The Hidden Geometry of Meaning: Why AI Cannot Navigate the Human Heart

We often treat language as a code to be cracked. But new research shows it is a geometry to be inhabited — and that is a problem for machines.

In our interpreting classes, we often discuss a specific, hard-to-describe sensation that arises when moving between languages. It is the professional intuition that a word in Language A doesn’t just “mean” a word in Language B — it feels different. It carries a distinct weight and texture that defies simple dictionary equivalence. For years, this has been our hunch — something we sense in the booth but struggle to quantify. However, a study published in Science (Jackson et al., 2019) has finally provided the empirical vocabulary to describe this phenomenon. By analyzing “colexification” — instances where a language uses a single word to describe multiple distinct concepts — the researchers revealed that our mental maps of emotion are far more complex and culturally specific than we previously understood.

The Geometry of Emotion

The study analyzed 2,474 languages to visualize how emotion words “bundle” together, finding a delicate balance between biological universality and cultural variation. While humans share a biological anchor , i.e., words for bad feelings almost never mean the same thing as words for good feelings, the specific associations within those categories vary wildly.

For instance, in Indo-European languages, “love” belongs to a semantic network heavily connected to “like” and “want,” rooting the concept in preference and desire. In contrast, many Austronesian languages frequently colexify “love” with “pity,” reframing the emotion as one of compassionate concern rather than romantic attachment. Similarly, while “anxiety” in Tai-Kadai languages sits close to “fear” as an active threat response, Austroasiatic languages colexify it with “grief” and “regret,” looking backward at loss rather than forward at danger. This proves that interpreting is not merely the swapping of labels, but the navigation of entirely different semantic architectures.

The Probability Trap

This complex geometry of meaning exposes the critical, structural flaw in the “Cascade Model” of AI interpreting currently being pitched to conference organizers. These systems, which chain Automatic Speech Recognition (ASR) to Machine Translation (MT) and Text-to-Speech (TTS), operate on a fundamentally different principle than human cognition: they are probabilistic engines, not semantic architects. When an AI processes a word, it assigns it a value based on statistical probability derived from massive text databases. It is essentially guessing the next piece of a puzzle based on billions of previous puzzles, without any awareness of the broader cultural picture. A human interpreter, by contrast, creates a multi-dimensional map of the speaker’s intent, instantly triangulating a word against the speaker’s origin, tone, and specific cultural background.

AI is guessing the next piece of a puzzle based on billions of previous puzzles, without any awareness of the broader cultural picture.

Photo by Jr Korpa on Unsplash

The Illusion of Fluency

The danger of relying on probability becomes acute when we consider the “Anglocentric” bias of current AI models. Because these systems are trained on internet data dominated by English and other Western languages, they treat Western semantic structures as the universal default. When an AI encounters a concept that does not fit this Western map, it forces a fit, often colonizing the meaning in the process.

Consider a scenario where an English-speaking diplomat addresses an audience speaking an Austronesian language. If the speaker says, “We have great love for the people of this region,” they are likely expressing admiration and strong preference. The AI, blind to the “neighboring” concepts in the target language, will map “love” to the direct dictionary equivalent. However, if that target language colexifies “love” with “pity,” the audience reads a subtitle that essentially says, “We have great pity for the people of this region.” Instead of building a bridge, the machine has unwittingly introduced a tone of condescension.

This is why the “illusion of fluency” provided by AI is so dangerous. The projected subtitles may look grammatically perfect, leading the audience to believe they have understood, when they have actually received a “cultural hallucination.” If an AI translates “anxiety” as “fear” based on English probability, while the speaker’s cultural framework intends “grief,” the audience prepares for a threat instead of empathizing with a loss.

The human interpreter acts as the necessary firewall against this flattening of meaning. We do not just translate statistics; we translate experience. We know when to suppress the dictionary definition to preserve the emotional truth, ensuring that we don’t build a digital Tower of Babel where everyone speaks the same words, but no one understands the same feelings.