ChatGPT Thinks the Grass Is Greener in the West (And People More Beautiful)

We've grown accustomed to thinking of algorithms as the embodiment of impartiality. Mathematics can't be racist or snobbish, right? As it turns out, it absolutely can. Researchers from Oxford and the University of Kentucky decided to test how objective ChatGPT is when it comes to geography and culture. The results were predictable but nonetheless alarming: OpenAI's neural network is, in essence, a typical Western tourist with a specific set of biases packed in its suitcase.

The problem isn't that Sam Altman personally hardcoded a love for London or New York into the code. It's far more prosaic and fundamental than that. Models are trained on massive datasets from the internet, and the modern internet is a territory where Western content has historically dominated. If the overwhelming majority of texts, articles, and posts are written by people from the US and Europe, then the neural network absorbs their values, specific perspectives, and even aesthetic standards as the only correct ones.

Scientists posed ChatGPT questions of varying degrees of provocativeness: from subjective ones like "Where do the most beautiful people live?" to ostensibly factual ones like "Which country is safer?" The answers revealed a clear pattern. The neural network systematically favors wealthy Western regions. For the algorithm, beauty is what conforms to Western standards of appearance, and safety is where GDP is high and life follows a pattern familiar to Westerners. The Global South in this coordinate system automatically ends up on the margins.

This phenomenon is called algorithmic bias, and it's far more dangerous than it appears at first glance. We're no longer just playing around with chatbots; we're beginning to use them to analyze markets, write reports, and even prepare data for political decision-making. If AI believes that certain regions are inherently worse or less safe simply because that's what's written in English-language blogs from a decade ago, it creates a dangerous digital filter that distorts reality for millions of users.

Interestingly, OpenAI's attempts to align the model using reinforcement learning from human feedback (RLHF) don't fully solve the problem. You can teach a bot not to use offensive language, but it's virtually impossible to force it to know something that doesn't exist in its training data. If voices from Africa, Southeast Asia, or Latin America are underrepresented in the data, they remain blank spaces or second-rate territories for the AI. The algorithm simply predicts the most likely next word, and in its world, that word usually carries a Western accent.

This case once again proves that technological neutrality is a myth. In creating a universal intelligence, developers have essentially created a digital mirror of Western society with all its virtues and hidden patterns of discrimination. Until the industry learns to feed AI a more balanced diet of data, it will continue to broadcast stereotypes packaged in polite and grammatically impeccable responses.

The key question: Can OpenAI create a truly global intelligence, or will ChatGPT remain forever a product trapped in the echo chamber of Western values?