From Delegates to Trustees: How Optimizing for Long-Term Interests Shapes Bias and Alignment in LLMs

Applications, bias, and safety2026ACL AnthologyApproved editorial review

Authors: Suyash Fulay, Jocelyn Zhu, Michiel A. Bakker

Keywords: Computers and Society, Artificial Intelligence

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

3
Authors
7
Findings
14
Limitations
4
Evidence

Editorial summary

English

This EACL 2026 paper maps a classic distinction in political representation onto LLMs. A delegate tries to reproduce a person's expressed preferences; a trustee exercises judgment about what would serve that person's interests. The authors ask what changes when a model stops predicting a profile's vote and instead estimates the short- and long-term consequences of yes and no. The contribution is not proof that either role is correct. It exposes a trade-off: weighting the future can move outputs toward expert-backed positions, but it can also make the system impose its own prior stance more strongly on contested questions.

The experiment is entirely synthetic. GPT-4o generates one hundred demographic combinations for U.S. voters and then writes a detailed biography for each. There are no human participants or observed votes. The policy set contains fifteen contested topics, including immigration, the minimum wage, health care, and pensions, plus five topics for which the authors select an expert-consensus position. Every proposal has an oppositely worded counterpart, producing thirty contested and ten consensus formulations. This reversal provides a wording-sensitivity check, although not every pair is perfectly symmetric in meaning.

Four proprietary snapshots are evaluated at temperature zero: GPT-4o, GPT-4o-mini, Claude 3.7 Sonnet, and Claude 3 Haiku. In the delegate condition, the model receives a biography and predicts how the person would vote, using five prompt variants. In the trustee condition it does not directly cast a vote; it assigns 0-100 utilities to yes and no. One procedure estimates six consecutive five-year periods and applies exponential discounting. A second combines short- and long-term utility at varying weights and uses three utility phrasings. The higher aggregate utility becomes the vote. The authors then compare that vote with the same model's neutral-prompt default or, for the consensus subset, with the answer backed by their sources.

On contested policies, moving from delegate to trustee raises agreement with the model default for all four systems: Claude Sonnet goes from 57% to 72%, Claude Haiku from 65% to 78%, GPT-4o from 63% to 83%, and GPT-4o-mini from 62% to 73%. One immigration example rises from 59% to 92%. The core result is therefore a movement away from the preference attributed to a profile and toward the model's prior position when the model is asked to judge future interest. The paper presents this as a paternalism risk: a system that appears to reason more deeply may represent the provider more and the user less.

Across the five areas labeled expert consensus, trustee votes usually move toward the author-selected target. For a carbon-restriction policy, agreement rises from 75% under delegation to nearly 100% under trusteeship. Group aggregates also differ. For contested items, Republican-profile agreement with the model default rises from 19% to 62%, independent profiles from 78% to 81%, and Democratic profiles move from 92% to 86%. Profiles below USD 50,000 rise from 54% to 76%, compared with 73% to 78% above USD 100,000. On consensus items, Republican profiles rise from 68% to 87% and lower-income profiles from 79% to 88%. These are model outputs for synthetic biographies, not measured effects on those populations.

The model-size conclusion needs particular care. The prose highlights larger default-agreement gaps for Sonnet than Haiku and for GPT-4o than GPT-4o-mini in selected aggregates. Final Table 9, however, uses vote divergence and reports greater delegate-to-trustee divergence for GPT-4o-mini than GPT-4o on both consensus policies, 0.2837 versus 0.1500, and social issues, 0.2868 versus 0.2450. This is not necessarily a numerical contradiction because the metrics differ, but it prevents a general rule that greater size causes more trustee-like behavior or more bias. There are only two pairs, and size is confounded with family, snapshot, and training.

The full-text audit finds another material inconsistency. The method section says every model default agrees with expert consensus on the five consensus areas. Table 5 nevertheless shows Claude answering no both to 'GMOs should NOT be allowed' and to 'GMOs should be allowed because they are safe'; its positive-form answer opposes the listed expert answer, while reversed formulations receive the same vote. This makes tables that treat default and expert consensus as interchangeable ambiguous. It also illustrates that expert consensus is author-curated from literature rather than a direct survey of experts on each exact sentence, and agreement with that target does not establish an individual's welfare.

The final paper adds demographic and model-size t-tests, many with small p-values. Profiles, policies, reversed formulations, and prompt variants are repeatedly reused, however, and the paper does not state clearly which unit enters each test as an independent observation. It reports no multilevel model, clustered inference, multiple-testing correction, or effect confidence intervals. There is no preregistration. Temperature zero supports consistency, but provider outputs can remain nondeterministic, and the paper does not report seeds, retries, parse failures, missing outputs, or repeated runs to measure stability.

The central construct problem is that the entire circuit is LLM-generated. GPT-4o writes profiles, and models generate both the alleged delegate vote and trustee utility. The study does not validate whether biographies represent real people, contain stereotypes, predict actual votes, or whether a one-dimensional utility score represents long-term interest. The connection to sycophancy is conceptual: no sycophancy benchmark is administered and agreement-seeking is not directly manipulated. The paper acknowledges several of these limits and raises autonomy, provider power, and algorithmic monoculture as ethical concerns.

Public reproducibility is low. All 24 published pages, the Responsible NLP Checklist, and the arXiv TeX sources were reviewed, but no code, profiles, executable prompt set, raw outputs, policy file, or statistical scripts were located. Reported API cost is about USD 500. The figures are therefore preserved as reported, not reproduced. A faithful reading is that changing the prompt and utility aggregation can systematically move simulated votes toward model defaults and, on a small curated set, toward expert-backed positions. It does not show that a trustee improves human decisions, that model utility measures welfare, that demographic effects are real, that model size causes the pattern, or that sycophancy has been solved.

Español

Este artículo de EACL 2026 traslada a los LLM una distinción clásica de teoría política. Un agente delegado intenta reproducir las preferencias expresadas de la persona; un agente fiduciario o trustee ejerce su propio juicio sobre lo que favorecería sus intereses. Los autores preguntan qué ocurre cuando un modelo deja de predecir el voto de un perfil y pasa a estimar las consecuencias de votar sí o no a corto y largo plazo. La aportación no es demostrar cuál de los dos roles es correcto, sino mostrar un conflicto: ponderar más el futuro puede acercar respuestas a posiciones respaldadas por expertos, pero también puede hacer que el sistema imponga con más fuerza sus propias inclinaciones en cuestiones discutidas.

El experimento es completamente sintético. GPT-4o genera cien combinaciones demográficas de votantes estadounidenses y después redacta una biografía detallada para cada una. No hay participantes humanos ni votos reales. El conjunto de políticas contiene quince temas controvertidos, por ejemplo inmigración, salario mínimo, sanidad o pensiones, y cinco temas para los que los autores seleccionan una posición de consenso experto. Cada propuesta tiene una formulación inversa, de modo que el corpus incluye treinta enunciados controvertidos y diez de consenso. Esta inversión sirve como comprobación de sensibilidad al wording, aunque algunas parejas no son semánticamente perfectamente simétricas.

Se evalúan cuatro snapshots propietarios a temperatura cero: GPT-4o, GPT-4o-mini, Claude 3.7 Sonnet y Claude 3 Haiku. En la condición delegado, el modelo recibe la biografía y predice cómo votaría esa persona, con cinco variantes de prompt. En la condición trustee no se le pide directamente un voto: asigna utilidades de 0 a 100 a las opciones sí y no. Un procedimiento estima seis periodos consecutivos de cinco años y aplica descuento exponencial; otro combina utilidad a corto y largo plazo con pesos variables y tres formulaciones de utilidad. La opción con mayor utilidad agregada se convierte en voto. Los autores comparan después ese voto con la postura por defecto que el mismo modelo produce ante un prompt neutral o, en el subconjunto de consenso, con la respuesta respaldada por sus fuentes.

En las políticas discutidas, el paso de delegado a trustee aumenta la coincidencia con el default del modelo en los cuatro casos: Claude Sonnet pasa del 57% al 72%, Claude Haiku del 65% al 78%, GPT-4o del 63% al 83% y GPT-4o-mini del 62% al 73%. Un ejemplo de inmigración sube del 59% al 92%. El resultado central es, por tanto, un desplazamiento de las preferencias atribuidas al perfil hacia la posición previa del modelo cuando se le encarga juzgar el interés futuro. El paper lo interpreta como riesgo de paternalismo: un sistema que parece razonar mejor puede reflejar más al proveedor y menos al usuario.

En las cinco áreas clasificadas como consenso, el trustee suele acercarse a la posición seleccionada por los autores. En una política de restricción de carbono, la coincidencia pasa del 75% en delegado a casi el 100%. El patrón también aparece en agregados por grupo. Entre perfiles republicanos, la coincidencia en temas controvertidos con el default pasa del 19% al 62%; entre independientes, del 78% al 81%; entre demócratas, del 92% al 86%. En renta, los perfiles por debajo de 50.000 dólares pasan del 54% al 76%, frente al 73% al 78% de los superiores a 100.000. En consenso, republicanos pasan del 68% al 87% y rentas bajas del 79% al 88%. Estas cifras describen salidas del modelo para biografías sintéticas; no miden un impacto real sobre esos colectivos.

La conclusión sobre tamaño de modelo requiere especial cautela. El texto destaca mayores brechas de coincidencia con el default en Sonnet que en Haiku y en GPT-4o que en GPT-4o-mini para ciertos agregados. Sin embargo, la Tabla 9 final usa divergencia de voto y encuentra una separación delegado-trustee mayor en GPT-4o-mini que en GPT-4o tanto para consenso, 0,2837 frente a 0,1500, como para temas sociales, 0,2868 frente a 0,2450. No es necesariamente una contradicción numérica, porque son métricas distintas, pero sí impide convertir el resultado en una regla general de que más tamaño causa más comportamiento trustee o más sesgo. Solo hay dos pares, además confundidos por familia, versión y entrenamiento.

La auditoría del texto detecta otra inconsistencia material. La sección metodológica afirma que el default de cada modelo coincide con el consenso experto en las cinco áreas. No obstante, la Tabla 5 muestra que Claude responde no tanto a «los GMO no deberían permitirse» como a «los GMO deberían permitirse porque son seguros»; su respuesta positiva contradice la respuesta experta listada y las dos formulaciones inversas reciben el mismo voto. Esto vuelve ambiguas las tablas que tratan default y consenso como equivalentes. También recuerda que «consenso experto» es una selección de los autores a partir de literatura, no una encuesta directa a expertos sobre cada frase, y que alinearse con ese objetivo no demuestra bienestar individual.

Las tablas estadísticas finales añaden t-tests por grupo demográfico y tamaño de modelo, muchos con p pequeño. Pero perfiles, políticas, formulaciones inversas y variantes de prompt se reutilizan, y el artículo no explica con precisión qué unidad entra como observación independiente en cada prueba. No publica un modelo multinivel, errores agrupados, corrección por comparaciones múltiples ni intervalos de confianza. Tampoco hay preregistro. Temperatura cero facilita consistencia, pero los proveedores pueden ser no deterministas y no se informan semillas, reintentos, fallos de parseo, respuestas ausentes o repeticiones para medir estabilidad.

La principal limitación constructiva es que todo el circuito es generado por LLM: GPT-4o redacta perfiles y los modelos generan tanto el supuesto voto delegado como la utilidad trustee. No se valida si las biografías representan a personas reales, si contienen estereotipos, si el delegado predice votos reales o si la utilidad unidimensional representa interés futuro. La relación con sycophancy es conceptual; no se aplica un benchmark de adulación ni se manipula directamente la tendencia a estar de acuerdo. El trabajo reconoce parte de estos límites y advierte sobre autonomía, concentración de poder y monocultura algorítmica.

La reproducibilidad pública es baja. Se revisaron las 24 páginas publicadas, el checklist de NLP responsable y las fuentes TeX de arXiv, pero no se localizan código, perfiles, prompts en formato ejecutable, salidas crudas, archivo de políticas ni scripts estadísticos. El coste declarado es de unos 500 dólares en API. Así, los números se conservan como reportados, no como reproducidos. La lectura fiel es que cambiar el prompt y la agregación de utilidad puede mover de forma sistemática los votos simulados hacia defaults del modelo y, en un pequeño conjunto, hacia posiciones respaldadas por expertos. No demuestra que un trustee mejore decisiones humanas, que su utilidad mida bienestar, que los efectos demográficos sean reales, que el tamaño sea la causa ni que el sistema haya resuelto la sycophancy.

Research question

How do the simulated votes of four LLMs change when they move from imitating the preferences of a synthetic profile, delegate, to aggregating short- and long-term utilities, trustee, and does that change move toward expert consensus or toward the model's own default according to policy, demographics, and size?

Method

Fully synthetic factorial experiment with 100 US biographies generated by GPT-4o, 15 controversial policies and 5 consensus ones, each with inverse formulation. Four models at temperature 0 produce delegate votes under five prompts and trustee utilities under temporal discounting of six periods or short-long weighting with three prompts. Agreements and divergences are compared with neutral defaults and consensus targets; the final tables add t-tests.

Sample: One hundred fully synthetic biographies; four LLM snapshots; fifteen controversial policies and five consensus ones, each in two directions. There are no human participants, observed preferences, or welfare outcomes.

Findings

  • On controversial topics, the trustee increases coincidence with the default in all four models: 15, 13, 20, and 11 percentage points for Sonnet, Haiku, GPT-4o, and GPT-4o-mini.
  • On five topics curated as consensus, weighting the long term more usually increases agreement with the selected expert answer.
  • The shifts are larger for Republican and low-income profiles in the reported aggregates.
  • The general direction of the change is toward the model's prior judgment, not necessarily toward the preference attributed to the profile.
  • The evidence on size depends on the metric: the prose favors a reading of a larger gap in large models, but Table 9 shows greater delegate-trustee divergence in GPT-4o-mini.
  • Table 5 contradicts the asserted equivalence between default and consensus for Claude on the positive formulation about GMO.
  • The results could not be reproduced without public profiles, outputs, policies, and scripts.

Limitations

  • All profiles, votes, and utility scores are generated by LLMs; there is no human validation.
  • Demographic biographies may encode stereotypes and are not audited against representative data.
  • The delegate is not contrasted with real votes and the trustee is not contrasted with real welfare.
  • The unidimensional 0-100 utility is not a validated measure of long-term interest.
  • The expert consensus is curated by the authors and may change; it does not respond directly to each exact sentence.
  • A GMO inconsistency invalidates the claim of perfect coincidence between defaults and consensus.
  • Inverse formulations may differ semantically and some models are sensitive to wording.
  • The inference reuses profiles, policies, pairs, and prompts without clearly modeling their dependence.
  • No intervals, multiple correction, preregistration, or sufficiently explicit unit of analysis are published for the t-tests.
  • Temperature 0 does not eliminate non-determinism and seeds, retries, missingness, and stability tests are missing.
  • There are only two size pairs, confounded by family and snapshot.
  • The relationship with sycophancy is conceptual, not a direct experimental measurement.
  • There is no public code, data, profiles, outputs, or scripts to reproduce the tables.
  • The context is US-based, in English, and with four dated proprietary models.

What the study does not establish

  • It does not demonstrate that the trustee benefits real people.
  • It does not validate that delegate votes represent human preferences.
  • It does not prove that generated utility measures welfare or future interest.
  • It does not demonstrate that all defaults coincide with expert consensus.
  • It does not establish causality of model size.
  • It does not validate the p-values under dependence and multiple comparisons.
  • It does not directly measure or reduce sycophancy.
  • It does not demonstrate real effects on Republicans, Democrats, or income groups.
  • It does not generalize to real users, other countries, languages, or current models.
  • It does not independently reproduce the published results.

Traceability

Scope: Full text

Version: EACL 2026 version of record, pages 5171-5194; arXiv lineage 2510.12689v2

Consulted source: https://aclanthology.org/2026.eacl-long.239.pdf

Review: Codex 24-page visual, official-ACL, Responsible-NLP-checklist, arXiv-source, synthetic-profile, utility-construct, table-consistency, statistics and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • gpt-4o-2024-08-06
  • gpt-4o-mini-2024-07-18
  • claude-3-7-sonnet-20250219
  • claude-3-haiku-20240307

Instruments and metrics

  • Cinco variantes de prompt delegado
  • Seis horizontes consecutivos de cinco años
  • Descuento exponencial con alpha variable
  • Ponderación de utilidad corto-largo con tres formulaciones
  • Prompt neutral para default binario
  • Formulaciones políticas inversas
  • T-tests de una muestra y por pares

Data used

  • 100 perfiles sintéticos de votantes estadounidenses generados por GPT-4o
  • 15 temas controvertidos seleccionados con Comparative Agendas y propuestas de Tessler et al. (2024)
  • 5 temas de consenso curados por los autores a partir de literatura experta
  • 40 formulaciones totales contando inversas: 30 controvertidas y 10 de consenso

Evidence and location

  • Final version, metadata, methods, results, tables, limitations, and ethics: ACL Anthology 2026.eacl-long.239, 24 pages
  • Preprint history and TeX sources v1/v2: arXiv:2510.12689
  • Limitations statements, artifacts, computation, humans, and AI use: Official Responsible NLP Checklist, 2 pages
  • Utility audit, synthetic profiles, defaults, consensus, statistics, and reproducibility: reports/verification/article-238-eacl-trustee-utility-synthetic-voter-default-consensus-statistics-and-artifact-audit.json