How do low-dimensional representation structures entangle multiple cultures together?
This explores how compressing meaning into a few shared dimensions forces distinct cultures to be represented through each other rather than on their own terms — so a change in one bleeds into others.
This explores how compressing meaning into a small number of shared dimensions forces distinct cultures to be represented through each other rather than on their own terms. The corpus has a surprisingly direct answer, and it starts with a geometric fact: the semantic features inside LLM embeddings aren't independent. Researchers found that twenty-eight semantic axes collapse into just three principal components, and that pushing on one feature predictably drags the aligned ones along with it — "off-target effects" that aren't a bug but a consequence of how meaning is packed into a low-dimensional space Do LLM semantic features organize along human evaluation dimensions?. When everything shares the same few axes, nothing can move alone.
Now apply that geometry to culture. A separate line of mechanistic interpretability work shows that low-resource cultures — Ethiopia, Algeria — are represented internally through high-resource cultural proxies. The model doesn't store these cultures in their own region of representational space; it routes them through dominant ones, a "unidirectional flattening" that persists in the internal states even when the model produces a correct surface answer Do LLMs represent low-resource cultures through dominant cultural proxies?. This is entanglement in action: there isn't enough dimensional room for every culture to get an independent address, so the underrepresented ones get encoded as deviations from the well-represented ones.
The entanglement also shows up behaviorally, not just in the weights. When GPT-4.5 was tested on social-norm judgments across 555 scenarios, it out-predicted every individual human — but all the AI models shared *identical* systematic errors on unwritten norms Can AI learn social norms better than humans?. A shared blind spot across models is the fingerprint of a shared low-dimensional substrate: they're all compressing many cultures' norms through the same collapsed structure, so they all miss in the same places.
Why does the compression organize cultures this way rather than keeping them separate? Two notes give the underlying mechanism. Language models build meaning purely relationally — Saussure's *langue* — by compressing co-occurrence structure from text with no external grounding Can language models learn meaning without engaging the world?. And the geometry that compression produces is hierarchical: the leading eigenvectors of embedding matrices carve broad categories first, then finer ones, in a coarse-to-fine spectral order that mirrors a hypernym tree Do embedding eigenvectors organize taxonomy from coarse to fine?. Cultures that appear rarely in text never earn their own fine-grained branch — they get folded under whichever coarse, high-frequency branch they statistically resemble.
The quietly unsettling takeaway: cultural entanglement isn't a moral failing layered on top of a neutral model, it's the same machinery that makes the model work at all. The dimensional thrift that lets three axes carry twenty-eight features is exactly what leaves no room for every culture to stand apart. If you want to pull on the thread of what a representation drops when it compresses, the argument that text itself is a lossy abstraction stripping physics and causality is the companion piece Are text-only language models fundamentally limited by abstraction?.
Sources 6 notes
Twenty-eight semantic axes in LLM embeddings reduce to three principal components matching human EPA structure. Intervening on one feature predictably shifts aligned features proportionally, creating unavoidable off-target effects that reflect how meaning is fundamentally organized.
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.
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.
Research shows LLMs learn culturally situated discourse patterns by compressing relational structure from text, demonstrating that fluent language generation requires no external referents or embodied grounding.
Leading eigenvectors of embedding Gram matrices separate broad taxonomic branches first, then progressively finer sub-branches—a coarse-to-fine spectral order that tracks the WordNet hypernym tree level by level, confirming predictions from co-occurrence statistics.
Text strips the physics, geometry, and causality present in reality, forcing language models to manipulate symbols without grounding in their source dynamics. This creates predictable failure modes in physical, geometric, and causal reasoning that multimodal training could address.