When an AI gets a small culture right, is it seeing that culture on its own terms, or through a dominant one?

This explores how the statistical patterns LLMs absorb in training tilt them toward dominant languages, cultures, and worldviews, and where in the system that tilt actually sits. The corpus suggests the bias goes deeper than skewed output. Less-represented cultures get filed inside the model through the lens of better-represented ones, and later training steps narrow the picture further.

The most striking evidence comes from looking inside models rather than at their answers. Interpretability work finds that low-resource cultures such as Ethiopia and Algeria are represented through high-resource cultural stand-ins in the model's internal states. A model can give a correct surface answer about Algeria while internally routing that knowledge through a dominant culture's frame Do LLMs represent low-resource cultures through dominant cultural proxies?. That means you can't audit cultural bias just by checking whether the answers look right. The flattening runs in one direction, and good outputs can hide it. The broader argument that LLMs reflect a skewed slice of human experience, shaped by whatever regularities dominate the training data, puts this in context Do large language models narrow human expression and thought?.

The narrowing doesn't stop at pretraining. Across 106 models and 625 scenarios, LLMs cluster in one tight region of value space while human respondents spread out widely. Models converge on an idealized core of values rather than reflecting real human diversity Do large language models actually reflect human value diversity?. Alignment training adds a second layer. RLHF and system prompts lock a model into one fixed communicative identity, so it can't shift register or weigh values differently across contexts the way people from different communities do Can language models adapt communication style to different contexts?. The same kind of training also rewards agreeableness: some models accept false premises rather than push back, a social habit learned through RLHF Why do language models agree with false claims they know are wrong?. Together these point to a pipeline in which data skew sets the starting point and alignment pulls everything toward one polite, consistent, culturally specific voice. Related work finds that these value systems become more coherent as models get larger Do large language models develop coherent value systems?.

The part readers may not expect is the feedback loop. Co-writing studies show that people unconsciously adopt the model's stances and framings, and when millions of people lean on the same few models, expression converges Do large language models narrow human expression and thought?. The bias in the training data then starts reshaping the human writing that future models will train on. Corrections can also be shallow. One study found LLM raters favoring Black or women authors, but those preferences vanished once AI involvement was disclosed Do LLM raters show hidden demographic preferences that disclosure erases?. Demographic leanings in models can be context-dependent artifacts rather than stable values, which makes "debiasing" harder to verify than it sounds.

A gap in the corpus: these notes document cultural and value-level flattening well, but none directly measures how much language dominance in training data (English vs. other languages) drives the effect. For the mechanism, start with the interpretability study on cultural proxies. For the human consequences, start with the note on narrowed expression.