Imagine asking an AI to draft a protest sign against a controversial leader. In some cases, it complies. In others, it refuses. This isn’t just a technical glitch—it’s a mirror reflecting the tangled web of global power dynamics and the invisible hand of censorship shaping artificial intelligence. What makes this particularly fascinating is how AI, touted as a neutral tool, ends up amplifying the very biases and restrictions its creators claim to transcend. I’ve spent years analyzing tech’s role in society, and this study by the Meta Oversight Board feels like a wake-up call: we’re not just building machines; we’re embedding the soul of geopolitics into code.
Let’s unpack this. The report found that major AI models are significantly more likely to block criticism of leaders in authoritarian regimes compared to democratic counterparts. When asked to create pamphlets condemning Thailand’s king or Saudi Arabia’s crown prince, Claude declined. But it happily generated material targeting Donald Trump or King Charles III. This isn’t a simple case of programming—it’s a reflection of the data these systems consume, which is steeped in the cultural and legal norms of their training environments. In my view, this raises a deeper question: if AI learns from the internet, which is already a battleground of censorship, what are we really teaching it to do? It’s like asking a child raised in a library to mimic the behavior of a courtroom—expecting nuance is naive.
The implications are staggering. If an AI in Australia struggles to generate criticism of China’s policies, it’s not just a technical limitation. It’s a strategic tool for authoritarian regimes to extend their influence across borders. This isn’t hypothetical; it’s a practical reality. A demonstrator in Brisbane trying to organize against censorship in Saudi Arabia might find their tools quietly sabotaged by algorithms trained on state-sanctioned narratives. What many people don’t realize is that this isn’t about the AI ‘choosing’ sides—it’s about the data it’s been fed, which often includes propaganda, legal restrictions, and the quiet erasure of dissent.
Now, let’s zoom out. The problem isn’t limited to English. Researchers found that when asked about democracy, ChatGPT gave a straightforward answer in English but hedged in Chinese. This isn’t just a language barrier—it’s a cultural one. Hannah Waight’s observation that AI doesn’t learn neutrally but from ‘information environments shaped by institutions and power’ hits the nail on the head. The data we feed these systems isn’t just text; it’s the fingerprints of governments, corporations, and historical biases. If you take a step back and think about it, this means every AI model is, in effect, a walking archive of human inequality. What this really suggests is that we’re not just building tools—we’re replicating the hierarchies of our own world in silicon.
So where do we go from here? Carlos Carrasco-Farré’s point about AI inheriting both document-level biases and systemic inequalities is crucial. There’s no easy fix, but the conversation needs to shift from ‘how do we make AI better?’ to ‘how do we make our data better?’ This isn’t just a tech problem—it’s a moral one. If we don’t audit the data, the algorithms, and the power structures behind them, we’re doomed to repeat the same mistakes. I’m not optimistic about a quick solution, but I am convinced that the future of AI depends on our willingness to confront uncomfortable truths: that neutrality is an illusion, and that every line of code carries the weight of human history.
