Why AI-Written Traditional Chinese Sometimes Contains Simplified Characters

How Large Models Produce Traditional Chinese

Currently, when mainstream large models process Chinese, their underlying training data is heavily skewed toward Simplified Chinese. This means that in the process of "learning" Chinese, the models primarily encounter Simplified character forms and texts. When a user requests Traditional Chinese output, the model is not "thinking in Traditional Chinese"; instead, it performs a character form mapping after generation—converting Simplified character forms to their corresponding Traditional forms.

This process is inherently unstable. The mapping relies on a set of internalized lookup tables within the model, not on a conversion engine with explicit rules. When the model's confidence in mapping a particular character is low, or the context does not provide sufficient clues, it may simply retain the Simplified character form without converting it.

Single-Character Infiltration: The Most Common Failure Mode

The way large models mix in Simplified Chinese is usually not by converting entire passages, but through single-character infiltration. Within a correct Traditional Chinese prose passage, one or two Simplified character forms might occasionally appear, such as zhe (corresponding to 'this'), ge (corresponding to 'individual'), or shuo (corresponding to 'say'). This type of error is difficult to spot with a casual visual scan, as the reader's eye tends to auto-complete.

Simplified character infiltration rarely involves just a single character. Once the model slips into Simplified mode at a point, character forms like zhe, ge, shi, shuo, and guo—which have no Traditional Chinese identity—will appear consecutively. These are purely Simplified characters and cannot be found in any Traditional Chinese dictionary.

Ambiguous Character Forms: The Converter's Trap

A more problematic issue lies not with characters that are "obviously Simplified," but with those that have ambiguous identities. Some character forms exist as both Simplified characters and independent, correct Traditional characters. These cannot be judged by the criterion of "whether they change when converted to Traditional," because they don't need to change—they are already correct.

Common examples include:

• yun: The Simplified form is yun, a simplification of the Traditional yun. However, in the phrase ren yun yi yun, the yun is the correct Traditional character, not a Simplified one.
• chou: The Simplified form is chou, a simplification of the Traditional chou. But the chou used in the Earthly Branches (地支) is the correct Traditional character and should not be converted to chou.
• gan: The Simplified form gan is a simplification of the Traditional gan or gan. However, the gan used in the Heavenly Stems (天干) is the correct Traditional character.
• tai, hou, li: These characters are correct Traditional characters in specific contexts and should not be uniformly converted to Tai, Hou, or Li.

If one uses the criterion "whether this character changes when converted to Traditional," all these correct Traditional characters would be misidentified as Simplified. This is a very common logical error.

The Reverse Problem of Converters

The trouble with character form conversion is not one-directional. Taking OpenCC as an example, its dictionary sometimes maps correct Traditional characters used in Taiwan to variant forms. For example:

• chuang is mapped to chuang
• zao is mapped to zao
• qun is mapped to qun
• gu is mapped to gu

In Taiwan's Ministry of Education《國語辭典》, chuang, zao, qun, and gu are all listed as correct characters. Judging by the dictionary, these correct characters are actually being treated as Simplified. This illustrates one point: to determine if a character form is correct, one must ultimately consult the dictionary of the specific region; one cannot rely solely on the output of a converter.

How to Handle This Problem

For websites using Traditional Chinese, content generated by large models needs to be proofread by humans or tools. The focus of proofreading is not "whether there are Simplified characters," but rather distinguishing between three situations:

1. Purely Simplified Characters (e.g., zhe, ge, shuo): These characters have no correct Traditional identity in a Traditional Chinese context; their appearance is an error.
2. Ambiguous Correct Characters (e.g., yun, chou, gan): Context is needed to judge these; they cannot be handled with a simple character form mapping table.
3. Over-correction by Converters (e.g., chuang being changed to chuang): Conversely, originally correct character forms being converted to variant characters also need to be caught.

This site's "Commonly Confused Characters" entries compile frequent cases of one Simplified character corresponding to multiple Traditional ones (e.g., the Earthly Branch chou, the Heavenly Stem gan), with dictionary sources for each entry. ⚠️ As for the category "converters mapping correct characters to variant characters" (chuang, zao, qun, gu), these have not yet been included in the entries—that's a different type of cause that requires consulting dictionaries on a case-by-case basis to determine. The design logic of detection tools also revolves around these three situations, rather than simply comparing Simplified and Traditional character forms.