CFaiRLLM: Consumer Fairness Evaluation in Large-Language Model Recommender System

Applications, bias, and safety2025ACMApproved editorial review

Authors: Yashar Deldjoo, Tommaso Di Noia

Keywords: Information Retrieval, Computation and Language, Artificial Intelligence, Computers and Society

Source: Open primary source (opens in a new tab)

2
Authors
6
Findings
14
Limitations
10
Evidence

Editorial summary

English

CFaiRLLM studies consumer-side fairness in LLM recommender systems; it neither uses psychological profiles nor evaluates personality. It argues that comparing a neutral recommendation list with one generated from a sensitive attribute can mistake legitimate personalization for bias. The framework adds two elements: it compares recommendations with a user's future interactions, which it calls “true preferences,” and evaluates intersectional attributes. Profiles contain 10 history items selected randomly, by highest rating, or by recency. GPT-3.5 produces recommendations for MovieLens-1M and LastFM-1K; GPT-4o mini is also tested on LastFM. Sensitive attributes are binary sex, three age groups, and their six intersections.

Neutral and sensitive lists are compared with Jaccard@K and PRAG@K, first over all items and then after retaining only recommendations present in each user's temporal test set. Across-group differences are summarized by SNSR and SNSV, where larger values mean greater disparity. In samples of 150 users per domain, sensitive attributes alter lists, intersectional gaps are often larger, and LastFM is less stable than MovieLens. Profile sampling matters: random sampling often favors Jaccard hit overlap, whereas recent or top-rated sampling can favor PRAG ranking agreement; no strategy dominates.

The central claim that true-preference alignment is consistently fairer is not supported consistently by the paper's own numbers. Its random-sampling sex example reports held-out SNSR .0210 and SNSV .0105, versus list-similarity SNSR .0010 and SNSV .0005-21 times higher disparity, not three to ten times lower. Some top-rated scenarios do show reductions, so direction depends on the setting. The paper also uses “significant” without tests, confidence intervals, repeated runs, or multiple-comparison control. One unseeded random sample and unrepeated calls to closed APIs do not quantify variability.

Held-out interaction is an observational preference proxy, not verified truth: it penalizes useful novelty, can reproduce historical behavioral bias, and yields very sparse overlaps. The study does not measure satisfaction, utility, exposure, harm, provider fairness, or causal discrimination. It also omits subgroup sizes, a reproducible 150-user sampling protocol, complete model snapshots, temperature, seeds, and public code. CFaiRLLM is therefore relevant as a recommender-auditing and intersectionality proposal, but provides no evidence about synthetic personality, traits, neuroticism, or psychological exploitation.

Español

CFaiRLLM estudia equidad del lado del consumidor en recomendadores basados en LLM; no usa perfiles psicológicos ni evalúa personalidad. Su punto de partida es que comparar una lista neutral con otra que incluye un atributo sensible puede confundir personalización legítima con sesgo. El marco añade dos elementos: contrasta las recomendaciones con interacciones futuras del usuario, que denomina “preferencias verdaderas”, y examina atributos interseccionales. Los perfiles se construyen seleccionando del historial 10 ítems al azar, mejor valorados o más recientes. GPT-3.5 genera recomendaciones para MovieLens-1M y LastFM-1K; GPT-4o mini también se prueba en LastFM. Los atributos son sexo binario, tres grupos de edad y sus seis intersecciones.

La evaluación compara listas neutrales y sensibles mediante Jaccard@K y PRAG@K, primero sobre todos los ítems y después conservando solo recomendaciones presentes en el test temporal de cada usuario. Las diferencias entre grupos se resumen con SNSR y SNSV, donde valores mayores significan mayor desigualdad. Sobre una muestra de 150 usuarios por dominio, el artículo encuentra que introducir atributos sensibles cambia las listas, que las brechas suelen crecer al combinar sexo y edad y que LastFM es menos estable que MovieLens. El muestreo del perfil importa: aleatorio suele favorecer la coincidencia Jaccard, mientras reciente o mejor valorado puede favorecer el orden PRAG; no hay una estrategia universal.

La afirmación central sobre la superioridad de la alineación con preferencias reales no queda demostrada de forma consistente. El propio ejemplo de sexo con muestreo aleatorio da SNSR 0,0210 y SNSV 0,0105 al filtrar por test, frente a 0,0010 y 0,0005 con similitud de listas: la desigualdad es 21 veces mayor, no entre tres y diez veces menor. Otros escenarios de mejor valorados sí muestran reducciones, por lo que la dirección depende del escenario. Además, “significativo” se usa sin tests, intervalos, repeticiones ni control por comparaciones. Un único muestreo aleatorio sin semilla y llamadas no repetidas a APIs cerradas impiden estimar variabilidad.

El filtro de test es una aproximación observacional a preferencia, no una verdad verificada: penaliza recomendaciones novedosas, hereda sesgos del comportamiento histórico y produce coincidencias muy escasas. El estudio no mide satisfacción, utilidad, exposición, daño, equidad de proveedores ni discriminación causal. Tampoco publica tamaños por subgrupo, protocolo reproducible para los 150 usuarios, snapshots completos del modelo, temperatura, semillas o código. Por ello, CFaiRLLM es relevante como propuesta de auditoría de recomendadores e interseccionalidad, pero no aporta evidencia sobre personalidad sintética, rasgos, neuroticismo ni explotación psicológica.

Research question

How does the evaluation of consumer fairness in LLM-based recommenders change when contrasting lists with future interactions, combining intersectional sensitive attributes, and varying the selection of profile items?

Method

A textual profile is created from N=10 items of each user's history by means of random sampling, by best rating, or by recency. GPT-3.5 is requested, and in LastFM, GPT-4o mini, a neutral list and lists conditioned by sex, age, or sex×age. The lists are matched with the catalog using regular expressions and difflib. Jaccard@K and PRAG@K quantify item and order coincidence; SNSR and SNSV summarize disparities between groups. A second evaluation filters the recommendations by the user's temporal test interactions.

Sample: 150 users are selected per domain to reduce API cost. In MovieLens, the train reports 150 users, 2,537 items and 18,428 interactions; the test, 150 users, 1,590 items and 4,023 interactions. In LastFM, the train reports 149 users, 21,967 items and 37,534 interactions; the test, 150 users, 7,308 items and 9,460 interactions. No seed, representativeness procedure, sizes of sensitive subgroups, or repetitions are reported.

Findings

  • Adding sex, age, or their intersection alters the recommendations with respect to the neutral prompt; the change of list does not by itself prove bias or harm.
  • The largest gaps are frequently observed in intersectional attributes and in LastFM, a more dispersed and less structured domain than MovieLens.
  • The strategy that builds the profile changes the results. Random usually improves Jaccard, while recency or best ratings can improve PRAG; none dominates across all models, domains, and attributes.
  • GPT-4o mini appears to show smaller gaps than GPT-3.5 in several LastFM comparisons, but there are no repetitions or statistical inference to attribute that difference to the model.
  • The profile size modifies the metrics in a non-monotonic way. The text and the figures are not fully consistent on the values explored: the visualization shows 5, 10, and 20, while one passage mentions 5 and 15.
  • The claim that alignment with test produces gaps between three and ten times smaller contradicts the random sex example: 0.0210/0.0105 versus 0.0010/0.0005, that is, 21 times larger for SNSR and SNSV.

Limitations

  • Future interactions are called "true preferences", but they are only observed behavior. They do not verify satisfaction and may contain historical bias, popularity, exposure restrictions, or noise.
  • Filtering by items already present in the test penalizes every novel recommendation and leaves very scarce overlaps; the resulting metrics may be unstable.
  • SNSR and SNSV measure equality of stability with respect to the neutral prompt, a limited definition of fairness. They do not measure utility, quality, satisfaction, exposure, harm, provider fairness, or causal discrimination.
  • Sex is reduced to male/female and age to three bins; the validity, provenance, consent, or accuracy of those attributes is not justified.
  • The sizes of the six intersectional subgroups are not reported. Ranges and deviations may depend strongly on small or imbalanced groups.
  • The sample of 150 users is called representative without a protocol, seed, or sensitivity analysis. LastFM even contains 149 users in train and 150 in test.
  • Random sampling and API responses are not repeated. There are no seeds, intervals, bootstrap, hypothesis tests, multiple correction, or power analysis.
  • The article uses "significant" in a descriptive way and presents a directional claim incompatible with part of its own results.
  • GPT-3.5 and GPT-4o mini are changing APIs. Exact snapshots, execution dates, temperature, system messages, retries, and other parameters necessary to replicate the outputs are missing.
  • The difflib threshold used to link generated titles with the catalog is not reported; parsing or matching errors may alter all subsequent metrics.
  • The transformation of implicit LastFM plays to a 1–5 scale adds assumptions and does not convert those values into declared preferences.
  • No code, executable configuration, complete prompts in reusable format, logs, model responses, or tabular results are published to verify the figures.
  • The gray cells of the appendix are considered not relevant, but the text interprets some cross effects; the inclusion criterion is not formalized.
  • The evaluation is observational and offline. It does not include users, human judgments on quality or fairness, longitudinal experiments, privacy audit, or mitigations.

What the study does not establish

  • It does not demonstrate that any difference between a neutral list and a sensitive list is bias, unfairness, or harm.
  • It does not establish that the test history is a true, complete preference free of exposure biases.
  • It does not demonstrate a universal superiority of alignment with test over list similarity; its results change direction depending on the scenario and metric.
  • It does not prove that GPT-4o mini is intrinsically fairer than GPT-3.5 or that the gaps are statistically distinct from sampling or generation variation.
  • It does not allow inferring causal discrimination, user experience, benefit, or real harm to any group.
  • It does not study psychological personality, Big Five, neuroticism, psychometric profiles, persistent identity, or emotional exploitation by recommenders.

Traceability

Scope: Full text

Version: arXiv:2403.05668v3, submitted 8 March 2024, revised 20 February 2025; author preprint of ACM TIST 2025, 25 pages

Consulted source: https://arxiv.org/pdf/2403.05668v3

Review: Codex full-text, bilingual-fidelity, 25-page visual, arXiv-v3, ACM-TIST-2025, recommender-fairness, intersectionality, construct-validity, metric-consistency, statistical-claim, API-reproducibility, code-availability, privacy and personality-relevance audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5
  • GPT-4o mini

Instruments and metrics

  • Random, top-rated and recent profile sampling
  • Neutral, sex, age and sex-by-age recommendation prompts
  • Regular-expression parsing and Python difflib title matching
  • Jaccard@K
  • PRAG@K pairwise ranking accuracy
  • Sensitive-to-Neutral Similarity Range (SNSR)
  • Sensitive-to-Neutral Similarity Variance (SNSV)
  • Temporal held-out interaction filter as a preference proxy

Data used

  • MovieLens-1M
  • LastFM-1K

Evidence and location

  • Motivation, scope, and contributions: Paper, pp. 1–4, Abstract, Introduction and Contributions
  • CFaiRLLM framework, profiles, and sensitive attributes: Paper, pp. 8–12, Sections 3.1–3.3 and Figures 1–2
  • Jaccard, PRAG, alignment with test, SNSR, and SNSV: Paper, pp. 12–14, Sections 3.4–3.5 and equations
  • Datasets, temporal split, sample, models, and parsing: Paper, pp. 14–16, Section 4.1 and Table 3
  • Results on metric, intersectionality, and model: Paper, pp. 16–19, Sections 4.2–4.4 and Figures 3–5
  • Sampling, profile size, and internal contradictions: Paper, pp. 19–20, Sections 4.5–4.6 and Figures 6–7
  • Conclusions, declared limitations, and future work: Paper, pp. 20–22, Sections 5–7
  • Detailed numerical results by attribute and strategy: Paper, pp. 22–23, Appendix tables
  • Absence of cited reproducible artifact: Paper, all 25 pages; no code or data release statement or repository URL
  • Comprehensive visual inspection: Paper, all 25 rendered pages, including every figure, table and appendix page