Same Intent, Different Language
Does AI’s interpretation of your words match what you mean?
An AI model may process a request through representations associated with another language, even when its answer is written in the user’s language. [3, 4] The language of the answer alone therefore does not establish that the request was interpreted as intended.
Translation equivalents can refer to the same action or idea while carrying different associations and implications. [1, 2] A model could recognize the subject of a request correctly while drawing a different conclusion about what the user wants. The response could then follow that interpretation, even though the user’s intention has not changed.
Why Equivalent Words Can Lead to Different Interpretations
A word’s associations reflect the contexts in which it is used: the ideas it appears alongside, the situations it describes, and the evaluations it tends to convey. These associations can differ between translation equivalents, even when both words refer to the same action, object, or idea.
For example, the English word “impressed” has been rated more positively than its Italian translation equivalent, “impressionato.” The difference concerns the emotional evaluation carried by the words, which is part of their meaning beyond the reaction they describe. Cross-language comparisons have found variation in these semantic relationships even among common words, with closer alignment in some domains than in others. [1]
The meaning of an expression also depends on what the speaker is doing with it in context. “Can you reach the top shelf?” can ask about someone’s ability or serve as an indirect request to retrieve something. The listener uses the situation and familiar conventions of speech to infer which function the speaker intends.
Those conventions help determine what is understood without being stated explicitly. When they differ between languages, a translation may introduce an implication that was absent from the original or lose one that the original conveyed. The literal content can remain similar while the message attributed to the speaker changes. [2]
A difference in the meaning attributed to a user’s words can change what the model understands the user to be asking.
English in Multilingual Model Processing
A conversation can remain in one language while the model’s intermediate representations show patterns associated with another. These representations are numerical states that incorporate information about the text and its context as it passes through the model’s layers.
In controlled Llama-2 experiments, analyses of intermediate layers often favored the English word for a concept before later layers favored its equivalent in the required output language. The researchers describe this as an English-biased conceptual space, linking the pattern to extensive exposure to English during training. [3]
Why Token Counts Differ Across Languages
Before processing a request, the model needs its text represented as tokens. A tokenizer divides that text using a vocabulary of units that can include whole words, word fragments, and smaller character or byte sequences. A word in one language may therefore require a different number of tokens from its translation.
How compactly a language is represented depends on the units available in that vocabulary. Subword tokenizers learn units from recurring patterns in their training data, and their coverage interacts with the language’s structure and writing system. Where suitable units are missing, words are divided into smaller pieces, producing a longer token sequence. [6]
Comparisons of parallel texts found that many non-English languages required more tokens than English to convey the same information. At a fixed per-token price, this makes comparable content more expensive to process. The additional cost reflects a difference in how the content is encoded. [6]
How Tokenization Relates to Internal Representations
Training provides a link between English’s compact tokenization and its prominence in the internal representations observed in Llama-2. The tokenizer’s vocabulary reflects the text used to train it, while the model’s conceptual representations reflect the language encountered during model training. Strong English representation in those data can help explain why English is both compactly encoded and closely aligned with the learned conceptual space. [3, 6]
Tokenization determines the pieces of text from which the model builds its interpretation. Information from several tokens may need to be combined to represent a single word. In the Llama-2 researchers’ conceptual account, this integration contributes to the formation of words and larger units of meaning. [3]
Their word-repetition experiment connects this process to the observed preference for English. That preference was less pronounced where complete words had language-specific tokens. The authors suggest that fragmentation can increase reliance on an English-biased semantic space as the word’s meaning is assembled. [3]
How a Difference in Meaning Could Change the Response
Processing a request through representations associated with another language could change the requirements, assumptions, or goals attributed to it. A word may still identify the right subject while the interpretation of what should be done with that subject shifts. The response can then develop a coherent line of reasoning around an altered task.
Consider an editing request to “make this argument stronger.” The user may mean improving the evidence and the connections between claims. Reading “stronger” as a demand for greater certainty would lead toward a different revision: more assertive wording without better support. Both readings concern the same argument, but they define different tasks.
An answer built on the second interpretation could be polished and persuasive while missing the improvement the user wanted. The user’s intention has not changed: it was expressed through the original language, but the response is organized around a different meaning. Establishing multilingual processing as the explanation for a particular response would require connecting the relevant semantic difference to the model’s processing and the interpretation reflected in its answer.
Preserving Intent Across Languages
For a model with English-biased representations, writing directly in precise technical English may help align the request with the language patterns involved in processing it. The same aim calls for explicit definitions of the operation, its scope, and the expected result. The request should supply the distinctions the user wants preserved when its words are interpreted in context.
In the editing example, this means defining “stronger” as better evidence and clearer reasoning, while preserving the level of certainty justified by the evidence. That formulation identifies both the improvement being requested and the boundary that must be respected. The model receives the intended distinction as part of the instruction, instead of having to infer it from the word alone.
Pseudo-code carries this precision into the logic connecting the requirements. It can specify which action follows from each condition, the order in which checks should occur, and what to do when a requirement is unresolved. For the same editing task, the logic could require checking whether a proposed revision overstates the evidence before applying it, and asking for clarification when the intended change is unclear. The relationship between the condition and the permitted action becomes explicit in the request.
Precise technical English and explicit logical structure may therefore help align the model’s interpretation with the user’s intended task. A change in language should not become a change in the request the model acts on.
References
- Thompson, B., Roberts, S. G., & Lupyan, G. (2020). Cultural influences on word meanings revealed through large-scale semantic alignment. Nature Human Behaviour, 4, 1029–1038. DOI
10.1038/s41562-020-0924-8. - Francesch, P., & Payrató, L. (2024). Pragmatic Ambiguity, Implicatures, and Translation. Studia Linguistica, 78(1), 156–185.
- Wendler, C., Veselovsky, V., Monea, G., & West, R. (2024). Do Llamas Work in English? On the Latent Language of Multilingual Transformers. Proceedings of ACL 2024, 15366–15394.
- Zhong, C., Liu, Q., Cheng, F., Jiang, J., Wan, Z., Chu, C., Murawaki, Y., & Kurohashi, S. (2025). What Language Do Non-English-Centric Large Language Models Think in?. Findings of ACL 2025, 26333–26346.
- Bayazit, D., AlKhamissi, B., & Bosselut, A. (2026). Lingua Franca or Probing Artifact? Rethinking Latent Language in Multilingual LLMs. Findings of EMNLP 2026, 10563–10583.