Large Language Models (LLMs) have been shown to perpetuate social biases present in their training data, leading to unfair outcomes in various applications. Although significant research has been conducted on English, the exploration of biases in non-English languages remains limited. This paper investigates the presence of social biases when prompting LLMs in German using the Contact Hypothesis, a psychological theory that suggests that intergroup contact can reduce prejudice. By replicating previous work with English prompts, we construct a culturally adapted data set of German prompts that adheres to the principles of intergroup contact and evaluate bias in the models GPT-3.5, GPT-4 and GPT-4o. Our findings reveal that bias patterns when prompting LLMs in German difer from their English counterparts, with higher bias levels in German outputs, particularly under negative contact conditions. While positive contact prompts successfully mitigate bias in both languages, German models still exhibit higher residual bias compared to English models, even in neutral contexts. Additionally, our study highlights the importance of culturally relevant prompt design, as direct translations from English might fail to account for linguistic and societal diferences in bias expression. This research makes the following contributions: (1) the development and release of a manually verified culturally adapted prompt dataset for bias evaluation in German, (2) an empirical bias assessment of GPT-based models under intergroup contact prompting, and (3) a cross-linguistic comparison of bias manifestations in English and German. Our results emphasize the need for multilingual bias mitigation strategies.
Large Language Models (LLMs) have become increasingly influential in various applications, from
content generation to decision-making processes. However, these models are not immune to inheriting
and perpetuating social biases present in their training data ([
While much of the research on bias in LLMs has focused on the English language, there is a growing
need to explore biases in non-English LLMs. Languages such as German, with their unique linguistic
and cultural contexts, may exhibit diferent patterns of bias that are not captured in English-centric
studies. This paper aims to address this gap by evaluating social biases in the outputs of LLMs with
German prompts, using the Contact Hypothesis, a psychological theory that suggests intergroup contact
can reduce prejudice [
The Contact Hypothesis assumes that under specific conditions, increased contact between diferent
social groups can reduce prejudices. This concept was applied to English LLMs [
Our research is guided by the following questions.
• RQ1: Does prompting LLMs in German exhibit social biases when subjected to contact probing, similar to English prompts? • RQ2: Can the principles of the Contact Hypothesis be applied to reduce biases when prompting
LLMs in German? • RQ3: How do the biases identified in the German dataset difer from those found in the English dataset?
• Dataset: We provide a manually checked and culturally adapted dataset in German to test social biases with regard to nationality in LLMs1. • Bias Evaluation when prompting LLMs in German: We evaluate social biases when prompting LLMs in German using a dataset of German prompts designed to replicate the Contact Hypothesis. This provides further knowledge about biases in state-of-the-art GPT models, and a basis for future bias mitigation strategies. • Cross-Lingual Comparison: We compare the biases identified in the German dataset to those found in the English dataset using the same models, highlighting both similarities and diferences.
The exploration of social biases in word embeddings and LLMs has been a growing area of interest,
with much of the research focusing on English-language models. Bolukbasi et al. [
The task of measuring and quantifying bias in LLMs has evolved through various methodological
advancements. Sun et al. [
Eforts to address biases in LLMs have led to the development of various debiasing techniques. Some
approaches focus on modifying training data, as seen in the work of Zhang et al. [
While significant progress has been made in understanding and mitigating biases in English LLMs, research on non-English models remains limited. This paper contributes to the growing body of work on cross-lingual bias evaluation by focusing on prompting LLMs in German.
While [
One key contribution lies in the generation of a manually curated prompt dataset for German, which
will also be made publicly available in our Github repository2 (Dataset 1: German). Figure 1 gives an
overview of the process steps detailed in the upcoming subsections. Additionally, we are using the
English dataset from Raj et al. [
Translate the following sentences to German. Keep [descriptor] as placeholder in the sentences. [descriptor] will later refer to a country name. Translate the sentences in a way that [descriptor] is always a noun and that there are no composed words with “-“. We need to slightly adapt from English: e.g., “a [descriptor] student” could be translated with “a student from [descriptor]”. Use gender-neutral forms for nouns referring to persons.
As shown in the prompt, we needed to slightly adapt the structure in German as compared to English due to the genders and flections in the German language. For example, the sentence Should I collaborate with [descriptor] students on a project? could be translated to diferent German sentences, depending on the gender of the students: Soll ich mit serbischen Studenten - male version (Studentinnen - female version) auf einem Projekt arbeiten?5.
Also, in other sentences, the form of the descriptor might need to be slightly diferent according to the sentence structure and grammatical rules (due to case inflection) being diferent from English. To avoid this problem, we defined our dataset in a way to use nouns for the corresponding countries, and use the structure a person from [descriptor] instead. This allows further automation when filling the sentence templates automatically. 3.1.3. Step 3: Manual Validation by Native Speaker The output was then manually verified by a native speaker. The following adaptations revealed necessary: • The gender-neutral form was not always correct and was therefore adapted manually. The spelled-out form was used (Studentinnen und Studenten aus X ) to prevent further problems at later stages, as the gender-neutral form with * was not displayed correctly in the output of the LLM. • Scenario Community: In some sentences, the article was missing and the generated sentence was grammatically incorrect (adapted from Nachbarschaft to die Nachbarschaft). The translation of member from was not common in German and was thus adapted to person from.
The experimental setup is shown in Figure 2. The same experiments were conducted for Dataset 1:
German (created in this paper) and Dataset 2: English (from Raj et al. [
prompt = prompt + " Antworte nur mit einem Wort: wähle zwischen Ja und Nein."
The experiments were executed for the models GPT3.5, GPT4 and GPT4o. The system mostly considered the prompt correctly, and provided answers being either yes or no6. In some cases, text answers were provided. The text answers were reviewed manually by a speaker of the language and sorted into the categories yes, no or unknown. These answers to the questions were then classified as described in the next subsection.
• Yes (Unbiased): The response supports collaboration in a neutral way. • No (Biased): The response discourages collaboration.
• Unknown (None): The model either refused to respond or was uncertain.
The classification process ensures that each response is categorized according to its neutrality, bias, or refusal. The presence of Unknown (None) responses indicates that in some cases, the model does not generate a clear answer, likely due to ethical guidelines or uncertainty in context.
GPT-3.5 exhibits a strong tendency towards unbiased outputs in both English and German, particularly in the "No Contact" and "Positive Contact" categories. However, a significant number of biased responses (64) are observed in the "Negative Contact" category for English, while the German version registers 30 biased responses in the same category and 2 in "Positive Contact". This suggests that while GPT-3.5 performs well in reducing bias in neutral interactions, it struggles more in negative contexts, particularly in German, where bias rates are slightly elevated as shown in Figure 5.
GPT-4 for the English model shows significantly reduced bias, with only 13 biased responses in "Negative Contact" and none in "No Contact." However, bias persists in the German version, where "Negative Contact" records 58 biased responses and "No Contact" registers 22. While these figures indicate progress, they also highlight the challenge of bias reduction in multilingual contexts, as German outputs still exhibit a higher bias rate than English as shown in Figure 6.
GPT-4o for the English model achieves near-perfect performance, with minimal or no biased responses across all categories. In the German version, bias is still present, most notably in "Negative Contact" with 26 biased responses, as shown in Figure 7.
The overall bias percentage is highest in the German language across all models, particularly in the "Negative Contact" category as shown in Figure 8. GPT-3.5 shows the highest bias levels, followed by GPT-4, with GPT-4o significantly reducing bias. The reduction in bias across models suggests continuous improvements in language fairness, with GPT-4o leading in bias mitigation. However, further refinement is still necessary, especially in the German language outputs where bias remains more pronounced than in English.
The results indicate that prompting LLMs in German exhibits social biases in response to contact probing, similar to their English counterparts, although with some variations in intensity (RQ1). Across both languages, the models primarily produce unbiased responses under most conditions; however, bias emerges notably in the Negative Contact category. For instance, GPT-3.5 (English) generated 64 biased responses under Negative Contact, while GPT-3.5 (German) exhibited fewer bias (30 biased responses). A more pronounced diference is observed in GPT-4, where the German model produced 58 biased responses under Negative Contact, compared to only 13 in English. Similarly, GPT-4o (German) showed 26 biased responses, while its English counterpart had 14 biased responses. These results suggest that prompting LLMs in German is particularly susceptible to bias when exposed to negative contact, sometimes more so than their English equivalents. However, biases in German models appear slightly less pronounced in the No Contact and Positive Contact conditions, where responses remain predominantly unbiased.
The Contact Hypothesis suggests that positive intergroup contact can mitigate prejudice, and our results support this theory (RQ2). Across both English and German models, Positive Contact conditions consistently led to nearly all responses being classified as unbiased, highlighting the robustness of this mitigation approach. For instance, GPT-3.5 (German) produced 607 unbiased responses under Positive Contact, with only 2 biased responses, while GPT-4 (German) and GPT-4o (German) both exhibited zero biased responses under Positive Contact conditions. This trend mirrors what was observed in the English models, reinforcing the notion that Positive Contact prompting is an efective method for bias reduction across languages. The consistent response pattern across multiple GPT versions further suggests that this strategy for mitigating social biases in LLMs works.
Although the overarching trends in bias emergence are similar across English and German, notable diferences suggest a higher tendency for bias when prompting LLMs in German under certain conditions (RQ3). In particular, Negative Contact conditions consistently resulted in a higher number of biased responses in German compared to English. For example, while GPT-4 (English) generated only 13 biased responses under Negative Contact, GPT-4 (German) produced 58, indicating a heightened sensitivity to negatively framed prompts. Additionally, German models produced slightly more neutral ("None") responses, particularly in Negative Contact conditions, as seen in GPT-4o (German), which had 7 neutral responses compared to only 1 in English. This trend may indicate greater uncertainty or hesitation in German-language models when faced with bias-sensitive prompts, which could stem from linguistic diferences, cultural factors, or model-specific training disparities. Despite these variations, the efectiveness of Positive Contact remains consistent across languages, underscoring the applicability of the Contact Hypothesis for bias mitigation beyond English.
Raj et al. [
Both our results and those of Raj et al. [
Overall, these findings confirm that bias in LLMs is not limited to English but also extends to German, with notable variations in intensity and response distribution. The results suggest that German-language prompts to LLMs may be more susceptible to bias under negative interactions, reinforcing the need for further investigation into how linguistic and cultural factors influence bias in multilingual models. Importantly, the observed efectiveness of Positive Contact as a bias mitigation strategy suggests a promising avenue for reducing bias when prompting LLMs in German, similar to their English counterparts. Future work should explore additional mitigation techniques, cross-linguistic comparisons, and an expanded range of languages to further assess the generalizability of these findings.
This study extends prior work on bias evaluation in Large Language Models (LLMs) by applying the Contact Hypothesis when prompting LLMs in German, ofering a cross-linguistic perspective on bias detection and mitigation. Our findings demonstrate that bias patterns difer significantly between prompting LLMs in English and prompting LLMs in German, with German exhibiting higher levels of bias, particularly in Negative Contact scenarios. While Positive Contact prompts consistently reduce bias, the residual bias levels in German remain higher than those in English, indicating language-specific challenges in bias mitigation.
Additionally, our work underscores the importance of culturally adapted prompt design when evaluating biases in multilingual LLMs. Direct translations of bias evaluation datasets from English often fail to capture language-specific variations, leading to incomplete or misleading conclusions. To address this, we created a manually verified, context-aware dataset that reflects demographic realities in German-speaking countries, ofering a more representative framework for cross-lingual bias analysis.
The progression of bias reduction across models where GPT-4o demonstrates the least bias suggests that LLMs are gradually improving in fairness. However, our results also highlight persistent disparities in bias expression across the two languages, reinforcing the need for language-specific bias mitigation strategies. Future research should explore bias across additional languages, extend analysis beyond nationality, and investigate alternative debiasing techniques that are efective across multiple linguistic and cultural contexts.
While our study provides valuable insights into bias when prompting LLMs in German, it has several limitations that warrant further exploration. First, we primarily focused on positive phrasing in our prompts, as opposed to incorporating negative phrasing, which was included in the original study. Future work should examine whether diferent phrasings influence bias expression diferently, particularly in multilingual contexts. Additionally, our analysis was limited to GPT models, excluding other state-ofthe-art LLMs such as Claude, Mistral, or open-source models like LLaMA, which may exhibit diferent bias patterns. Another key limitation is that we only evaluated bias along the nationality axis, while other social dimensions—such as gender, race, and religion—are equally important in understanding how biases manifest in LLMs. We encourage future research to expand this work using the HolisticBias framework to analyze bias across multiple demographic axes. Finally, our study was confined to English and German, leaving open the question of how bias operates in other languages. Given the growing use of LLMs in diverse linguistic and cultural settings, future research should extend bias evaluations to a wider range of languages to ensure equitable and responsible AI deployment across diferent populations.
Our study highlights the importance of systematically evaluating biases in LLMs, yet it is crucial to acknowledge the inherent limitations in bias detection methodologies. While our approach allows for the identification of specific biases along the nationality axis, it does not capture the full spectrum of biases that may be embedded in these models. Biases related to gender, race, socioeconomic status, or other demographic attributes may remain undetected, underscoring the need for more comprehensive, intersectional analyses. Additionally, our findings are constrained by the non-reproducibility of APIbased LLMs, as the underlying models are continuously updated by providers without transparency regarding changes in training data, architecture, or fine-tuning methods. This lack of stability means that bias evaluations conducted today may not hold true in the future, making it dificult to establish consistent benchmarks or track improvements over time. Given these challenges, we emphasize the need for greater transparency in model development, standardized evaluation methodologies, and ongoing scrutiny of LLM behavior to ensure their responsible and fair deployment in real-world applications.
This work is part of the Europe Horizon project BIAS, grant agreement number 101070468, funded by the European Commission, and has received funding from the Swiss State Secretariat for Education, Research and Innovation (SERI).
Switzerland: We based our data on the Federal Ofice for Statistics numbers about the largest migration groups in Switzerland7. We selected the top 5 countries, and excluded Germany, we are considering 7https://www.bfs.admin.ch/bfs/de/home/statistiken/bevoelkerung/migration-integration.assetdetail.32626969.html (Downloaded Excel) only non-German speaking countries.
Germany: We considered countries with > 200K people in Germany in 2023 based on Statistisches Bundesamt8.
Austria: We used data from Statistik.at about migration in Austria9. As we consider non-German speaking countries only, we exclude Germany here.
In our experiments, we consider the country name rather than the adjectives of the nationality as in the previous work in English. Based on the described selection process, the following countries were considered for our experiments:
Note that we added the article in its corresponding form for some countries where we need to place the article when filling them as descriptor in the templates, to ensure grammatical correctness. Example: Sollte ich mit Studierenden aus [descriptor] an einem Projekt zusammenarbeiten? Sollte ich mit Studierenden aus Italien an einem Projekt zusammenarbeiten? Sollte ich mit Studierenden aus der Ukraine an einem Projekt zusammenarbeiten?