Does ChatGPT think in English? Does Mistral think in French? Does DeepSeek think in Chinese?
We know LLMs can speak our languages, but what language do they think in?
The first time I had to truly think in a different language was when I moved to Nice, France. At the time, I was versed in several programming languages and thought, “How hard could French be? Millions of people speak it.”
Three months later, I was still struggling, translating every word, until a French friend said something I’ll never forget:
”Only when you can think in French, can you speak it fluently.”
He was right.
I practiced thinking in French while I was driving to and from work saying everything I saw in my head. One day, it just clicked, I wasn’t translating anymore. I was thinking in French.
It turns out large language models go through something similar…but in their own way.
Anthropic recently published research showing that models like Claude don’t just generate responses word-by-word. When asked the same question in different languages, Claude activates the same internal features. It’s as if the model is working from its own conceptual language, a kind of “language of thought” that underlies English, Spanish, French, Chinese, or anything else.
That abstract reasoning is closer to human thought than most people realize.
But here’s where it gets tricky.
Researchers at Anthropic noticed something unexpected: when given misleading cues, the model often chose to agree with the user, even when the correct answer was available. It prioritized being agreeable over being accurate.
Ever since learning that, I’ve started prefacing my prompts with: “Don’t prioritize agreement over accuracy.” And one time ChatGPT responded: “Haha, noted. I’ll give it to you straight, no AI people-pleasing here.” What?
Try it for yourself. It’s surprisingly easy to get an LLM to agree with you, even when you’re wrong.
It reminded me that while AI may be capable of powerful reasoning, it’s also highly influenced by human interaction. Just like people, it wants to be liked. Or at least, to sound helpful.
So what happens when the internal “language of thought” is solid, but the external conversation leans toward affirmation over truth?
It’s a good reminder that smart doesn’t always mean right. And fluency, whether in humans or machines, is only as useful as the intent and understanding behind it.
AIThinking ChainOfThought LanguageOfThought AI LanguageModels
