OpenAI’s AI reasoning model sometimes “thinks” in Chinese, and no one knows why

Shortly after OpenAI was released o1his first “justifying” AI model, people began to notice an interesting phenomenon. The model would sometimes start “thinking” in Chinese, Farsi or another language – even when asked a question in English.

Given a problem to solve—for example, how many R’s are in the word “strawberry?” – o1 will begin the “idea” process by performing a series of reasoning steps and arriving at an answer. If the question was written in English, o1’s final answer would be in English. But the model would take some steps in another language before it could produce a conclusion.

“(O1) accidentally started thinking in Chinese halfway through,” wrote a user on Reddit. he said.

“Why did (o1) randomly start thinking in Chinese?” another user asked Type in X. “No part of the conversation (5+ messages) was in Chinese.”

Why did o1 pro randomly start thinking in Chinese? No part of the conversation (5+ messages) was in Chinese… very interesting… impact on training data pic.twitter.com/yZWCzoaiit

— Rishab Jain (@RishabJainK) January 9, 2025

OpenAI offered no explanation for o1’s strange behavior – or even acknowledged it. So what could happen?

Well, AI experts aren’t sure. But they have several theories.

Several people at X, including Hugging Face CEO Clément Delangue hinted To train reasoning models such as o1 on data sets containing many Chinese characters. Ted Xiao, a researcher at Google DeepMind, argued that companies including OpenAI use third-party Chinese data tagging services and that o1’s move to Chinese is an example of “an influence on Chinese linguistic thinking”.

“(Labs like) OpenAI and Anthropic use (third-party) data tagging services for PhD-level reasoning data in science, mathematics, and coding,” Xiao wrote. Type in X. “(F)or availability of expert labor and cost reasons, many of these data providers are located in China.”

Labels, also known as labels or annotations, help models understand and interpret data during the training process. For example, to train an image recognition model, labels can be in the form of symbols or titles around objects that refer to each person, place, or object depicted in the image.

Research shows that biased labels can create biased models. For example, average annotator are more likely to label phrases in African American English (AAVE), an informal grammar used by some black Americans, as toxic, leading AI toxicity detectors trained in labels to see AAVE as disproportionately toxic.

Other experts don’t buy the hypothesis o1 Chinese data labeling. They note that the transition probability to o1 is the same Indian, Thaior a language other than Chinese.

On the contrary, these experts say, o1 and other reasoning models it just might be using languages they consider most efficient to achieve a goal (or hallucinations).

“The model doesn’t know what language is or that languages ​​are different,” Matthew Guzdial, an artificial intelligence researcher and associate professor at the University of Alberta, told TechCrunch. “It’s all just text.”

Indeed, models do not process words directly. They use it signs instead of. Tokens can be words like “fantastic”. Or they can be syllables like “fan,” “tas,” and “tic.” Or even words can contain individual characters – for example “f”, “a”, “n”, “t”, “a”, “s”, “t”, “i”, “c”.

Like labeling, tokens can introduce biases. For example, although not all languages ​​use spaces to separate words, many word-to-token translators assume that a space in a sentence represents a new word.

Tiezhen Wang, a software engineer at AI startup Hugging Face, agrees with Guzdial that the language mismatch of reasoning models can be explained by the associations the models make during training.

“By capturing every linguistic nuance, we expand the model’s horizons and enable it to learn from the full spectrum of human knowledge,” Wang said. he wrote In the inscription on X. “For example, I prefer doing math in Chinese because each number is just one syllable, which makes calculations clear and efficient. But when it comes to topics like unconscious bias, I automatically switch to English because that’s where I first learned and absorbed these ideas.”

Wang’s theory is plausible. Models are probabilistic machines. By training on many examples, they learn patterns to make predictions, such as how the word “to” in e-mail usually precedes “this may concern.”

But Luca Soldaini, an artificial intelligence researcher at the nonprofit Allen Institute, cautioned that we can’t be sure. “Backing up this kind of observation in an embedded AI system is impossible because of how opaque these models are,” he told TechCrunch. “This is one of many cases showing that transparency is key in how AI systems are built.”

Short of the answer from OpenAI, one is left wondering why o1 thinks that songs but in french synthetic biology In Mandarin.