The Multilingual Illusion

Modern multilingual language models do something remarkable. They map entirely different languages into a shared vector space, creating a geometric landscape where words for numbers, animals, and abstract concepts form strikingly similar constellations regardless of the tongue used to express them. This shared geometry is what allows a model to learn a concept in English and instantly apply it in Swahili.

Yet this shared mental map does not mean language is merely a transparent translation layer. When you change the language of a prompt, you change how the model reasons. Performance shifts in unpredictable hierarchies across linguistic barriers, revealing that the choice of words alters the underlying computational path.

Even more peculiar is what happens when these models encounter cognitive traps. LLMs exhibit human-like cognitive biases, falling prey to phenomena like the illusory truth effect, where frequent repetition breeds high confidence in falsehoods. They mimic human error patterns in heuristic reasoning tasks, stumbling precisely where humans stumble, even as they easily outperform us on rigid, rule-based deductive logic.

This creates a strange parallel. We often think of machine intelligence as alien and purely mechanical, liberated from the sloppy shortcuts of human thought. But by forcing diverse human languages through a single mathematical geometry, these models do not strip away our cognitive flaws. Instead, they learn them.