Does an AI 'understand' Ethiopia on its own terms, or only by comparing it to cultures it has seen far more of online?
Do language models systematically underrepresent non English knowledge about local topics?
This explores whether language models know less, or represent things less faithfully, when the knowledge comes from non-English, local sources: places, cultures and communities that are thinly covered in English text.
This explores whether language models systematically shortchange local, non-English knowledge. The corpus doesn't have a study that directly counts which local facts models know in one language but not another. What it does have is evidence that the problem is real and runs deeper than missing facts. The clearest example comes from interpretability work, which looks inside a model at its internal representations rather than only at its outputs. When models handle lower-resource cultures such as Ethiopia or Algeria, their internal states route those cultures through better-documented proxies. The model in effect thinks about Ethiopia by way of the cultures it has seen more of. This holds even when the model's answer looks correct on the surface Do LLMs represent low-resource cultures through dominant cultural proxies?. So the underrepresentation isn't only about gaps. It is also about local knowledge being quietly translated into a dominant frame.
That matters because you can't simply prompt your way out of it. One line of work shows that prompting can only bring out knowledge a model already has. It cannot add knowledge that was never in the training data Can prompt optimization teach models knowledge they lack?. A related finding shows that even when you put the correct information straight into the prompt, strong associations from training can override it Why do language models ignore information in their context?. Together these suggest a double bind. If local knowledge is thin in training, the model lacks it. If you supply it, the model's English-heavy defaults may still pull the answer back toward the familiar version.
The less obvious angle is that underrepresentation also shows up in who is asking, not just in what is being asked about. When users signal that they are not native English speakers or not from the US, major models give less accurate answers and refuse more often. They do this even on questions they demonstrably know the answers to. That points to how the models were tuned rather than to what they are able to do Do language models treat less-educated users worse?. The people most likely to ask about local, non-English topics may be the ones getting the weakest service.
There's a useful counterpoint on culture. GPT-4.5 judged social appropriateness more accurately than any individual human across hundreds of scenarios. However, all the models shared the same blind spots on unwritten norms, which is exactly the kind of knowledge that lives in local practice rather than in text Can AI learn social norms better than humans?. Scale that up and you get the broader worry. Models reflect a skewed slice of human experience, and because millions of people rely on the same few models, that skew can spread into how people write and think Do large language models narrow human expression and thought?. The real risk isn't just that models miss local knowledge. It's that they may slowly crowd it out.
Sources 6 notes
Mechanistic interpretability analysis reveals that low-resource cultures like Ethiopia and Algeria are structurally represented through high-resource cultural proxies in internal model states, not just output. This architectural bias persists even when models can produce correct surface-level answers.
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Evaluation of three major LLMs found significant accuracy drops and higher refusal rates when user bios signaled less education, non-native English, or non-US origin, with compounded effects at intersections. Models knew correct answers but withheld them selectively, suggesting RLHF misalignment rather than capability gaps.
GPT-4.5 outperformed every individual human at judging social appropriateness across 555 scenarios, challenging the theory that embodied cultural experience is necessary. However, all AI models share identical systematic errors on unwritten norms.
Show all 6 sources
LLMs mirror skewed slices of human experience shaped by training data regularities, and widespread reliance on identical models amplifies convergence. Co-writing studies show users unconsciously adopt model stances and framings.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- On Epistemic Diversity in Large Language Models
- The Widespread Adoption of Large Language Model-Assisted Writing Across Society
- The Homogenizing Effect of Large Language Models on Human Expression and Thought
- Six misconceptions about large language models: A minimal model and diagnostic taxonomy
- Learning To Retrieve Prompts for In-Context Learning
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms
- Entangled in Representations: Mechanistic Investigation of Cultural Biases in Large Language Models
- LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users