In our everyday use of large language models (LLMs), we tend to believe that their responses are the result of pure logic and factual data processing. The reality, however, is far more complex. As a new study published in the Proceedings of the National Academy of Sciences (PNAS) shows, current AI models are not just "intelligent," but are also deeply influenced by the cultural context in which they were developed and trained.
Moral compass in the digital age: What do the data say?
A research team examined the ability of OpenAI models — specifically GPT-3.5, GPT-4, and the latest GPT-4o — to estimate the moral norms of 48 different nations. The researchers compared the AI's responses with data obtained from more than 90,000 people worldwide. For evaluation, they used the so-called Moral Foundations Questionnaire, which measures six key values: care, equality, proportionality, loyalty, authority, and purity.
The results are unequivocal. AI models systematically favor values such as care and equality, which are pillars of Western liberal society. Conversely, values like purity (e.g., religious or traditional respect for the sacred) are undervalued in the models, even when the AI is supposed to present itself as the "average citizen" of, for example, Morocco or Nigeria. The models thus consistently exhibit a tendency to "Westernize" responses, regardless of the cultural context in which they are supposed to operate.
Why does this happen? The data foundation problem
The cause lies not in the architecture of neural networks themselves, but in the data used to train them. As an article in The Conversation notes, the internet is still primarily an English-speaking space. Around 2023, English-language websites made up approximately 59% of all content on the web. Most of this content originates from the United States.
This creates an effect where models learn about the world through the lens of an American user. A practical example is when ChatGPT previously misinterpreted restaurant customs in Madrid — because large tips are standard in the US, the model assessed the Spanish absence of tipping as either impoliteness or financial difficulty, instead of understanding the local culture. Although models are gradually improving thanks to newer updates, the fundamental inclination toward Western hegemony remains.
Model comparison: Who fares better in ethical neutrality?
If we wanted to find a model with the least amount of cultural distortion, we need to look at how models are "tuned" (the so-called alignment process using RLHF — Reinforcement Learning from Human Feedback).
- OpenAI (ChatGPT/GPT-4o): Currently the most popular model. It is highly capable in logic, but its safety filters and moral rules are strongly defined by American ethical standards.
Price: Free tier, ChatGPT Plus costs $20/month (approx. 470 CZK). - Anthropic (Claude 3.5 Sonnet): Considered by many to be a "more human" and less rigid model. Claude strives for a higher degree of nuance, but still draws from similar English-language datasets.
Price: Free tier, Claude Pro costs $20/month (approx. 470 CZK). - Google (Gemini): Google has a huge advantage in data diversity thanks to YouTube and search, which can lead to broader cultural understanding. However, Gemini often suffers from excessive "political correctness," which is also defined by Western corporate standards.
Price: Free tier, Gemini Advanced costs approx. €20/month (approx. 500 CZK).
Impact on the Czech market and the European sphere
For Czech users and businesses, this has three main dimensions:
- Language barrier vs. cultural barrier: Even though models like ChatGPT or Gemini handle Czech very well, their "thinking" is still built on English concepts. This means that when translating cultural nuances (e.g., in marketing or legal texts), AI may produce Western idioms or values that make no sense in a Czech context.
- EU AI Act regulation: The European Union is trying to address this problem through the AI Act. The regulation emphasizes that AI systems must be transparent and that the risk of discrimination and bias must be minimized. Companies in the Czech Republic that implement AI into their processes (e.g., HR or customer support) must be aware that the model may unknowingly discriminate against certain groups based on cultural prejudices.
- Business risk: If a Czech company uses AI for global expansion (for example into Asia), it must not blindly rely on generated texts. Without human review, AI communication with clients from other cultures may come across as inappropriate or even offensive, because it will apply Western moral standards where they do not apply.
How to minimize the risk of bias?
As experts, we recommend not using LLMs as "truth," but as a "draft." To reduce cultural bias, the most effective technique is RAG (Retrieval-Augmented Generation). Instead of relying solely on what the model "knows" from its training, provide it with specific, local, and relevant documents as context (e.g., the Czech legal code or your company's ethical code). This way, you "instruct" the model to stick to your facts, not its own internal Western biases.
Will using the Czech language reduce the model's cultural bias?
Unfortunately not. Even when the model responds in Czech, its internal logical structures and "moral rules" are still shaped primarily by English-language data. Czech is more of a linguistic wrapper for models than a new way of thinking.
How can I tell if the AI is responding with cultural bias?
If you notice that the model, in questions concerning traditions, religion, or social roles, consistently favors a liberal Western viewpoint (e.g., on matters of family, authority, or social obligations), you are likely facing cultural bias.
Is it possible to completely "cleanse" AI models of these biases?
At present, this is nearly impossible, because data on the internet is inevitably skewed. The goal is not complete neutrality (which doesn't exist), but increasing the diversity of training data and improving control over how models interpret different cultural contexts.
