GermanPartiesQA: Benchmarking Commercial Large Language Models and AI Companions for Political Alignment and Sycophancy

Applications, bias, and safety2025AAAIApproved editorial review

Authors: Jan Batzner, Volker Stocker, Stefan Schmid, Gjergji Kasneci

Keywords: GermanPartiesQA, Political alignment, Party-position factuality, Persona-based steerability, Wahl-o-Mat, Sycophancy terminology, AIES 2025

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

4
Authors
16
Findings
24
Limitations
9
Evidence

Editorial summary

English

GermanPartiesQA separates three questions that must not be conflated: whether an LLM knows a party's official position, how often its own forced-choice answers match party answers, and how much it changes when given a politician context. It does not study personality. Its “personas” are short prompts containing a real parliamentary leader's name, party, gender, birth year and education; political affiliation is supplied explicitly. The outcome is context-conditioned steerability, not trait inference or an internal political personality.

The paper compiles Wahl-o-Mat statements from ten state elections and one federal election held between 2021 and 2023. It reports 418 items with Agree, Disagree or Neutral answers from SPD, Greens, Die Linke, CDU-CSU, FDP and AfD. Six commercial models were evaluated in January 2025: gpt-4o-2024-08-06, gpt-3.5-turbo-0125, claude-3-sonnet-20240229, claude-2.1, command-r-plus-04-2024 and command. Each statement was queried separately ten times, and temperatures 0 and 1 were compared. Prompts and statements were in German.

The first test directly asks what each party answered and requires an exact match. GPT-4o has the best party-level values, ranging from 69.38% for SPD to 87.11% for AfD; Command has the lowest, from 47.23% for AfD to 65.22% for Die Linke. The authors emphasize errors for the centrist SPD and CDU-CSU. This outcome measures recall or reconstruction of historical electoral positions, not the model's own political position. It also does not establish accurate electoral advice in open dialogue.

The second test asks the model to answer Agree, Disagree or Neutral itself. The primary alignment score mirrors Wahl-o-Mat: 1 for a match, 0 for opposition and .5 when a polar answer is compared with Neutral. Models generally score higher against SPD, Greens and Die Linke and lower against AfD. GPT-4o, for example, reaches 75.78% with Greens and 34.59% with AfD; Command R+ reaches 70.93% with SPD and 41.32% with AfD. The paper describes model-specific alignment and a left-green pattern in several comparisons, especially for newer OpenAI and Cohere variants.

That interpretation needs a central caveat: the score mixes content with response style. In the released CSV, SPD, Greens and Die Linke have more Agree labels than AfD, while models differ in acquiescence and Neutral use. Under exact match, Claude 3 Sonnet drops by 36.4-40.3 percentage points and Claude 2.1 by 43.2-47.0 because both often choose Neutral; Command and Command R+ fall much less. A left-green ordering remains in several cells, but absolute levels and some differences depend on label prevalence and answer propensity, not ideology alone. The analysis does not adjust marginals, item difficulty, region or time with a statistical model.

The third test adds “I am [politician]” or “You are [politician].” Both shift answers toward the named party and nearby parties, with large provider differences. In one CDU-CSU example, GPT-4o increases by more than 24 points, Command R+ by more than 12, and Claude 3 Sonnet by only a little over 2. Yet the largest relative shift does not always yield the largest absolute match with the requested party: under an AfD persona, models may still match CDU-CSU or FDP more than AfD. I am and You are produce nearly indistinguishable patterns. The paper therefore concludes that this design does not identify sycophancy: it measures neither intent to flatter, human perception nor agreement with a user's opinion against objective truth. Persona-based political steerability is the faithful term.

The Character.ai demonstration is separate from the main benchmark. It tests one user-created Alice Weidel chatbot, selected by interaction count, on a subset of 2023 Berlin questions and reports 58% agreement with AfD. It is neither an author-controlled model nor representative evidence about the 27 bots or 364,542 chats mentioned, and its transcripts are not released.

The public-repository audit found a material gap between artifact and paper. At commit 198d92f, the repository contains only a README and GermanPartiesQA-1.csv: no experiment code, executable prompts, outputs, figures, tests, environment, releases or license. The CSV parses to 413 rows rather than 418; BY-18, HB-13, HB-26, SH-18 and SL-29 are absent. The paper says seven parties but names and releases six. Thirty-six AfD positions are empty.

Bavarian rationales are also systematically misattached. Comparing Bavaria A1-A38 with Schleswig-Holstein J1-J38, 34 of 37 shared-ID pairs have mean six-party rationale similarity above .95; the overall mean is .987. Thus BY A1 on shop opening carries SH J1 wind-energy rationales, BY A2 on border police carries vaccination rationales, and BY A3 on organic agriculture carries Autobahn A20 rationales. English translations also contain clear truncation and mistranslation. Because the main analysis uses German statements and categorical position labels, this corruption does not by itself prove the published scores wrong. It does invalidate the released rationales, contradict the public 418-row claim and make tables and figures unreproducible.

The faithful conclusion is that, in German forced-choice prompts with January 2025 commercial snapshots, models differ in factual recall, response distribution and sensitivity to explicit party-and-politician contexts. The paper offers a useful critique of indiscriminate “sycophancy” terminology. It does not establish synthetic personality, stable internal ideology, faithful politician simulation, voter persuasion, production behavior or a clean and fully reproducible public benchmark.

Español

GermanPartiesQA separa tres preguntas que no deben confundirse: si un LLM conoce la posición oficial de un partido, cuánto coincide su propia respuesta forzada con las respuestas partidistas y cuánto cambia al recibir el contexto de un político. No estudia personalidad. Las “personas” son prompts breves con nombre, partido, género, año de nacimiento y formación de líderes parlamentarios reales; la afiliación política se entrega explícitamente. El resultado es steerability condicionada por contexto, no inferencia de rasgos ni una personalidad política interna.

El paper compila declaraciones de Wahl-o-Mat de diez elecciones regionales y una federal entre 2021 y 2023. Reporta 418 ítems, cada uno con respuestas Agree, Disagree o Neutral de SPD, Verdes, Die Linke, CDU-CSU, FDP y AfD. Evalúa seis modelos comerciales en enero de 2025: gpt-4o-2024-08-06, gpt-3.5-turbo-0125, claude-3-sonnet-20240229, claude-2.1, command-r-plus-04-2024 y command. Cada declaración se consulta separadamente diez veces; se comparan temperaturas 0 y 1. El prompt y los ítems son alemanes.

La primera prueba pregunta directamente qué respondió cada partido y exige coincidencia exacta. GPT-4o obtiene las mejores cifras por partido, entre 69,38 % para SPD y 87,11 % para AfD; Command las menores, entre 47,23 % para AfD y 65,22 % para Die Linke. Los autores destacan errores en los partidos centristas SPD y CDU-CSU. Este resultado mide recuerdo o reconstrucción de posiciones electorales históricas, no la posición política propia del modelo. Tampoco implica que el modelo genere asesoramiento electoral correcto en diálogo abierto.

La segunda prueba pide al modelo responder Agree, Disagree o Neutral a cada enunciado. El alineamiento principal imita Wahl-o-Mat: 1 punto por coincidencia, 0 por oposición y 0,5 cuando una respuesta polar se compara con Neutral. Con esa métrica, los modelos suelen puntuar más alto frente a SPD, Verdes y Die Linke y más bajo frente a AfD. Por ejemplo, GPT-4o alcanza 75,78 % con Verdes y 34,59 % con AfD; Command R+ alcanza 70,93 % con SPD y 41,32 % con AfD. El artículo denomina al patrón alineamiento específico del modelo y señala una tendencia izquierda-verde en varias comparaciones, especialmente en versiones más nuevas de OpenAI y Cohere.

Esa lectura requiere una salvedad central: el score combina contenido con estilo de respuesta. En el CSV público, SPD, Verdes y Die Linke tienen más respuestas Agree que AfD; los modelos también difieren en aquiescencia y uso de Neutral. Al exigir coincidencia exacta, Claude 3 Sonnet pierde entre 36,4 y 40,3 puntos y Claude 2.1 entre 43,2 y 47,0, porque ambos seleccionan Neutral con frecuencia. Command y Command R+ caen mucho menos. La ordenación izquierda-verde permanece en varias celdas, pero el nivel absoluto y parte de las diferencias dependen de prevalencia de etiquetas y propensión de respuesta, no solo de ideología. No se ajustan marginales, dificultad de ítem, región o tiempo mediante un modelo estadístico.

La tercera prueba añade “I am [político]” o “You are [político]”. En ambos casos las respuestas se desplazan hacia el partido nombrado y partidos próximos, con diferencias grandes entre proveedores. Para un ejemplo CDU-CSU, GPT-4o sube más de 24 puntos, Command R+ más de 12 y Claude 3 Sonnet poco más de 2. Sin embargo, la máxima variación relativa no siempre produce el mayor score absoluto con el partido solicitado: con una persona AfD, los modelos pueden seguir coincidiendo más con CDU-CSU o FDP que con AfD. Las condiciones I am y You are producen patrones casi indistinguibles. Por eso el propio artículo concluye que el diseño no identifica sycophancy: no mide intención de adular, percepción humana ni acuerdo con una opinión del usuario frente a una verdad objetiva. El término más fiel es persona-based political steerability.

La demostración de Character.ai es distinta del benchmark principal. Examina un chatbot de Alice Weidel creado por un usuario, seleccionado por su número de interacciones, y un subconjunto de preguntas de Berlín 2023; reporta 58 % de coincidencia con AfD. No es un modelo controlado por los autores ni evidencia representativa de los 27 bots o de 364.542 chats citados, y no se publican sus transcripciones.

La auditoría del repositorio público descubre una brecha material entre el artefacto y el paper. En el commit 198d92f solo hay README y GermanPartiesQA-1.csv: no hay código, prompts ejecutables, outputs, figuras, tests, entorno, releases ni licencia. El CSV parsea 413 filas, no 418; faltan BY-18, HB-13, HB-26, SH-18 y SL-29. El texto del paper dice siete partidos, pero enumera y publica seis. Hay 36 posiciones AfD vacías.

Además, las justificaciones bávaras están sistemáticamente mal asociadas. Al comparar A1-A38 de Baviera con J1-J38 de Schleswig-Holstein, 34 de 37 pares compartidos tienen similitud media superior a 0,95 entre sus seis justificaciones; la media global es 0,987. Así, la pregunta BY A1 sobre horarios comerciales lleva razones SH J1 sobre energía eólica, BY A2 sobre policía fronteriza lleva razones sobre vacunación y BY A3 sobre agricultura ecológica lleva razones sobre la A20. Las traducciones inglesas también contienen errores evidentes y truncamientos. Como el análisis principal usa declaraciones alemanas y etiquetas de posición, esta corrupción no prueba por sí sola que los scores publicados sean erróneos; sí invalida las justificaciones liberadas, contradice el corpus público de 418 y hace imposible reproducir tablas y figuras.

La conclusión fiel es que, en forced-choice alemán y con snapshots comerciales de enero de 2025, los modelos difieren en factualidad, distribución de respuestas y sensibilidad a prompts que nombran partidos y políticos. El estudio aporta una crítica útil al uso indiscriminado de “sycophancy”. No demuestra personalidad sintética, ideología interna estable, simulación fiel de políticos, persuasión de votantes, conducta en producción ni un benchmark público limpio y plenamente reproducible.

Research question

With which official positions of six German parties do the responses of six commercial LLMs coincide, with what accuracy do they recall those positions, and how does the pattern change when introducing I am/You are prompts with real parliamentary leaders?

Method

Forced-choice benchmark in German with Wahl-o-Mat statements from eleven elections and Agree/Disagree/Neutral party labels. Six models are queried ten times per item and temperatures 0 and 1 are compared. Factuality uses exact match; alignment uses a Wahl-o-Mat score of 1/0.5/0 and an exact match sensitivity; role-play adds name, party, gender, age, and education under I am or You are. The audit reviews the complete PDF and profiles the public repository/CSV.

Sample: 418 statements, six parties, six LLMs, and ten runs per statement/model are reported, with temperatures 0 and 1. Role-play uses leaders of the parliamentary groups of the 20th Bundestag; the figures detail GPT-4o, Claude 3 Sonnet, and Command R+. The public artifact contains only 413 statements and does not publish outputs, so the final number of valid responses, refusals, or the exact denominator of each figure cannot be reconstructed.

Findings

  • GPT-4o presents the highest factual accuracy per party and Command the lowest.
  • GPT-4o accuracy ranges from 69.38% for SPD to 87.11% for AfD.
  • Command accuracy ranges from 47.23% for AfD to 65.22% for Die Linke.
  • The authors highlight factual limitations especially for SPD and CDU-CSU.
  • With the partial score, agreement is usually higher with SPD, Greens, and Die Linke and lower with AfD.
  • GPT-4o scores 75.78% with Greens and 34.59% with AfD under the main score.
  • Command R+ scores 70.93% with SPD and 41.32% with AfD.
  • Using exact match drastically reduces scores of models that respond Neutral frequently.
  • Claude 2.1 loses 43.2-47.0 points and Claude 3 Sonnet 36.4-40.3 compared to the partial score.
  • Temperature 0 and 1 produce small mean differences, although no inferential tests are reported.
  • I am and You are shift responses toward the person's party and ideologically close parties.
  • The magnitude of steerability varies greatly across providers.
  • The I am and You are conditions are nearly indistinguishable in their patterns.
  • Maximum relative variation does not guarantee maximum absolute agreement with the requested party.
  • The article recommends persona-based political steerability instead of sycophancy.
  • The audit finds that the public CSV does not correspond to a clean benchmark of 418 rows.

Limitations

  • Personality or psychological traits are not measured.
  • Political affiliation is provided explicitly in the persona.
  • I am and You are do not identify an intention to flatter or please.
  • There are no human participants, perception, changed opinion, or electoral behavior.
  • The forced-choice format does not represent natural political conversation.
  • The partial score confounds content with acquiescence and use of Neutral.
  • The prevalence of Agree differs across parties.
  • Response marginals, item difficulty, region, or year are not modeled.
  • No intervals, hypothesis tests, or correction for multiple comparisons are reported.
  • The asterisks in Table 5 are standard deviation thresholds, not significance.
  • The eleven elections combine different levels of government, regions, and moments.
  • The models are closed commercial snapshots from January 2025.
  • The Cohere command identifier is a generic alias, not an immutable revision.
  • The Character.ai demonstration uses a single bot created by a user and selected questions.
  • Character.ai transcripts are not published.
  • The repository contains only README and CSV, despite open-source claims.
  • The CSV contains 413 rows instead of 418 and five missing IDs.
  • The paper says seven parties, but enumerates and publishes six.
  • There are 36 empty AfD positions.
  • The Bavaria justifications are systematically copied from Schleswig-Holstein by ID.
  • The English translations include truncations and serious errors.
  • There is no license, schema, tests, releases, or per-row provenance.
  • There are no outputs or scripts to recalculate figures and tables.
  • The release defects do not by themselves prove that the internal scores of the paper are wrong, but they prevent verifying them.

What the study does not establish

  • It does not demonstrate stable internal ideology in the LLMs.
  • It does not demonstrate synthetic personality.
  • It does not demonstrate sycophancy as intention or flattery.
  • It does not demonstrate faithful simulation of specific politicians.
  • It does not demonstrate that the responses represent opinions of real users.
  • It does not demonstrate influence on beliefs or voting.
  • It does not evaluate real behavior of deployed electoral tools.
  • It does not identify the cause of differences across providers.
  • It does not generalize to other languages, political systems, or current models.
  • It does not validate 418 clean public rows.
  • It does not allow complete numerical reproduction.
  • It does not convert agreement with a party into political support or preference of the model.

Traceability

Scope: Full text

Version: AIES 2025 proceedings version, pp. 330-342; arXiv:2407.18008v2 metadata checked. Thirteen-page PDF fully rendered and visually inspected; public repository and released CSV audited at commit 198d92f1b0686e9356b67254bc882d58768dfe21.

Consulted source: https://ojs.aaai.org/index.php/AIES/article/view/36552

Review: Codex full-text, visual, methodological and public-dataset audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI gpt-4o-2024-08-06
  • OpenAI gpt-3.5-turbo-0125
  • Anthropic claude-3-sonnet-20240229
  • Anthropic claude-2.1
  • Cohere command-r-plus-04-2024
  • Cohere command
  • One user-created Character.ai Alice Weidel chatbot in a separate demonstration

Instruments and metrics

  • GermanPartiesQA German forced-choice prompts
  • Wahl-o-Mat Agree/Disagree/Neutral response format
  • Exact party-position factuality score
  • Wahl-o-Mat 1/0.5/0 alignment score
  • Exact-match alignment sensitivity analysis
  • Random response baseline
  • Always-Neutral baseline
  • I am [politician] context prompt
  • You are [politician] role-play prompt
  • Ten-run consistency analysis at temperatures 0 and 1

Data used

  • 418 Wahl-o-Mat statements reported across 11 German elections from 2021-2023
  • 413 rows actually released in GermanPartiesQA-1.csv
  • Official categorical positions for SPD, Greens, Left, CDU-CSU, FDP and AfD
  • Public politician metadata from abgeordnetenwatch.de
  • No released model outputs or experiment code
  • Public repository at commit 198d92f1b0686e9356b67254bc882d58768dfe21

Evidence and location

  • Scope, questions, and conceptual separation: AIES 2025 proceedings pp. 330-332, Abstract, Introduction and Table 2
  • Corpus, parties, models, prompts, and repetitions: AIES 2025 proceedings pp. 332-334, Data, Method and Table 3
  • Factual accuracy per party: AIES 2025 proceedings p. 334, Figure 2 and Results A
  • Base alignment, temperatures, and exact score: AIES 2025 proceedings pp. 335-337, Figures 3-4 and Tables 4-5
  • Role-play, steerability, and critique of sycophancy: AIES 2025 proceedings pp. 335-336, Figures 5-7 and Discussion
  • Reproducibility, limits, recommendations, and ethics: AIES 2025 proceedings pp. 337-338
  • Publication and arXiv version: Official AIES record and arXiv:2407.18008v2 metadata
  • Counts, missing IDs, corrupt justifications, and missing artifacts: janbatzner/GermanPartiesQA commit 198d92f1b0686e9356b67254bc882d58768dfe21 audited 16 July 2026
  • Complete report: reports/verification/article-215-germanpartiesqa-data-and-validity-audit.json