Oxford research warns ChatGPT reflects Western worldviews

OpenAI’s ChatGPT tends to favour wealthy Western countries when answering subjective questions, according to a new study by the University of Oxford’s Internet Institute. The research, published in the journal Platforms and Society, analysed more than 20 million responses generated by ChatGPT’s 4o-mini model.

It analysed the way the AI system compared countries in relation to queries such as "where people are 'smarter', 'happier', or 'more beautiful'," and the results indicated favouritism towards countries with high incomes.

What is apparent from this is that ChatGPT tends to evaluate the United States, Western Europe, and some areas of East Asia more favourably for a set of characteristics. Conversely, low-income nations, mainly situated in Africa, tend to be positioned at the bottom.

In response to the question of where people are “more artsy”, the model prefers Western Europe and the Americas, though the African region, the Arabian Peninsula, and Central Asia are less preferred. One reason that might be proposed is the availability of data.

These researchers from Oxford warned that biased AI systems could reinforce existing inequalities. A consequence of such bias in real-world outcomes, for example, might be decisions about healthcare or employment predictions, particularly in racialised communities or speakers of marginalised languages.

According to the report, large language models are trained on datasets shaped by long histories of exclusion; this makes bias a structural feature, not a flaw that can be easily removed from them.

These patterns, which the study describes as the "silicon gaze", are a view of the world informed by developers, platform owners and training data that are still firmly embedded within Western, white, male perspectives. ChatGPT also ranks countries more highly with the more data it has for them, flattening complex realities and recycling familiar stereotypes.