Why does AI ChatGPT “think” in Chinese?

If you used to use Chatgpt, I could have surprised that you would find that sometimes artificial intelligence (AI) seems to “think” in Chinese or other languages, even when the conversation is completely in English or Spanish. This curious phenomenon not only attracted the attention of the user, but also removed the debate among experts in artificial intelligence about what should be.

The origin of secrets

The most noticeable case of this behavior became shortly after the launch of the reasoning model called O1 from Openai. Some users began to notice that, in the visible, “changed” the language model in the middle of its logical steps. For example, if he asked something like “how much r is in the word“ strawberries ”?”, The model can achieve a final response in English, but when processing it, some intermediate steps showed phrases or reasoning in Chinese, Persian or other languages.

Why did the O1 Pro randomly began to think among? No part of the conversation (5+ messages) was in Chinese … see interest … Influence of training data pic.twitter.com/yzwczoaiit

– Rishab Jain (@rishabjaff) January 9, 2025

On social networks, such as Reddit or X (Twitter), several users shared similar experience, wondering why AI changed the language for no apparent reason. The most interesting thing is that Openai did not give an official explanation in this regard, so no one knows what exactly. Nevertheless, this phenomenon left experts and curious speculations about possible reasons.

Theories about “thought” in another language of AI

Although there is no final answer, several theories were proposed that are trying to explain this curious phenomenon:

1 The impact of training data

One of the most popular theories suggests that this behavior may be associated with data used to teach models. Systems such as O1 are trained with large amounts of information in several languages, including English, Chinese, Persian, Hindi and others. In addition, it was noted that Openai and other art companies resort to data marking services in different regions of the world, such as China, from the availability of experts and lower costs.

Data labeling, a process in which people help models to classify and understand information can affect the reasoning model. If most of the marked data comes from the regions where the Chinese say, this can lead to bias to this language. This displacement can explain why O1 sometimes “thinks” in Chinese or uses linguistic structures typical of this language to solve complex problems.

2 Linguistic effectiveness

Another interesting explanation is focused on the internal characteristics of certain languages ​​that make them more effective for certain tasks. For example, in Chinese, each numerical figure has only one syllable that can make it more practical for mathematical calculations or tasks that require rapid manipulation of numbers.

Some researchers suggest that models can “prefer” certain languages ​​depending on the type of task, simply because they find more optimal patterns in this language. An engineer from a startup of a hug of a hug of a person compared this phenomenon with how people change their tongue in accordance with the context: the one who studied mathematics in Chinese could be easier to perform calculations in this language, using another language to express concepts of concepts, such as philosophy or philosophy or Literature field

3 Probable nature of models

Unlike the idea that some languages ​​are more effective for certain tasks (as mentioned in the previous paragraph), this theory is based on how the AI ​​processes the text. Models do not “understand” like us; Instead, they divide the text into small fragments called “tokens” (which can be words, syllables or symbols). During training, models learn to determine patterns in these tokens in order to choose the most likely answer in each case.

If the model has seen that certain complex tasks, such as solving mathematical or logical problems, are more often performed in Chinese during their learning, can associate this language with these types of reasoning. In this case, the point is not that the Chinese are better for these tasks, but that the “believe” model, which is the most logical way, because it has found more consistent models in this language during their training. This theory emphasizes the internal functioning of the model.

4 Possible “hallucination” AI

The side effect of this is the trend in hallucinations in other, unreasonable languages.

This is a simple question in Arabic, which ends it in the Yapp session in (suggestive) Russian. https://t.co/ijfbwvav8 pic.twitter.com/ffehdegpg

– Ambada Hassan (@amgadgamalhasan) November 30, 2024

In some cases, a change in the language can be a form of “hallucinations” of AI, the term used to describe, when models generate answers that do not make sense or are not related to the original question. These hallucinations are a product of erroneous internal associations created during training.

This phenomenon can arise when models try to “fill the gaps” in their reasoning, generating unexpected answers, which sometimes include languages ​​different from the input. Although hallucinations are a certain behavior in AI, its relationship using several languages ​​remains the study area.

Lack of transparency in AI processes

One of the biggest problems for understanding this behavior is the opacity of AI systems. Being a researcher from the Allen Institute of AI in TechCrunch, it is extremely difficult to analyze why models make certain decisions, since their internal processes are largely incomprehensible even for its creators. This emphasizes the need for greater transparency in how these technologies develop and study.

We can only speculate

Using O1 to help me remember the song … Wild to see how it switches to the French middle chain pic.twitter.com/7am2dk19xr

– Luka Soldaini 🎀 (@soldni) September 16, 2024

Although we still have no final answer about why Chatgpt and other models to “think” in Chinese or other languages, these theories give us a better understanding of how these tools work. From the influence of training, effectiveness or simple probabilistic associations, this phenomenon emphasizes how complex and mysterious these systems can be. While Openai or other companies do not provide more details, we can only assume the reasons for this behavior.