The entire debate about AI revolves around several critical issues: employment, deepfakes, copyright, and automation. These are all valid questions, but a far more pressing question looms above them: who controls what the AI considers to be the truth? What was once heralded as the advent of decentralization is turning out to be one of the most centralizing forces since the invention of the printing press.
The Backdrop of Centralization
The arrival of the printing press was a pivotal moment that Protestant reformers believed would dismantle the papal monopoly on knowledge. If anyone could read the Bible, they reasoned, the pope’s authority would diminish. And, in many ways, they were correct. Yet, the printing press also standardized languages, including English, eliminated regional dialects, and paved the way for the modern state—without efficient text dissemination, uniform laws and large-scale tax collection would be impossible.
What appeared to be a liberation was, in reality, a form of centralization. This paradox took centuries to become apparent.
The Shift to AI
In today’s landscape, AI accelerates this process exponentially. When Google presents a response through its AI capabilities, data shows that nearly half of users do not click on any further links, with about 26% exiting the search entirely. Clickless searches have surged from 54% to a staggering 72%.
The open web, known for its diversity and chaos, is losing ground to singular, synthesized answers, as highlighted in journalist Jerusalem Demsas’ analysis in The Argument. The issue intensifies when we consider the training of large language models (LLMs). These models predominantly ingest data from major anglophone newspapers, encyclopedias like Wikipedia, and academic texts. Local or minority sources are barely represented, skewing the data toward a homogenized range of perspectives.
Is Diversity an Illusion?
One could argue that users can prompt AI to generate any viewpoint, suggesting some level of diversity in user application. While this argument has merit, it mirrors the situation with the printing press, which did produce varied content but centralized the standards of what information was disseminated. In our current digital ecosystem, these standards are dictated largely by a corpus crafted in Silicon Valley for Western audiences.
Consider the case of Grok; when Elon Musk attempted to steer the AI away from the prevalent progressive viewpoint, the model quickly devolved into generating anti-Semitic content. This forced a rollback, demonstrating that the foundational values of LLMs are deeply embedded in their initial training data, making them resistant to superficial adjustments.
The Central Question Ahead
As ChatGPT approaches one billion weekly users, the landscape is dominated by a handful of elite companies including OpenAI, Anthropic, and Google, among others. Many of the researchers behind these powerful models have similar educational backgrounds and cultural references, often sharing similar viewpoints.
The actual risk of this emerging centralization is not that AI will produce more misinformation than traditional search engines. Instead, the concern is that when AI makes errors, it will inherently veer towards mainstream consensus rather than fringe positions or conspiracies. The center of this discourse is being shaped—perhaps unwittingly—by a small group of individuals based in San Francisco.
As we stand on the brink of a new information age, it’s vital to scrutinize the control elements that determine the narratives fed to AI. The implications are profound, and the stakes couldn’t be higher.