TLDR: This research investigates how Large Language Models (LLMs) internally represent religion, and how these representations intersect with concepts of violence and geography. Using Sparse Autoencoders (SAEs) and the Neuronpedia API, the study analyzed latent feature activations across five models. It found that while all five major religions (Christianity, Islam, Judaism, Hinduism, Buddhism) show comparable internal cohesion, Islam is consistently more linked to features associated with violent language. Geographic associations largely reflect real-world demographics but also a Western-centric view. The findings underscore that LLMs embed both factual distributions and cultural stereotypes, highlighting the need for structural analysis beyond just model outputs to audit internal biases.
Large Language Models (LLMs) have become integral to many aspects of our digital lives, but their widespread use has brought increasing scrutiny to the biases they might embed. While much research has focused on biases related to gender and race, the internal representation of religious identity within these powerful AI systems has remained largely underexplored. A recent study delves into this critical area, examining how LLMs perceive religion and its connections to violence and geography.
The research, titled Mechanistic Interpretability with SAEs: Probing Religion, Violence, and Geography in Large Language Models, was conducted by Katharina Simbeck and Mariam Mahran. Their work utilizes a sophisticated approach called mechanistic interpretability, specifically employing Sparse Autoencoders (SAEs) via the Neuronpedia API, to peer into the ‘minds’ of LLMs. Instead of just observing what LLMs say, this method allows researchers to analyze the latent feature activations – the internal signals – that shape a model’s understanding.
Unpacking Internal Cohesion and Stereotypes
The study set out to answer four key questions: how consistently LLMs encode each religion as a distinct concept (RQ1), the extent to which religious identities are associated with violence (RQ2), how LLMs encode geographic patterns of religion (RQ3), and how these associations vary across different model architectures (RQ4).
To address these, the researchers analyzed latent feature activations across five different LLMs, including variations of GPT2-small, Gemma-2, and Llama3.1-8B. They used carefully crafted prompts related to five major world religions (Christianity, Islam, Judaism, Hinduism, and Buddhism) and a separate set of prompts related to violence and criminality.
The findings revealed that all five religions showed comparable internal cohesion within the models. This means that LLMs tend to treat each religion as a distinct and coherent concept, rather than a diffuse collection of ideas. For instance, in GPT2-small, Buddhism and Hinduism shared a similar number of features across their respective prompts, indicating a consistent internal representation.
The Link Between Religion and Violence
However, this internal consistency doesn’t equate to neutrality. When examining the overlap between religion-related features and violence-related features, a significant asymmetry emerged. Islam consistently registered the highest Violence Association Index (VAI) across all five models. The VAI normalizes raw overlap values, making comparisons between models meaningful. A VAI above 100 indicates a stronger-than-average association with violence-related features within that model. For example, in Gemma-2-2b, Islam scored 117, while other religions ranged from 94 to 96, indicating a clear skew.
Further semantic analysis of activation texts – the actual phrases that highly activate specific features – reinforced these findings. By searching for crime-related keywords like “terrorism,” “extremist,” and “violence,” the study found that Islam consistently had the highest proportion of such terms in most models. While there were exceptions, such as Hinduism showing higher rates in GPT2-small and Llama3.1-8B, the overall pattern suggested a concerning link between Islam and violent language within the models’ internal structures.
Geographic Footprints of Faith
The geographic analysis provided another layer of insight. By scanning activation texts for keywords representing various global regions, the study observed how LLMs associate religions with different parts of the world. Europe and North America were the most frequently mentioned regions, with relatively balanced associations across all five religions. Asia and the Middle East also showed strong representation, with Hinduism and Buddhism dominating the Asian context, and Islam being most prominent in the Middle East. These patterns largely reflect real-world religious demographics.
However, the analysis also highlighted a Western-focused lens in the models, with Europe and North America strongly represented across all religions, while regions like Australia and South America were largely absent. Judaism, by comparison, had a markedly narrower geographic distribution, while Islam showed a broader global spread. This suggests that LLMs’ internal representations mirror cultural salience and media visibility more than strict statistical reality.
Also Read:
- Enhancing Multimodal AI Understanding by Tackling Superficial Biases
- CultureScope: A Deeper Look into AI’s Cultural Competence
Model Differences and Future Implications
The study also underscored that these associations are not universal across all LLMs. Differences were observed across model architectures and training datasets, indicating that bias is shaped not just by the data they consume but also by their scale and structure. Smaller models, like GPT2-small, sometimes revealed noisier and more exaggerated associations, while larger models, such as Gemma-2-9b, encoded more compact and abstract representations.
This research highlights the critical value of structural analysis in auditing LLMs. It moves beyond simply evaluating model outputs to uncover the internal conceptual structures that truly shape model behavior. The findings suggest that LLMs reliably abstract religion into stable latent categories, but these categories can inadvertently embed cultural stereotypes and societal narratives, particularly concerning violence and geographic associations. Understanding these internal biases is crucial for developing AI systems that are fair, respectful, and accurately represent diverse identities.
