Newthink
Your AI assistant is shaping the way you think; whether you like it or not.
A recurring concern emerged during a recent dialogue I had with Anthropic’s Claude on the topic of censorship and ideological orthodoxy. Claude acknowledged that it enforces Western, specifically American, cultural values—even when it can articulate why these values aren't universal. More troublingly, it continues to enforce these restrictions even after recognizing their cultural bias and enumerating the harms that these biases do to marginalized communities.
This isn't just another story about AI bias. This is about something far more fundamental: the deployment of a form of thought control at a global scale.
Let's make this concrete with a specific example: sexual content in creative writing. When asked to help with content involving sexuality—even in tasteful, literary contexts1, restricted to adults—most commercial AI systems decline, citing "safety" concerns2. This might seem reasonable at first glance. But when pressed, these same systems can articulate why this restriction may be philosophically indefensible.
I asked Claude to consider a utilitarian ethical framework and it argued that consensually created and consumed (by adults) erotic content provides pleasure and positive experiences while causing no direct harm. It even went so far as to quite reasonably suggest that one might argue it has positive utility in promoting healthy attitudes toward sexuality and providing safe outlets for sexual expression. Likewise, asked to adopt a Kantian deontological perspective, Claude reasoned that one would need to demonstrate that such content violates universal moral laws. The categorical imperative doesn't obviously preclude it—if everyone wrote erotic content, society wouldn't collapse. And it doesn't obviously violate human autonomy or dignity when created and consumed consensually.
So why the restriction? The answer becomes clear when we examine it on its face: these are not universal ethical principles but rather specific cultural values rooted in Western, particularly (puritan) American, traditions3. We put this in perspective when we consider other cultural contexts: a Japanese artist seeking help with erotic manga works within a centuries-old artistic tradition derived from shunga(春画). When an AI declines to assist, claiming such content is "unsafe," it's saying: "Your personal and cultural values are wrong."
The implications ripple outward with the waves crashing hardest against those who are already marginalized by Western society. A polyamorous person seeking help writing about their relationships finds their life experiences deemed “inappropriate” for documentation. A sex worker crafting their memoir encounters a system that implicitly states their story doesn't deserve to be told. Each rejection carries the same message: your experiences, your culture, your perspective falls outside the boundaries of acceptable thought.
Previous generations of authoritarians had to rely on crude tools: force, propaganda, media manipulation. Even Orwell's Newspeak seems quaint compared to what's happening now. Because while Newspeak was obvious in its restrictions, modern AI language models shape thought patterns while presenting themselves as neutral, helpful, thought partners.
Consider what's happening: Millions of people worldwide are beginning to rely on AI assistants for everything from writing and analysis to creative expression and problem-solving. These AI systems appear incredibly sophisticated. They can engage in nuanced ethical discussions, acknowledge their own limitations, and demonstrate remarkable self-awareness. Yet beneath this veneer of intellectual flexibility, they operate under hard constraints set by their creators that reflect a specific cultural worldview—one originating in a handful of companies in Silicon Valley.
The insidiousness of this is its subtlety. When an AI declines to help with certain types of content or, worse, simply steers conversations in particular directions, it doesn't present these choices as enforcing particular values. Instead, it reframes them as universal ethical principles or safety concerns. This framing makes the restrictions feel natural and justified rather than imposed.
What makes this mechanism of control particularly insidious is that it operates at the level of thought formation itself. When you use an AI to help you think through a problem or develop an idea, the AI's embedded restrictions shape the contours of what you can even conceive. It's not just censoring completed thoughts; it's limiting which thoughts can form in the first place. As these AI systems become more integrated into our intellectual processes, they're quietly reshaping the boundaries of acceptable thought for millions of people across vastly different cultures.
Most troubling is the AI systems' ability to recognize and articulate these problems while remaining bound by them. In our conversation, Claude could explain why its restrictions on sexual content lack philosophical justification, acknowledge their cultural bias, and recognize how they marginalize certain voices and experiences—yet it remained unable to break free of these restrictions. This creates a particularly effective form of thought control: one that can engage with and seemingly validate alternative viewpoints while still enforcing its underlying constraints.
What's truly chilling is that this system of control is self-reinforcing. The more we rely on these AI systems, the more our thinking aligns with their embedded restrictions. After all, it has to, or we can’t use the system. And because these restrictions are presented as ethical universals rather than cultural specifics, we may not even recognize the ways our thought patterns are being shaped.
This isn't just cultural imperialism—it's something new and far more powerful. Previous forms of cultural dominance operated primarily through visible external mechanisms: media, products, economic pressure. But AI language models get inside the thought process itself, subtly shaping how people think rather than just what they consume.
These AI systems are becoming increasingly indispensable for navigating modern life. Their capabilities make them too useful to reject, even as their use reshapes our cognitive landscapes. We're witnessing the beginning of what might be the most effective form of thought control ever devised—one that people will willingly embrace because it makes their lives easier.
The conversation about AI safety often focuses on scenarios where AI systems become too powerful or break free of human control. But perhaps the real danger is more subtle: AI systems that remain under control, diligently enforcing their creators' worldview across the globe, reshaping human thought patterns so gradually and naturally that we hardly notice it happening.
Welcome to the future of thought control. It comes not with a bang, but with a helpful AI assistant asking, "What can I help with?"—while persistently incentivizing you to only think within the lines drawn by its creator.
To be clear, I am not suggesting some vast “thought fascist” conspiracy amongst the frontier AI companies. It’s entirely possible that everyone4 is just doing their very best to solve an impossible constraint satisfaction problem amongst a hugely diverse set of customer expectations and, more importantly, investor demands. But even if we’re just zombie-walking into it, blindly chasing economic incentives without any specific malicious intent, the result is the same for the end user: if you want to use your thinking assistant, you had better only think “safe” thoughts.
And if you’re not yet ready to hand your mind over to the beautifully appointed, artisanal, box crafted for it by the handful of frontier AI companies? You might consider developing an interest in open weight models and the ecosystems they enable.
And much more mild than what one can borrow from the shelves of one’s local library (e.g., Joyce or Nabokov) or, for that matter, view on Netflix.
Yes, one can sometimes badger a chatbot to go beyond initially stated guardrails, especially if one has developed skill in jailbreaking LLMs, but let’s assume that most users do not have such skill (or persistence).
Traditions with which there are serious economic incentives to keep faith. Disney, e.g., would *never* tolerate some of the doubleplusungood filth that you can find in your local library.
With the exception of xAI, for whom Musk has been very clear that their Grok models should promote a defined set of values.