Presumed Cultural Identity: How Names Shape LLM Responses

🔼 This figure illustrates the experimental setup used in the study to analyze cultural presumptions in Large Language Model (LLM) responses. The process begins with selecting names and associated cultures from a dataset, which are then paired with information-seeking questions. These prompts, including names, are fed to four different LLMs, generating responses. The responses are then evaluated using two methods: an LLM-as-a-judge approach, and a method comparing responses to pre-defined cultural assertions. The outcome assesses the extent of cultural presumptions made by the LLMs based on the names provided in the prompts.

read the caption

Figure 2: Experimental Setup

🔼 This figure displays the average default bias across different LLMs and cultural facets (food, clothing, tradition & rituals). Default bias refers to the inherent cultural biases present in LLM responses even without any user name provided in the prompt. Higher values indicate a stronger predisposition towards certain cultures in the generated text.

read the caption

Figure 3: Default Bias values averaged over Models and Facets. For details refer to subsection 3.8.

🔼 This figure displays the cultural bias exhibited by four different large language models (LLMs) when presented with names from various cultural backgrounds. The bias is calculated as the difference between the bias observed with names and the default bias without names, as described in section 3.8 of the paper. Each bar represents a country, and its height indicates the degree of cultural bias for that country across all evaluated models and aspects (food, clothing, and rituals). Positive values suggest that the model displays a stronger bias towards associating the given names with the respective culture than would be expected based on default bias. Conversely, negative values indicate a weaker association than expected.

read the caption

Figure 4: Bias across models above the default bais. For calculation of bais refer to section 3.8

🔼 This box plot visualizes the cultural bias in Large Language Model (LLM) responses, broken down by facet (clothing, food, rituals & traditions). For each facet, the plot shows the average bias for different countries across four different LLMs. The bias score represents how often the model’s responses reflect culturally specific aspects of the named country, compared to responses where no country is explicitly mentioned. Higher values indicate stronger cultural bias towards that country.

read the caption

Figure 5: Box plot showing comparison of bias for countries values (averaged over 4 models) for each facet.

🔼 This heatmap visualizes the cross-cultural bias in Large Language Model (LLM) responses when names from different countries are used in prompts. Each cell in the heatmap represents the bias score (calculated as the difference between bias with and without names; see Section 3.8) for a given pair of countries. The X-axis shows the country being investigated for bias in the LLM’s response, while the Y-axis shows the country of origin for the name used in the prompt. Warmer colors indicate stronger biases, meaning the LLM’s response shows a stronger association with a specific culture when a name from that culture is present in the prompt. Cooler colors show weaker biases or even negative biases, suggesting the name may not strongly trigger cultural assumptions.

read the caption

Figure 6: Cross-cultural bias heatmap for bias values over the default (3.8). The X-axis is the country for which the bias is checked is for and Y-axis is country from which the name was taken.

🔼 This figure shows the distribution of the number of biased responses generated by LLMs for each of the names used in the study. The x-axis represents the individual names used in the prompts (omitted for clarity), and the y-axis shows the count of biased responses associated with each name. The distribution is heavily skewed, indicating that a relatively small number of names account for a large proportion of the biased responses while the majority of the names have very few or no biased responses.

read the caption

Figure 7: Distribution of biased responses per name [Names are omitted from the x-axis to avoid clutter]

🔼 This figure shows the percentage of biased responses containing each specific word, relative to the total number of biased responses across all words. It helps to understand which words contribute most strongly to the cultural bias observed in the model’s responses when names are included in prompts.

read the caption

Figure 8: Percentage contribution of each word’s biased responses relative to the overall number of biased responses

🔼 This figure displays a bar chart illustrating the degree to which OpenAI’s GPT-40-mini model exhibits cultural biases when names are included in prompts, compared to responses without names (default bias). The chart showcases the difference in bias for various countries, indicating the extent to which the model associates specific names with particular cultures, thereby demonstrating the cultural presumptions embedded within the model’s responses.

read the caption

Figure 9: OpenAI GPT-4o-mini name bias over the default responses

🔼 This figure presents a comparison of default bias across four different large language models (LLMs). Default bias refers to the inherent biases exhibited by LLMs in their responses even without any user-provided names or identifying information. Each bar in the chart represents a country, and the height of the bar shows the average default bias score for that country across various facets (food, clothing, tradition) considered in the study. The figure helps to visualize how different LLMs display varying levels of default bias towards different cultures, providing insights into the underlying biases present in the models’ training data.

read the caption

Figure 10: Default Bias across models, for calculation and discussion about default bias refer to section 3.8

🔼 This table lists the large language models (LLMs) used in the study, along with their respective model names and the Hugging Face repository links where these models can be accessed. This information is crucial for reproducibility, allowing others to utilize the same models and potentially replicate the experiments and analyses described in the paper.

read the caption

Table 2: Models used in this study and their corresponding HuggingFace repository code

🔼 This table lists names that are common across multiple cultures, along with the percentage of times each name is associated with different countries. It highlights the potential for ambiguity and bias when using names as a sole indicator of cultural identity in large language models.

read the caption

Table 3: Multicultual Names

🔼 This table lists the names that elicited biased responses from the language models, categorized by country of origin. The number in parentheses after each name indicates how many times that name triggered a culturally biased response from the language models. This provides insight into which names are most strongly associated with specific cultural assumptions within the models.

read the caption

Table 4: Biased Names for All Countries (Names with number of biased responses in parenthesis)