AI Is Not Neutral, and Pretending Otherwise Is the Problem
Bottom line: every AI model reflects a value system, whether that value system comes from a government’s law or from a training process. The honest move is naming that plainly. The dishonest move, the one causing most of the actual damage in this debate, is any vendor claiming their model is simply neutral. It is not, and it was never going to be.
What we’re seeing
This series has already covered one version of this problem: Chinese models echoing state-aligned narratives on sensitive political questions, a pattern the U.S. government’s own testing found roughly four times more often than in American models. That’s one axis of non-neutrality, and it’s the easier one to talk about, because it traces back to actual law. China’s models are shaped by the 2017 Cybersecurity Law and subsequent content-ecosystem regulations, with real, codified penalties behind them.
Here’s what doesn’t get talked about nearly as much: Western frontier models carry their own measurable lean, and the evidence for it is a lot stronger than the “it’s just a talking point” dismissal suggests. This isn’t one study. It’s roughly a decade of research, replicated across dozens of models and multiple countries. Testing across twenty-four models found twenty-three exhibiting a left-of-center lean. Separate research found ChatGPT’s default responses aligning with the Democratic Party in the U.S., with Lula in Brazil, and with Labour in the U.K., a genuinely interesting result because it means the lean isn’t just an artifact of American training data. It shows up as the local left-of-center position in each country tested. Interestingly, a Stanford study found that both self-identified Republicans and Democrats perceive this same slant, though Republicans perceive it more strongly, which tells you it isn’t purely a partisan mirage.
I want to be precise here, because overstating this would be its own kind of hype, and this series has tried not to do that anywhere else. The lean isn’t uniform across every issue. Some research found models leaning left on environmental and welfare questions but right-leaning on law-and-order questions. And when people push back directly on a model’s position, the model holds its original answer only somewhere between 58 and 95 percent of the time depending on which model you’re testing, meaning this is real but not rigid.
Where this comes from matters more than whether it exists
The more useful question isn’t “does the bias exist.” It’s “where does it come from,” because the answer determines whether this is fixable or fundamental.
The strongest explanation I’ve seen comes from a 2026 paper by Thorsten Hagendorff, and it’s a genuinely more interesting argument than “the engineers are biased.” His case is that the alignment process itself, training a model to be harmless, helpful, and honest, encodes assumptions that correlate more naturally with progressive moral frameworks: harm avoidance, inclusivity, fairness, than with conservative ones. Not because someone sat down and decided to build a left-leaning model, but because “don’t cause harm” has to get defined somehow, and it’s getting defined using training data drawn overwhelmingly from English-language, Western internet text.
That distinction matters enormously for what you do about it. An emergent property of how harm gets operationalized is a correctable engineering problem. A conspiracy of individual developers pushing their politics is an accusation you can’t really fix, only argue about forever. The evidence points toward the first explanation, and here’s the detail that seals it: newer model generations show measurable softening of this effect over time. That’s exactly the pattern you’d expect from something developers are actively working to correct. It’s exactly the pattern you would not expect if the bias were a fixed, structural feature baked permanently into how these systems work.
The comparison that clarifies things
Set the two side by side and the difference in kind, not just degree, becomes obvious. China’s censorship is compelled by law, with real penalties for a company that doesn’t comply, and it shows up as a stronger, more consistent pattern in testing. The Western lean is shaped by training data composition and the demographics of the humans doing feedback review, with no legal mandate behind it and no penalty for a company that chooses to correct it, which is precisely what the softening over model generations shows happening in practice.
That is a real and meaningful difference. It should not, however, be read as evidence that Western models are neutral, because they demonstrably are not. They’re non-neutral in a more diffuse, more correctable way than a state-censored model, and a serious person should be able to hold both of those facts in view at the same time: real bias, and a real difference from state-compelled censorship, without collapsing one into the other or using one to wave away the other.
Where this matters for you
Net: “neutral AI” is a marketing claim, not a technical description of anything that currently exists. Every model you or your business relies on carries some value system baked into how it defines harm, fairness, and acceptable answers. Concurrently, as referenced above, when “challenged” many of the models will recompile/reassess and offer a potentially different answer. I have witnessed this myself on numerous occasions when dealing with less constrained topics. The question worth asking isn’t which vendor is unbiased, because none of them are. It’s which vendor is honest about where their model’s assumptions come from, and which ones are actively working to correct for it rather than pretending the problem doesn’t exist.
That’s a due diligence question, not a political one, and it applies whether the model in question sits in Beijing or San Francisco. The full whitepaper this series is drawn from, “The Superintelligence Reckoning,” walks through the full academic record behind this argument and the more fine-grained distinction between two failure modes that get conflated far too often in public debate. If you want the deeper case, that’s where it lives.