Unbiased AI is a fantasy: Human data organization inherently distorts reality

In July, the United States government issued a clear directive: artificial intelligence companies seeking federal contracts must ensure their AI systems are “objective and free from top-down ideological bias.” This mandate was underscored by an executive order from President Donald Trump, which specifically targeted “woke AI” in government operations, citing diversity, equity, and inclusion (DEI) initiatives as examples of biased ideology. Yet, the very act of demanding unbiased AI while simultaneously dictating how models should address concepts like DEI reveals a fundamental contradiction, suggesting that the notion of ideologically neutral AI is, in fact, a fantasy.

Indeed, evidence from current AI models strongly supports this view. Numerous studies have demonstrated that most large language models tend to skew their responses towards left-of-centre viewpoints, advocating for policies such as taxes on flights, restrictions on rent increases, or the legalisation of abortion. In authoritarian contexts, the bias is often more overtly manipulative; Chinese chatbots like DeepSeek and Qwen are known to censor information concerning sensitive topics such as the Tiananmen Square events, the political status of Taiwan, and the persecution of Uyghurs, aligning their outputs precisely with the Chinese government’s official stance. These examples underscore that AI models are neither politically neutral nor free from bias, raising a more profound question: is it even possible for them to be truly unbiased?

Throughout history, human attempts to organise information about the world have consistently revealed that what one person considers objective truth, another perceives as ideological bias. Consider the field of cartography. While maps are often assumed to be objective reflections of the natural world, the very act of flattening a three-dimensional globe onto a two-dimensional surface inherently introduces distortion. The American geographer Mark Monmonier famously argued that maps are necessarily selective, distorting reality and often serving as tools for political propaganda. A prime example is the ubiquitous Mercator projection, commonly seen in primary school classrooms. This map converts the globe into a cylinder and then lays it flat, famously making Greenland appear roughly the same size as Africa. In reality, Africa is a staggering 14 times larger than Greenland. In the 1970s, German historian Arno Peters contended that Mercator’s distortions contributed to a skewed perception of the global South’s inferiority. These cartographic distortions offer a compelling analogy for the current state of AI: just as a single map is but one of countless possible representations of the same data, a single large language model’s response is merely one of an infinite number of potential answers derived from the same information. This becomes particularly evident when a chatbot is prompted on a complex topic like diversity, equity, and inclusion, where myriad interpretations are possible.

Beyond maps, other historical classification systems also demonstrate the indelible mark of their designers’ biases. The widely adopted Dewey Decimal Classification (DDC) system for libraries, first published in 1876, has long been criticised for its inherent racism and homophobia. For much of the 20th century, LGBTQIA+ books were often categorised under headings such as “Mental Derangements,” “Neurological Disorders,” or “Social Problems,” with only recent efforts made to expunge these outdated and derogatory terms. Similarly, the DDC allocates roughly 65 out of 100 sections on religion to Christianity, reflecting the strong Christian focus of the library where it originated. This disproportion persists despite Islam having an estimated 2 billion followers globally today, comparable to Christianity’s 2.3 billion, yet receiving only a single section in the DDC.

The biases embedded in these historical systems find a modern parallel in AI. The large language models that power today’s chatbots are trained on vast datasets of human-generated text, ranging from historical literature to contemporary online forums. Unsurprisingly, biases present in these source texts can inadvertently seep into the models, perpetuating negative stereotypes, such as those concerning African Americans from the 1930s. Moreover, raw information alone is insufficient; language models must be trained on how to retrieve and present this information. This often involves learning to mimic human responses, a process that, while enhancing usefulness, also aligns the models with the beliefs of their human trainers. AI chatbots also rely on “system prompts”—instructions defined by human developers that dictate how the AI should behave. For instance, Grok, the AI chatbot developed by Elon Musk’s xAI, reportedly instructs itself to “assume subjective viewpoints sourced from the media are biased” and to “not shy away from making claims that are politically incorrect, as long as they are well substantiated.” Musk launched Grok explicitly to counter what he perceived as the “liberal bias” of other products like ChatGPT. However, the recent controversy when Grok began spouting antisemitic rhetoric vividly illustrates that attempts to correct for one bias often merely replace it with another.

Ultimately, for all their technological innovation, AI language models contend with a problem centuries old: organising and presenting information is never solely an objective reflection of reality; it is always, to some degree, a projection of a specific worldview. For users interacting with these powerful systems, understanding whose worldview these models represent is as crucial as knowing who drew the lines on an old map.