Cultural Knowledge Inside AI Models
- Marina Pantcheva
- Jul 30
- 4 min read
This is the fourth edition of AI Research Explained, where we break down important AI research in simple terms and discuss what it means for our industry.
This edition looks at a research paper titled “Localized Cultural Knowledge is Conserved and Controllable in Large Language Models,” published in April 2025. It starts with a simple question:
When an AI model speaks a language, does it adopt the culture associated with that language?
To find out the answer, you can keep reading or watch the video.
The pumpkin test
If you ask an AI model, in Japanese:
What color is a pumpkin?
The model will answer "Orange." That makes sense if we picture the kind of pumpkins common in the United States and Europe.
In Japan, however, pumpkins are commonly green.
Now if we change the question to:
What color is a pumpkin in Japan?
The model will correctly answer "Green."
So the model actually knows the culturally appropriate answer — but it cannot retrieve it when the only cultural clue is the language of the prompt. It needs to be explicitly prompted to answer for the context of Japan.
The explicit–implicit localization gap
The researchers call this gap in performance the explicit–implicit localization gap. Implicit localization means that the culture is suggested indirectly, for example by the language of the prompt. Explicit localization means the culture is clearly specified, as in “What color is a pumpkin in Japan?”
To test how big this gap is, the researchers studied five languages and cultural settings:
English and the United States,
French and France,
Russian and Russia,
Turkish and Turkey, and
Bengali and Bangladesh.
They tested seven AI models on cultural questions, aiming to find out how much better a model performs when explicitly prompted, compared to when simply prompted in the local language.
The result was a clear gap. Across most models, explicitly naming the cultural context improved performance by more than ten percentage points. In one case, the gap reached sixty-eight percentage points.
In other words, prompting an AI model in a local language does not provide enough of a cultural signal. The model must have the culture clearly indicated in the prompt to produce an answer that fits the culture.
To indicate the culture, though, one does not need to be very explicit. Just a single culturally specific word is enough to snap the model to the right cultural frame. When the researchers prefixed the prompts with culturally loaded words such as baguette, baklava, or borscht, the models produced more culturally aligned responses.
Prompting an AI model in a local language does not provide enough of a cultural signal.
Was the problem solved?
Not really. Explicit prompting improved cultural accuracy, but the answers became more similar and stereotypical. The explicit specification of the culture pushed the model toward creating culture caricatures and producing less diverse responses.
Looking inside the model
The researchers therefore asked the next question:
Can we activate cultural knowledge without placing an explicit cultural label in the prompt?
They looked inside one model, Gemma 2 9B, using a technique called activation patching, which allowed them to trace where cultural localization actually happened inside the network.
And they found it. Cultural localization was concentrated in a specific band of middle layers, roughly between layers 23 and 30. That is where the model converted its default world model into one with a cultural flavor. Crucially, this same region lit up across different languages and different tasks, indicating that the mechanism is universal.
Steering toward a culture
That led to the paper’s most interesting experiment: steering. The researchers computed what is called a steering vector — basically a vector encoding a "cultural direction." They found it by comparing the model’s internal state with explicit cultural context and without cultural context, then isolating the difference to obtain a cultural vector.
Then they simply added that cultural vector while the model was generating responses. When they added the "Turkey vector," the model localized toward Turkish culture on its own; when they subtracted it, the model drifted toward US culture.
Steering recovered much of the accuracy of explicit prompting, but at the same time, it kept the natural diversity of implicit prompting, produced fewer stereotypes, and stayed more faithful to the real culture.
The researchers also found evidence of a more general, or "universal," culture steering vector, which signaled to the model to simply follow the culture associated with the prompt’s language. This improved performance, yet remained weaker than culture-specific steering and explicit prompting. So, not the one magical culture switch we were hoping for.
What it means for localization
The central message of this paper is that large language models store some cultural knowledge but fail to surface it. The challenge, then, is how to activate that cultural knowledge without reducing it to stereotypes and cultural clichés.
From this perspective, localization becomes the know-how of how to steer the cultural knowledge encoded deep within the model to produce culturally aligned translations.
