Why can an AI language model split a short Chinese phrase and a short English word into different numbers of small text pieces?

Answer: The model learns common text pieces from the training text of each language.

Each written character is always kept as exactly one small piece. — A character can become one piece, so this captures a common case. But the model does not promise that every written character stays separate. A frequent phrase may be stored as one larger piece, while a rare character or symbol may be split into smaller pieces. The count comes from learned text patterns, not from a one-character-one-piece rule.

The model learns common text pieces from the training text of each language. ✓ — The model’s splitter is built from patterns in training text, not from a fixed rule that maps one character to one piece. Because languages repeat different character combinations and word boundaries, the common pieces it learns differ by language. That lets a short Chinese phrase become one or two pieces while a short English word may need several, or the reverse. The lengths look unrelated because the learned pieces, not visible length alone, set the count.

Only the spaces between words decide where the small pieces are cut. — Spaces are a real clue in writing systems that use them, so this gets part of the story right. But the model’s pieces are learned from training text, and Chinese text often has no spaces between words. Punctuation, common phrases, and repeated character combinations also affect where cuts fall. Spaces can influence the result, but they do not decide it alone.