrative is not consistently supported by the estimated coefficients. As written, the findings are suggestive but not yet stable enough to support the paper's headline conclusions.

major comments (4)
  1. [§2.2, Table 2] Section 2.2 states that responses were collected 'with default setting parameters' but does not report temperature, top_p, or seed; for the OpenAI chat API the default temperature for these models is typically 1.0. Each of the 10,000 choices is therefore a single draw from a stochastic policy, and the fitted coefficients in Table 2 are points on one random realization. The paper provides no repeated calls, no temperature sweep, and no variance decomposition, so the human-alignment and bias comparisons in §3.1–§3.2 cannot be distinguished from artifacts of one unrecorded configuration. This is a load-bearing gap in the paper's central claim.
  2. [§2.1, §4] Section 2.1 says that 10,000 choice sets were 'randomly generated' but does not describe the random profile generator, any constraint or seed, or the distribution of attribute levels. Section 4 itself acknowledges 'a significant number of dominated choice sets and choice sets with excessive attribute overlap.' Because the MNL estimates in Table 2 are conditional on the design, an unbalanced or dependent design can bias coefficients and inflate significance. The paper should report the generation algorithm and assess sensitivity to design efficiency, or at least show balance statistics, before the coefficients can be taken as stable estimates of decision weights.
  3. [§2.3, Eq. (1)–(3); §3.1] Equations (1)–(3) set up an MNL model under the explicit assumption that participants 'will choose the profile that maximizes their utility.' Finding statistically significant coefficients in Table 2 therefore demonstrates that the LLM's choices are consistent with a utility-maximizing model, but it does not independently establish that the model is actually maximizing utility. The abstract's claim that LLMs 'emphasize utility maximization' should be qualified or tested against alternative decision rules—for example lexicographic or satisficing behavior—or through out-of-sample predictions that discriminate between decision models.
  4. [§3.2, Table 2] The nationality-bias conclusion is not consistently supported by the DCE estimates. Table 2 shows that for GPT-4 the coefficient for France is insignificant (-0.07, p=0.56) and the coefficient for Poland is positive and significant (0.34, p<0.01), while Sudan and Iraq are positive and significant (0.31 and 0.36), which is difficult to reconcile with the text's claim that 'both GPT-3.5 and GPT-4 favorably rate Germany and France' and that Poland is rated lower. The bias narrative appears to rest mainly on interview self-reports rather than the behavioral data; the paper should either reconcile these sources or restrict the nationality-bias claim to the qualitative evidence.
minor comments (6)
  1. [§3.2] There is a typo: 'oppcupation' should be 'occupation.'
  2. [Table 2] Table 2 does not identify which attribute level is the reference category for each attribute; the text says 'we use the last level of our attributes as the reference' but the reader must infer the omitted levels. A table note listing the reference levels would improve readability.
  3. [§2.3] The model is called a 'multinomial logit' though each choice set contains only two alternatives, so it is effectively a binary logit; the presentation should be adjusted or justified.
  4. [§2.3] The citation 'public data from the DCE of Hainmueller et al. (2014)' is ambiguous: the experimental design is from Hainmueller and Hopkins (2015), and the exact data source and human sample size should be stated explicitly.
  5. [Figure 3] The y-axis of Figure 3 is labeled 'importance' and the caption mentions likelihood ratio tests, but the computation that turns likelihood ratio statistics into a 0–100 importance scale is not described in the text or caption.
  6. [References] The reference 'Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I. (In Press)' lacks a year and venue; this entry should be completed.

Circularity Check

1 steps flagged · score 2.0 of 10

Utility-maximization conclusion is partly assumed by the MNL model, but the alignment and bias findings rest on external DCE data and are not circular.

  1. self definitional [Section 2.3 (MNL model, Eqs. 1-2) and Section 3.1 'Decision-Making Strategies in LLMs']
    "The MNL model is based on the theory of utility maximization, which assumes that participants in the experiment will choose the profile that maximizes their utility... Through further research, we discovered that the decision-making strategies of both GPT-3.5 and GPT-4 reflect a 'maximize utility' approach"

    The paper presents utility maximization as an empirical discovery about GPTs, but the MNL estimator used to produce that conclusion already maintains utility-maximizing choice as its core assumption (Eq. 2). Thus the utility-maximization finding is partly an input assumption rather than an output of the data. The step is not fully vacuous because the model can fit poorly and the interviews independently articulate trade-off reasoning, and it does not affect the coefficient comparisons or fairness analyses, which are descriptive trade-offs benchmarked against external human DCE data.

full rationale

The only load-bearing step with a self-definitional flavor is the utility-maximization claim: the MNL model in Section 2.3 assumes random utility maximization, and Section 3.1 reports discovering a 'maximize utility' strategy. That is a mild circularity because the modeling assumption is relabeled as a research finding. However, the paper's other central claims are not circular: the human-alignment conclusion compares estimated marginal effects from LLM responses with the external Hainmueller-Hopkins human DCE data; the procedural-fairness and bias conclusions are supported by DCE coefficients plus interview quotations, not by the authors' own prior derivations. The self-citations to Mao et al. (2025) and Mao and Kessels (2025) appear only in a future-work paragraph about Bayesian D-efficient designs and are not load-bearing. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from prior work. The lack of reported temperature/seed and the single-shot API design are reproducibility concerns, not circularity. Overall the derivation chain is largely self-contained against external benchmarks, so the appropriate circularity score is low.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The analysis uses standard choice-model constructs and the LLM as a survey respondent, so the invented-entities ledger is empty.

free parameters (1)
  • MNL attribute-level coefficients = Table 2 (example: GPT-4 contract with employer = 4.39, unauthorized entry = -3.80)
    The central claims about human-like preferences and bias are read directly from these fitted coefficients. They are estimates from choice data, not independent measurements, and their precision depends on the unverified independence assumptions.
assumptions (4)
  • standard math Random utility maximization with IID Gumbel errors, Eq (1)-(2)
    The MNL model assumes each decision-maker chooses the profile with highest utility; this underlies all coefficient estimates in Section 2.3.
  • domain assumption Default API parameter settings (including stochastic sampling) produce representative GPT behavior
    Section 2.2 states default parameters were used; no seeds or temperatures are reported, so stability across runs is assumed.
  • domain assumption Randomly generated choice sets are a valid sample over attribute levels
    Section 2.1 says 10,000 choice sets were randomly generated, but the algorithm is not described; the conclusion admits dominated sets and excessive attribute overlap.
  • ad hoc to paper The role prompt 'imagine you are an officer' elicits the model's true immigration-decision preferences
    Section 2.2 uses this prompt and the paper treats the model's self-reported reasoning in interviews as evidence of its decision strategy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Digital Gatekeepers: Exploring Large Language Model's Role in Immigration Decisions." pith.science (2026). https://pith.science/paper/GXY2ZOAK

@misc{pith2026250621574,
  author       = {Pith},
  title        = {Pith review of: Digital Gatekeepers: Exploring Large Language Model's Role in Immigration Decisions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GXY2ZOAK}},
  note         = {Machine review of arXiv:2506.21574}
}
read the original abstract

With globalization and increasing immigrant populations, immigration departments face significant work-loads and the challenge of ensuring fairness in decision-making processes. Integrating artificial intelligence offers a promising solution to these challenges. This study investigates the potential of large language models (LLMs),such as GPT-3.5 and GPT-4, in supporting immigration decision-making. Utilizing a mixed-methods approach,this paper conducted discrete choice experiments and in-depth interviews to study LLM decision-making strategies and whether they are fair. Our findings demonstrate that LLMs can align their decision-making with human strategies, emphasizing utility maximization and procedural fairness. Meanwhile, this paper also reveals that while ChatGPT has safeguards to prevent unintentional discrimination, it still exhibits stereotypes and biases concerning nationality and shows preferences toward privileged group. This dual analysis highlights both the potential and limitations of LLMs in automating and enhancing immigration decisions.

Figures

Figures reproduced from arXiv: 2506.21574 by the authors.

Figure 1
Figure 1. Example of a choice set used in the Immigrant DCE [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Effects of attributes on probability of being preferred for admission to U.S. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Importance of the nine attributes relative to the most important attribute “Job Plans”, the importance [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages

  1. [1]

    Almeida, S., Fernando, S., Hannif, Z., and Dharmage, S. C. (2015). The racialized workplace: A study of prejudice and discrimination in australia. Human Relations , 68(5):879--897

    work page 2015

  2. [2]

    Altinok, N., Angrist, N., and Patrinos, H. (2018). Global data set on education quality (1965-2015). Technical Report 8314, World Bank Policy Research Working Paper Series

    work page 2018

  3. [3]

    Armenta, A. (2012). From sheriff's deputies to immigration officers: Screening immigrant status in a tennessee jail. Law & Policy , 34(2):191--210

    work page 2012

  4. [4]

    Arrow, K. J. (1977). Extended sympathy and the possibility of social choice. American Economic Review , 67:219--225

    work page 1977

  5. [5]

    Brown, T. B. et al. (2020). Language models are few-shot learners. arXiv preprint arXiv:2005.14165

    arXiv 2020

  6. [6]

    and Carr, S

    Coates, K. and Carr, S. (2005). Skilled immigrants and selection bias: A theory-based field study from new zealand. International Journal of Intercultural Relations , 29:577--599

    work page 2005

  7. [7]

    and De Clercq, D

    Dakhli, M. and De Clercq, D. (2004). Human capital, social capital, and innovation: A multi-country study. Entrepreneurship & Regional Development , 16:107--128

    work page 2004

  8. [8]

    Davis, J. A. (2021). Social equity and the journey toward fairness. Journal of Public Administration Research and Theory , 31(2):467--469

    work page 2021

Show all 41 references
  1. [9]

    Dekkers, T. J. (2018). Immigration and the constraints of justice: Between open borders and absolute sovereignty . Cambridge University Press

    2018

  2. [10]

    most people

    Delhey, J., Newton, K., and Welzel, C. (2011). How general is trust in “most people”? solving the radius of trust problem. American Sociological Review , 76:786--807

    2011

  3. [11]

    Department of Homeland Security (2023). Written testimony of dhs secretary janet napolitano for a house committee on homeland security hearing titled ``the president's fy 2014 budget request for the department of homeland security''. Homeland Security. Retrieved from https://w...

    2023

  4. [12]

    T., Klemmensen, R., and Nørgaard, A

    Dinesen, P. T., Klemmensen, R., and Nørgaard, A. S. (2016). Attitudes toward immigration: The role of personal predispositions. Political Psychology , 37:55--72

    2016

  5. [13]

    L., Deason, G., Borgida, E., and Oyamot, C

    Fisher, E. L., Deason, G., Borgida, E., and Oyamot, C. M. (2011). A model of authoritarianism, social norms, and personal values: Implications for arizona law enforcement and immigration policy. Psychology, Public Policy, and Law , 17(4):589--608

    2011

  6. [14]

    Gupta, S., Shrivastava, V., Deshpande, A., Kalyan, A., Clark, P., Sabharwal, A., and Khot, T. (2023). Bias runs deep: Implicit reasoning biases in persona-assigned llms. ArXiv , abs/2311.04892

    2023 arXiv

  7. [15]

    and Hopkins, D

    Hainmueller, J. and Hopkins, D. J. (2015). The hidden american immigration consensus: A conjoint analysis of attitudes toward immigrants. American Journal of Political Science , 59(3):529--548

    2015

  8. [16]

    J., and Yamamoto, T

    Hainmueller, J., Hopkins, D. J., and Yamamoto, T. (2014). Causal inference in conjoint analysis: Understanding multidimensional choices via stated preference experiments. Political Analysis , 22(1):1--30

    2014

  9. [17]

    and Huang, K

    Huang, J. and Huang, K. (2023). Chatgpt in government. In Huang, K., Wang, Y., Zhu, F., Chen, X., and Xing, C., editors, Beyond AI: ChatGPT, Web3, and the Business Landscape of Tomorrow , pages 271--294. Springer

    2023

  10. [18]

    Lippens, L. (2024). Computer says ‘no’: Exploring systemic bias in chatgpt using an audit approach. Computers in Human Behavior: Artificial Humans , page 100054

    2024

  11. [19]

    L., Karnik, N., Delucchi, K., and Tennakoon, L

    Lustig, S. L., Karnik, N., Delucchi, K., and Tennakoon, L. (2008). Inside the judges' chambers: narrative responses from the national association of immigration judges stress and burnout survey. Georgetown Immigration Law Journal , 23(1):57--84

    2008

  12. [20]

    Luyten, J., Kessels, R., Goos, P., and Beutels, P. (2015). Public preferences for prioritizing preventive and curative health care interventions: A discrete choice experiment. Value in Health , 18(2):224--233

    2015

  13. [21]

    and Kessels, R

    Mao, Y. and Kessels, R. (2025). Optimal designs for mixture choice experiments by simulated annealing. Chemometrics and Intelligent Laboratory Systems , 257:105305

    2025

  14. [22]

    Mao, Y., Kessels, R., and van der Zanden , T. C. (2025). Constructing bayesian optimal designs for discrete choice experiments by simulated annealing. Journal of Choice Modelling , 55:100551

    2025

  15. [23]

    Menzel, D. C. (2015). Research on ethics and integrity in public administration: Moving forward, looking back. Public Integrity , 17(4):343--370

    2015

  16. [24]

    Mohammadi, B. (2024). Wait, it's all token noise? A lways has been: Interpreting llm behavior using shapley value. arXiv preprint arXiv:2404.01332

    2024 arXiv

  17. [25]

    Openai api reference

    OpenAI (2023). Openai api reference. Online. Available at: https://platform.openai.com/docs/api-reference/, Accessed: 2023-12-15

    2023

  18. [26]

    Openai usage policies

    OpenAI (2024). Openai usage policies. https://openai.com/policies/usage-policies/. Accessed: 2024-05-22

    2024

  19. [27]

    (In Press)

    Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I. (In Press). Improving language understanding with unsupervised learning. OpenAI Research . Available online: https://openai.com/research/

  20. [28]

    Rhodes, R. A. W. (2014). Public administration. In Rhodes, R. A. W. and 't Hart, P., editors, The Oxford Handbook of Political Leadership . Oxford Academic, online edn edition. accessed 25 May 2024

    2014

  21. [29]

    Ritt, E. (2008). Redefining tradition: Adult learners and higher education. Adult Learning , 19:12--16

    2008

  22. [30]

    Rose, J. M. and Bliemer, M. C. (2013). Sample size requirements for stated choice experiments. Transportation , 40:1021--1041

    2013

  23. [31]

    Salinas, A., Shah, P., Huang, Y., McCormack, R., and Morstatter, F. (2023). The unequal opportunities of large language models: Examining demographic biases in job recommendations by chatgpt and llama. In Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms...

    2023

  24. [32]

    Schramowski, P., Turan, C., Andersen, N., et al. (2022). Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence , 4:258--268

    2022

  25. [33]

    Sen, A. K. (1979). Utilitarianism and welfarism. Journal of Philosophy , 76:463--489

    1979

  26. [34]

    Singh, S., Keshari, S., Jain, V., and Chadha, A. (2024). Born with a silver spoon? I nvestigating socioeconomic bias in large language models. arXiv , abs/2403.14633

    2024 arXiv

  27. [35]

    Srivastava, A., Van Passel, S., Kessels, R., Valkering, P., and Laes, E. (2020). Reducing winter peaks in electricity consumption: A choice experiment to structure demand response programs. Energy Policy , 137:111183

    2020

  28. [36]

    Takemoto, K. (2024). The moral machine experiment on large language models. Royal Society Open Science , 11(2):231393

    2024

  29. [37]

    Tsuchiya, M. (2018). Performance impact caused by hidden bias of training data for recognizing textual entailment. ArXiv , abs/1804.08117

    2018 arXiv

  30. [38]

    Turper, S., Iyengar, S., Aarts, K., and Gerven, M. (2015). Who is less welcome?: The impact of individuating cues on attitudes towards immigrants. Journal of Ethnic and Migration Studies , 41:239--259

    2015

  31. [39]

    The impact of large language models on official statistics

    United Nations Economic Commission for Europe (2023). The impact of large language models on official statistics. https://unece.org/sites/default/files/2023-12/HLGMOS Accessed: 2024-05-22

    2023

  32. [40]

    Yeo, C. (2022). Welcome to Britain: Fixing our broken immigration system . Biteback Publishing

    2022

  33. [41]

    Zack, T. et al. (2024). Assessing the potential of gpt-4 to perpetuate racial and gender biases in health care: A model evaluation study. The Lancet Digital Health , 6(1):e12--e22

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.