Why Does ChatGPT Sometimes "Reason" in Chinese?

When using ChatGPT, some users have been surprised to notice that sometimes the artificial intelligence seems to "think" in Chinese or other languages, even when the conversation is conducted entirely in English or Spanish. This phenomenon has sparked curiosity and generated debate among professionals in the field of artificial intelligence.

One of the most notable cases occurred after the release of the reasoning model known as "o1" by OpenAI. Several users began to observe that, during the process of reaching an answer, the model could "switch" languages in the middle of its reasoning. For example, when asking "How many R's are in the word ‘strawberry’?", the model would give the final answer in English, but some intermediate steps included phrases in Chinese, Persian, or other languages.

This behavior has been discussed on social platforms, where numerous users have shared similar experiences and expressed their curiosity about why the artificial intelligence alternates between languages without an obvious reason. What is most intriguing is that OpenAI has not provided an official explanation.

Various theories have been formulated to try to clarify this phenomenon:

1. Influence of training data: One of the most accepted explanations suggests that this behavior could be related to the nature of the data used to train the models. Since models like o1 are fed with large amounts of information in multiple languages, including English, Chinese, Persian, and others, it is possible that there is a bias towards certain languages if the data in those languages is more predominant. This could lead the model to "think" in Chinese when solving problems, as it has found patterns that are more familiar to it in that language.

2. Linguistic efficiency: Another theory suggests that certain characteristics of specific languages might make them more efficient for certain tasks. For example, in Chinese, each number has only one syllable, which could facilitate tasks involving calculations. It has been proposed that models choose a language based on the task at hand, similar to how humans switch languages according to context.

3. Probabilistic nature of the models: Instead of claiming that some languages are more effective, this theory emphasizes how AI processes text. The models break down information into small chunks and learn to identify patterns. If they have observed that complex problems are often solved in Chinese during their training, they might associate that language with such reasoning, choosing what they consider the most logical path based on their previous experiences.

4. Possible AI "hallucination": In some cases, the language switch can be seen as a type of AI "hallucination," where the model produces nonsensical answers related to the original query. This can happen when it tries to fill in gaps in its reasoning, resulting in answers that include languages different from the initial conversation.

The lack of clarity regarding how these systems work presents a significant challenge. The complexity of AI's internal processes makes it difficult to analyze their decisions, highlighting the need for greater transparency in the development and training of these technologies. Although there is still no clear answer as to why models like ChatGPT "think" in other languages, these theories provide deeper insight into their functioning.

Until more clarifications are provided by OpenAI and other entities in the sector, the mystery surrounding this phenomenon will continue to be a subject of speculation.