{"id":"7d0dbf3e-660b-4a91-9880-f3227f705534","arxiv_id":"2506.21574","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"GPT-3.5 and GPT-4 approximate human immigration preferences in a discrete choice experiment but exhibit systematic biases toward privileged nationalities and occupations.","lead":"This paper tested whether ChatGPT can make fair immigration decisions by giving GPT-3.5 and GPT-4 10,000 hypothetical US immigration cases and comparing their choices to human preferences. The result is a double-edged picture: the models roughly mimic human utility-based priorities, but they also show clear favoritism toward wealthy countries and high-status jobs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fitted decision strategies rest on one-shot API calls with unreported temperature and prompt details; without robustness checks, the human-alignment and bias claims are not yet stable.","rationale":"The reader's CONDITIONAL verdict is appropriate. I agree with the reader's weakest_assumption but would sharpen it: the issue is not only that repeated calls might differ, but that a single prompt configuration and one stochastic realization are the sole evidence for every quantitative claim. The MNL estimates are conditional on the exact prompt and the one sampled response per set; without reporting temperature/seed and without a robustness replication, the comparison with human data in Table 2 cannot be separated from configuration noise. A temperature-0 deterministic rerun is the most direct check: if the large coefficients (e.g., -3.80 for unauthorized entry, 4.39 for contract with employer) are stable, the concerns are mitigated; if they move substantially, the paper overstates the robustness of GPT's 'strategies' and 'biases.' I do not think this requires rejection — the findings are plausible and the tests are feasible — so the verdict remains CONDITIONAL pending these checks.","tokens_in":12226,"tokens_out":11706,"duration_ms":135746,"concrete_test":"Re-run the 10,000 choice sets on GPT-4 with temperature=0 and a fixed seed, and separately with the order of Case 1/Case 2 reversed (same seed); refit the MNL and compare the Table 2 coefficients, especially unauthorized entry, contract-with-employer, doctor, and Germany/Poland. If any coefficient changes by more than 25% or a sign flips, the central strategy and bias claims are not robust to sampling or prompt configuration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing step is the conversion of 10,000 single-shot API responses into stable 'decision strategies.' Section 2.2 says responses were collected 'with default setting parameters' but does not report temperature, top_p, or seed. For the OpenAI chat API the default temperature is 1.0, so each recorded choice is a draw from a stochastic policy. One draw per choice set is not by itself fatal for MNL estimation, but the paper also provides no check that the fitted coefficients are robust to the exact prompt wording, to the order of Case 1 and Case 2, or to the sampling temperature. The headline conclusions — utility maximization, procedural fairness, and nationality/privilege bias — are all read off Table 2 and Figure 2. If a small change in prompt or a different random seed changes which coefficients are large or significant, then the 'strategies' are not properties of GPT-3.5/GPT-4 but artifacts of one unrecorded configuration. Because no code or data are released, this is currently unverifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":12416,"tokens_out":5487,"duration_ms":52629,"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":[{"comment":"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.","section":"§2.2, Table 2"},{"comment":"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.","section":"§2.1, §4"},{"comment":"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.","section":"§2.3, Eq. (1)–(3); §3.1"},{"comment":"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.","section":"§3.2, Table 2"}],"minor_comments":[{"comment":"There is a typo: 'oppcupation' should be 'occupation.'","section":"§3.2"},{"comment":"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.","section":"Table 2"},{"comment":"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.","section":"§2.3"},{"comment":"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.","section":"§2.3"},{"comment":"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.","section":"Figure 3"},{"comment":"The reference 'Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I. (In Press)' lacks a year and venue; this entry should be completed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of cs.CL and addresses a timely question, but the central claims depend on the stability of LLM responses, which the authors have not demonstrated. I would encourage the editor to require a robustness appendix with temperature/seed sensitivity analyses and the release of the choice-generation code (or at least a detailed description of the generator). The comparison with human data would also benefit from a clearer statement of the human sample and estimation details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper. Here's my take.\n\nThe core idea is solid and the paper has a real result: it's the first application of the Hainmueller–Hopkins immigration conjoint to LLM respondents, and it produces a genuinely interesting comparison with the public human data. The finding that GPT-4 is more accepting of Sudanese and Iraqi applicants than human subjects, and the extreme penalty GPT-4 applies to unauthorized entry, are new and policy-relevant. The MNL estimation is standard and the 10,000-choice sample is large. The paper is also candid about its limitations, including dominated choice sets and main-effects-only analysis.\n\nMy main concern is the one you flagged: the LLM sampling configuration is unreported. The paper says 'default setting parameters' but never gives temperature, top_p, or seed. The OpenAI API default temperature is 1.0, so each response is a stochastic draw. One draw per choice set is not fatal in itself, but without a robustness check—same prompts at a lower temperature, repeated runs with different seeds—the fitted coefficients could shift. The headline conclusions about GPT-4's harshness on unauthorized entry, for example, rest on a single coefficient (-3.80) that might be a fluke of one configuration. The paper also doesn't describe the random profile generator, and no code or data are released, so I can't check whether the design worked as intended. These are fixable, but they need to be fixed before the claims are fully supported.\n\nThe in-depth interviews are used as evidence of model reasoning, but they're just chat completions; I'd treat them as illustrative, not as a source of objective findings. Also, the 'utility maximization' conclusion is partly built into the MNL model, so the model fit doesn't independently prove the LLM is utility-maximizing—but the human-LLM comparison doesn't depend on that.\n\nI think this paper deserves a serious referee. The application is novel, the data are large, and the findings would matter for agencies considering LLM screening. The weaknesses are all addressable with more transparent reporting and robustness checks, not with a rethink of the design. So if I were the editor, I'd send it to review. I'd want the authors to release prompts, code, and data, and to report temperature/seed and repeat the estimation under a few configurations.","headline":"Worth evaluating seriously, but the missing sampling details and lack of code/data keep the bias claims from being fully stable.","tokens_in":12920,"tokens_out":2936,"would_cite":false,"duration_ms":33685,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GPT models judge immigrants with human-like strategies but carry their own biases.","keywords":["large language models","immigration decisions","discrete choice experiments","multinomial logit","algorithmic bias","GPT-4","GPT-3.5","fairness"],"falsifier":"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.","tokens_in":12010,"feed_emoji":"⚖️","tokens_out":7713,"duration_ms":75448,"temperature":0.7,"pith_summary":"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.","feed_headline":"GPT models mirror human immigration picks, then add bias","feed_subtitle":"Choice-experiment data show ChatGPT weighs jobs like humans but favors privileged groups.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the DCE design, the nine attributes and levels, and the human benchmark data the LLM responses are compared against.","marker":"Hainmueller and Hopkins (2015)"},{"why":"Provides the conjoint-analysis framework and the public human choice data used in the multinomial logit estimation.","marker":"Hainmueller et al. (2014)"},{"why":"The API through which the 10,000 GPT-3.5 and GPT-4 responses were collected, establishing the experimental instrument.","marker":"OpenAI (2023)"},{"why":"The stated safeguards against discrimination that the fairness analysis tests and qualifies.","marker":"OpenAI (2024)"},{"why":"Precedent for using discrete choice experiments to probe LLM moral and social decision-making.","marker":"Takemoto (2024)"},{"why":"Prior use of choice-based methods to interpret LLM behavior, supporting the method's validity.","marker":"Mohammadi (2024)"},{"why":"Supports the claim that biases in training data carry into model decisions.","marker":"Tsuchiya (2018)"},{"why":"Supports the claim that LLMs exhibit implicit reasoning biases even with safety instructions.","marker":"Gupta et al. (2023)"}],"fun_headline_variants":["LLM gatekeepers mirror human bias in immigration picks","ChatGPT immigration picks: human-like, but biased","LLMs align with human logic, then add nationality bias","GPT-4 harsher on unauthorized entry, favors privileged"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["LLM gatekeepers mirror human bias in immigration picks","ChatGPT immigration picks: human-like, but biased","LLMs align with human logic, then add nationality bias","GPT-4 harsher on unauthorized entry, favors privileged"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000145,"raw_usage":{"total_tokens":1144,"prompt_tokens":874,"completion_tokens":270,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":204}},"tokens_in":490,"tokens_out":270,"duration_ms":4010,"temperature":1.0,"reasoning_tokens":204,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:35:49.385572+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"and Hopkins, D","cited_arxiv_id":null,"evidence_quote":"Supplies the DCE design, the nine attributes and levels, and the human benchmark data the LLM responses are compared against."},{"cited_title":"J., and Yamamoto, T","cited_arxiv_id":null,"evidence_quote":"Provides the conjoint-analysis framework and the public human choice data used in the multinomial logit estimation."},{"cited_title":"Openai api reference","cited_arxiv_id":null,"evidence_quote":"The API through which the 10,000 GPT-3.5 and GPT-4 responses were collected, establishing the experimental instrument."},{"cited_title":"Openai usage policies","cited_arxiv_id":null,"evidence_quote":"The stated safeguards against discrimination that the fairness analysis tests and qualifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Precedent for using discrete choice experiments to probe LLM moral and social decision-making."},{"cited_title":"Performance Impact Caused by Hidden Bias of Training Data for Recognizing Textual Entailment","cited_arxiv_id":"1804.08117","evidence_quote":"Supports the claim that biases in training data carry into model decisions."}],"review_version":1}