Do AIs think differently in different languages?

Kelsey Piper with a great piece of empirical AI journalism which basically asks "does the Sapir-Whorf hypothesis hold for LLMs?" It doesn't really for humans and, it turns out it doesn't really for LLMs, either! Models are pretty much center-left across the board, with some edge cases specific to DeepSeek when used in Chinese.

The approach here is a good one, using questions from the World Values Survey, paying translators for questions & answers in Arabic, Hindi, and Chinese, and evaluating multiple state-of-the-art models.

One of the challenges of making any generalized assertions about how LLMs behave is that most of the academics who are motivated and have the skills to do so aren’t able to move quickly enough to study the newest models. Often, when findings come out, a new generation of models has already emerged.