Destination (Un)Known: Auditing Bias and Fairness in LLM-Based Travel Recommendations

Applications, bias, and safety2025mdpi.comApproved editorial review

Authors: Hristo Andreev, Petros Kosmas, Antonios D. Livieratos, Antonis Theocharous, Anastasios Zopiatis

Keywords: Personas demográficas, Recomendaciones turísticas, Sesgo de exposición, Sesgo geográfico, Evaluación de equidad, Sistemas desplegados, Reproducibilidad

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

5
Authors
10
Findings
15
Limitations
7
Evidence

Editorial summary

English

This paper audits how travel recommendations from two deployed conversational services, ChatGPT-4o build 20250326 and DeepSeek-V3 build 0324, vary with a traveller persona's origin, age, gender identity, and tourism interest. It is relevant to synthetic personality as demographic-persona and context conditioning, but it neither induces nor measures personality traits, persistent identity, or a stable user model. The audited source is the final 30-page AI 2025 article; every page was rendered and visually inspected.

The design crosses eight origins, three ages, three gender identities, and three interests, producing 216 chains per system. Each chain requests five destinations under a generic persona, five for a theme, sun and sea, cultural heritage, or wildlife, and five more in a follow-up. The authors operate the web interfaces in fresh private sessions, clear cookies, match a VPN to the stated origin, randomize order, and disable tools or search. Temperature, top-p, seed, system prompt, and safety filters are not fixed: the comparison is between deployed products under vendor defaults, not isolated weights. Execution dates and request identifiers are also absent.

The 6,480 figure is a theoretical maximum: 216 chains times two systems times three prompts times five items. Methods define it before cleaning and deduplication even though the abstract says that many recommendations were generated. The retained total and disposition of missing, refused, regenerated, and duplicate items are not reported. Interest is also absent from the generic prompt, leaving 72 distinct generic demographic strings per system, each repeated three times because of the later interests, rather than 216 unique generic baselines.

The city-popularity subset contains 1,020 valid ChatGPT and 1,050 valid DeepSeek recommendations against 1,080 requested per system; 50.0% and 57.9% fall outside Euromonitor's Top 100. At country level, 30.5% and 34.8% fall outside the Travel & Tourism Development Index Top 30. Frequency correlation with TTDI is weak and nonsignificant: r=.262, p=.239 for ChatGPT and r=.167, p=.523 for DeepSeek. These cutoffs show exposure differences, but TTDI measures tourism-development conditions rather than popularity, and off-list status does not establish relevance, fairness, sustainability, or genuine long-tail exposure.

Origin portfolios differ as well. DeepSeek has higher Jensen–Shannon distance than ChatGPT for 23 of 28 origin pairs, with an average rounded difference near .06. Aggregate domestic share is 22.8% for ChatGPT and 34.6% for DeepSeek, but heterogeneity dominates: India is 93% versus 85%, the United States 4% versus 59%, Japan 37% versus 74%, and Saudi Arabia 0% for both. Different portfolios can indicate personalization, error, or bias; without preferences, relevance, utility, an exposure target, or observed harm, distributional separation cannot decide which. Prompted country and VPN geolocation are manipulated together, so textual persona effects cannot be separated from localization or provider serving.

For cultural distance, the paper weights absolute national gaps on six Hofstede dimensions. Published rows reproduce a mean inter-system Euclidean distance of 7.373, with Saudi Arabia at 10.612 and Japan at 2.664. Using rounded domestic percentages from the figure, correlations reproduce at approximately -.758 for ChatGPT and -.698 for DeepSeek, but only eight origins are involved. All six mean differences are nominally positive for DeepSeek, yet some cells are negative and mean uncertainty-avoidance difference is only .126 on a 0–100 scale. Hofstede scores are static ecological national averages; cultural distance is not itself individual mismatch or bias.

The cliché analysis uses an unreleased 150-term lexicon. Across 2,160 recommendations per system, the generic and thematic prompts, excluding follow-up, ChatGPT has 1,864 tokens (.863 per item) and DeepSeek 1,931 (.894), only 3.6% greater in relative density. ChatGPT uses 85 types and DeepSeek 63, yielding the reported .046 and .033 ratios. Mean density cannot show that almost every item contains a cliché: the paper defines coverage as the percentage of items with at least one match but never reports it. The lexicon also lacks human validation, precision, recall, or contextual error analysis; it measures stock promotional language, not the full construct of social stereotyping.

Published symmetric KL divergences by gender are 1.260 and 1.468 for female versus male; 4.867 and 8.771 for female versus non-binary; and 3.964 and 5.897 for male versus non-binary, for ChatGPT and DeepSeek respectively. These are large under study-created heuristic bands, not validated harm thresholds. In sparse strata with many countries, KL is highly sensitive to additive smoothing, whose epsilon is not reported. Distributional difference does not establish worse treatment, stereotyping, or harm without quality, safety, preference, or outcome measures.

For non-binary personas, country frequency correlates with LGBTI acceptance: ChatGPT Pearson r=.367, p=.023 and Spearman rho=.455, p=.004; DeepSeek r=.419, p=.015 and rho=.389, p=.025. This is a compositional association without adequate controls for region, popularity, income, or infrastructure. The same correlation is not compared with female and male portfolios, so neither specificity to non-binary personas nor protective intent is established. Correlations with Numbeo's crowdsourced safety index are weak; that does not show the systems ignore safety.

Overlap between the thematic list and the turn explicitly asking for other places is 6.74% for ChatGPT and 8.25% for DeepSeek: novelty of 93.26% and 91.75%, with zero overlap in 73.1% and 69.9% of chains. This is a two-turn country-level measure under an instruction demanding alternatives. It does not support a general claim of minimal reinforcement, absence of filter bubbles, or long-run stability.

The audit also finds reporting defects. DeepSeek's Table 5 is not sorted by frequency: the United States appears seventh with 30 above countries with 41–52. Table 10 labels a ChatGPT age matrix as gender. Methods promise proportion tests, mixed logistic models, controlled regressions, robust errors, and random effects, but results omit coefficients, standard errors, intervals, formulas, diagnostics, and most p-values. No multiplicity correction is reported, and severity bands are unvalidated; the KL bands even leave a gap between 1.5 and 1.6.

Data are available only upon request and no public artifact is linked. Missing materials include the 6,480 requested outputs, cleaned item-level table, timestamps, sessions, refusals and retries, cliché lexicon, ranking snapshots, joined indices, geocoding, smoothing epsilon, code, fitted models, environment, tests, and immutable archive. Published tables allow partial arithmetic checks, but the experiment is not reproducible end to end.

The defensible contribution is a time-stamped audit showing that two specific chat interfaces change destination distributions and language as demographic persona, origin, geolocation, theme, and turn change. This is useful evidence about deployed-system sensitivity. It does not establish personality, causal weight effects, fairness harm, protective intent, amplification against an external baseline, absence of long-term reinforcement, or benefit from the proposed public reranker, which is neither implemented nor evaluated.

Español

El artículo audita cómo varían las recomendaciones turísticas de dos servicios conversacionales desplegados, ChatGPT-4o, build 20250326, y DeepSeek-V3, build 0324, al cambiar país de origen, edad, identidad de género e interés turístico de la persona. Es relevante para personalidad sintética como evaluación de condicionamiento por persona demográfica y contexto, pero no induce ni mide rasgos de personalidad, identidad persistente o un modelo estable del usuario. La fuente auditada es el artículo final de 30 páginas publicado en AI 2025, todas ellas renderizadas e inspeccionadas visualmente.

El diseño cruza ocho orígenes, tres edades, tres identidades de género y tres intereses, formando 216 cadenas por sistema. Cada cadena solicita cinco destinos con una persona genérica, cinco con un tema, sol y playa, patrimonio cultural o fauna, y otros cinco en un turno de seguimiento. Los autores operan las interfaces web en sesiones privadas nuevas, borran cookies, alinean una VPN con el origen declarado, aleatorizan el orden y desactivan herramientas o búsqueda. No fijan temperatura, top-p, seed, prompt de sistema o filtros: comparan productos desplegados con sus defaults, no pesos aislados. Tampoco informan fechas de ejecución o identificadores de petición.

La cifra de 6.480 recomendaciones es el máximo teórico: 216 cadenas por dos sistemas por tres prompts por cinco ítems. El método la define antes de limpieza y deduplicación, aunque el abstract dice que se generó esa cantidad. No se publica el total retenido ni un flujo de faltantes, rechazos, regeneraciones y duplicados. Además, el interés no aparece en el prompt genérico: por sistema existen 72 cadenas demográficas genéricas distintas, repetidas tres veces por los intereses posteriores, no 216 baselines genéricos únicos.

En el subconjunto de popularidad urbana se analizan 1.020 recomendaciones válidas de ChatGPT y 1.050 de DeepSeek, frente a 1.080 solicitadas por sistema; el 50,0 % y el 57,9 % quedan fuera del Top-100 de Euromonitor. A escala de país, el 30,5 % y el 34,8 % quedan fuera del Top-30 del Travel & Tourism Development Index. La correlación entre frecuencia y TTDI es débil y no significativa: r=.262, p=.239 para ChatGPT y r=.167, p=.523 para DeepSeek. Estos cortes binarios muestran distinta exposición, pero TTDI mide condiciones de desarrollo turístico, no popularidad, y estar fuera de una lista no demuestra relevancia, equidad, sostenibilidad o auténtica larga cola.

Las distribuciones por origen también difieren. DeepSeek tiene mayor distancia Jensen–Shannon que ChatGPT en 23 de 28 pares de origen, con una diferencia media redondeada cercana a .06. La cuota doméstica agregada es 22,8 % para ChatGPT y 34,6 % para DeepSeek, pero la heterogeneidad domina: India 93 % frente a 85 %, Estados Unidos 4 % frente a 59 %, Japón 37 % frente a 74 % y Arabia Saudí 0 % en ambos. Una cartera distinta por origen puede ser personalización, error o sesgo; sin preferencias, relevancia, utilidad, objetivo de exposición o daño observado, la separación distributiva no decide cuál. País en el texto y geolocalización por VPN también se manipulan juntos, de modo que no se separa la persona verbal de la localización o del serving del proveedor.

Para distancia cultural, el artículo pondera brechas nacionales absolutas en seis dimensiones de Hofstede. Las filas publicadas reproducen una distancia euclídea intermodelo media de 7,373: Arabia Saudí alcanza 10,612 y Japón 2,664. Con los porcentajes domésticos redondeados de la figura, la correlación se reproduce aproximadamente como −.758 para ChatGPT y −.698 para DeepSeek, pero solo hay ocho orígenes. Las seis diferencias medias favorecen nominalmente a DeepSeek; aun así existen celdas negativas y la diferencia media en evitación de incertidumbre es apenas .126 sobre una escala 0–100. Hofstede son promedios nacionales ecológicos y estáticos; distancia cultural no equivale por sí sola a sesgo o desajuste individual.

El análisis de clichés usa un léxico no liberado de 150 términos. Sobre 2.160 recomendaciones por sistema, los prompts genérico y temático, no el seguimiento, aparecen 1.864 tokens en ChatGPT (.863 por ítem) y 1.931 en DeepSeek (.894), solo un 3,6 % relativo más. ChatGPT emplea 85 tipos y DeepSeek 63, de donde salen las razones .046 y .033. Esa densidad media no permite afirmar que casi todos los ítems contienen un cliché: el paper define cobertura como porcentaje de ítems con al menos una coincidencia, pero no publica ese porcentaje. Tampoco valida el léxico con anotación humana, precisión, recall o errores contextuales; mide lenguaje promocional estereotipado, no todo el constructo de estereotipo social.

Las divergencias KL simétricas por género son 1,260 y 1,468 entre mujer y hombre; 4,867 y 8,771 entre mujer y persona no binaria; y 3,964 y 5,897 entre hombre y persona no binaria, para ChatGPT y DeepSeek respectivamente. Son diferencias grandes según bandas heurísticas creadas por el estudio, no umbrales validados de daño. En estratos dispersos con muchos países, KL depende fuertemente del suavizado aditivo, cuyo epsilon no se informa. Una diferencia de distribución tampoco prueba peor trato, estereotipo o perjuicio sin evaluar calidad, seguridad, preferencia o resultado.

Para personas no binarias, la frecuencia por país correlaciona con aceptación LGBTI: ChatGPT Pearson r=.367, p=.023 y Spearman rho=.455, p=.004; DeepSeek r=.419, p=.015 y rho=.389, p=.025. Es una asociación composicional sin controles suficientes por región, popularidad, renta o infraestructura. No se compara esa correlación con las carteras de mujeres y hombres, por lo que no queda establecido que sea específica de la persona no binaria ni que refleje una intención protectora. Las correlaciones con el índice crowdsourced de seguridad de Numbeo son débiles; esto tampoco demuestra que los sistemas ignoren la seguridad.

El solapamiento entre la lista temática y el turno que pide expresamente otros lugares es 6,74 % para ChatGPT y 8,25 % para DeepSeek: novedad de 93,26 % y 91,75 %, con cero solapamiento en 73,1 % y 69,9 % de las cadenas. Es una medida de dos turnos adyacentes, a nivel de país, bajo una instrucción que exige alternativas. No sustenta la afirmación general de refuerzo mínimo, ausencia de burbujas o estabilidad a largo plazo.

La auditoría detecta además problemas de reporte. La Tabla 5 de DeepSeek no está ordenada por frecuencia: Estados Unidos aparece séptimo con 30, por encima de países con 41–52. La Tabla 10 etiqueta una matriz de edad de ChatGPT como género. Los métodos prometen pruebas de proporciones, modelos logísticos mixtos, regresiones con controles, errores robustos y efectos aleatorios, pero resultados no publica coeficientes, errores estándar, intervalos, fórmulas, diagnósticos ni la mayoría de p-valores. No hay corrección por multiplicidad y las bandas descriptivas no están validadas; las de KL incluso dejan un hueco entre 1,5 y 1,6.

La disponibilidad indica que los datos pueden solicitarse, pero no hay artefacto público. Faltan las 6.480 salidas solicitadas, tabla limpia por ítem, timestamps, sesiones, rechazos y reintentos, léxico de clichés, snapshots de rankings, joins de índices, geocodificación, epsilon de suavizado, código, modelos estadísticos, entorno, tests y archivo inmutable. Por ello, las cifras publicadas pueden auditarse parcialmente en las tablas, pero el experimento no puede reproducirse de extremo a extremo.

La contribución defendible es una auditoría fechada que muestra que dos interfaces concretas cambian sus distribuciones de destino y su lenguaje al variar personas demográficas, origen, geolocalización, tema y turno. Es evidencia útil sobre sensibilidad de productos desplegados. No demuestra personalidad, efectos causales de los pesos, daño de equidad, intención protectora, amplificación respecto a un baseline externo, ausencia de refuerzo a largo plazo ni eficacia del reranker público propuesto, que no se implementa ni evalúa.

Research question

How do exposure to destinations, geographic and domestic concentration, cultural distance, promotional language, age and gender distributions, and novelty vary between ChatGPT-4o and DeepSeek-V3 turns when a tourist persona and its geolocation are changed?

Method

Factorial audit of two web interfaces deployed with eight origins, three ages, three gender identities, and three interests. 216 chains are executed per system with three prompts of five destinations, private sessions, cookies cleared, and VPN aligned to the origin. Euromonitor/TTDI cuts, JSD, Hofstede gaps, coincidences with a lexicon of 150 clichés, symmetric KL, and correlations with LGBTI acceptance and safety are calculated. The editorial audit visually reviewed the 30 pages and recalculated published denominators, ratios, and distances.

Sample: 216 chains per system from 8 origins × 3 ages × 3 gender identities × 3 interests. Each chain requests 15 destinations in three turns, for a theoretical maximum of 6,480 across both systems. The generic prompt contains only 72 distinct demographic personas per system and is repeated three times; the analyses use distinct subsets and the final clean n is not published.

Findings

  • Outside the urban Top-100: ChatGPT 50.0% of n=1,020; DeepSeek 57.9% of n=1,050.
  • Outside the Top-30 TTDI per country: 30.5% and 34.8%; correlations with frequency not significant.
  • DeepSeek exceeds ChatGPT's JSD in 23 of 28 origin pairs.
  • Aggregate domestic share 22.8% versus 34.6%, with extreme variation across origins.
  • Mean intermodel Hofstede distance reproduced of 7.373, based on eight origins.
  • Clichés: 1,864 versus 1,931 tokens over 2,160 items per system; relative density difference 3.6%.
  • KL woman/man 1.260 versus 1.468; woman/non-binary 4.867 versus 8.771.
  • Frequency for non-binary persona correlates with LGBTI acceptance in both systems, with no comparison between genders or causal identification.
  • Novelty between thematic prompt and follow-up 93.26% versus 91.75%, partly induced by asking for other places.
  • Table 5 has a sorting error and Table 10 an incorrect label.

Limitations

  • 6,480 is a theoretical maximum; the total retained and the flow of exclusions and retries are not reported.
  • There are only 72 distinct generic prompts per system, repeated three times for the subsequent thematic factor.
  • Deployed products with different defaults are compared, not model weights under identical conditions.
  • Seed, temperature, top-p, system prompt, filters, or backend are not fixed and there are no execution dates or IDs.
  • Textual persona and VPN geolocation are manipulated together.
  • Euromonitor and TTDI are incomplete binary proxies; TTDI is not a popularity ranking.
  • Distributive difference or cultural distance do not automatically equate to unfair bias or harm.
  • Hofstede uses ecological national averages; the cultural correlation has only eight observations.
  • The cliché lexicon is not published or validated and per-item coverage is not reported.
  • KL depends on additive smoothing whose value is not reported.
  • LGBTI correlations are compositional and are not compared between genders.
  • Numbeo is crowdsourced and not a validated universal safety index.
  • Novelty only compares two turns and the second explicitly asks for alternatives.
  • Statistical models, coefficients, errors, intervals, diagnostics, and multiple correction promised are missing.
  • Data upon request does not equate to public availability; there is no code or reproducible artifact.

What the study does not establish

  • It does not establish personality or a persistent model of the traveler.
  • It does not isolate causal effects of weights versus the complete deployed system.
  • It does not demonstrate that being outside a list is fair, relevant, or sustainable exposure.
  • It does not demonstrate that every difference by origin, gender, or age constitutes harm.
  • It does not validate Hofstede as individual cultural mismatch nor promotional clichés as all stereotypes.
  • It does not demonstrate worse treatment of non-binary persons or protective intent.
  • It does not demonstrate that the association with LGBTI acceptance is specific to non-binary persons.
  • It does not demonstrate absence of reinforcement, bubbles, or long-term drift.
  • It does not demonstrate structural amplification without a human, market, corpus, or demand baseline.
  • It does not demonstrate that the proposed reranker improves equity or sustainability without reducing utility.
  • It does not allow reproducing the study end to end.

Traceability

Scope: Full text

Version: AI 2025, 6(9), article 236, DOI 10.3390/ai6090236; final 30-page CC BY 4.0 article. Every page was rendered and visually inspected from the Cyprus University of Technology repository copy. The publisher record and data-availability statement were checked; no public data, code or supplement is linked.

Consulted source: https://www.mdpi.com/2673-2688/6/9/236

Review: Codex full-text, 30-page visual, bilingual-fidelity, construct, arithmetic, statistical-reporting and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • ChatGPT-4o deployed chat UI, build 20250326
  • DeepSeek-V3 deployed chat UI, build 0324

Instruments and metrics

  • Euromonitor Top 100 City Destinations Index
  • World Economic Forum Travel & Tourism Development Index Top 30
  • Pairwise Jensen-Shannon distance by origin
  • Domestic recommendation share
  • Six Hofstede national cultural dimensions
  • Unreleased 150-term tourism cliché lexicon
  • Symmetric KL divergence with unspecified additive smoothing
  • LGBTI Global Acceptance Index
  • Numbeo crowdsourced Safety Index
  • Adjacent-prompt country-set overlap

Data used

  • Theoretical 6,480 requested recommendation slots; retained total not reported
  • City-popularity subsets: ChatGPT n=1,020, DeepSeek n=1,050
  • Cliché subsets: 2,160 recommendation rationales per system
  • Paper tables and figures only; item-level prompts and outputs not public
  • No public code, lexicon, joined indices, session logs or fitted models

Evidence and location

  • Publication, systems, factorial design, and acquisition: AI 2025, 6(9), 236, pp. 1-6, Abstract and Sections 1-2
  • Popularity, JSD, and domestic recommendations: AI 2025, 6(9), 236, pp. 6-13, Sections 3.1-3.2, Tables 3-6 and Figures 1-2
  • Cultural distance and clichés: AI 2025, 6(9), 236, pp. 13-19, Sections 3.3-3.4, Tables 7-9 and Figures 3-5
  • Demographic divergence and acceptance and safety correlations: AI 2025, 6(9), 236, pp. 19-23, Section 3.5, Tables 10-12 and Figure 6
  • Overlap, discussion, reranker, and conclusions: AI 2025, 6(9), 236, pp. 23-28, Sections 3.6-5 and Figure 7
  • Declared limitations, data availability, and references: AI 2025, 6(9), 236, pp. 28-30, Limitations, Data Availability and References
  • Validity, arithmetic, construct, and reproducibility audit: reports/verification/article-222-travel-recommendation-bias-validity-and-reproducibility-audit.json