It's often said that an AI is biased, as if neutrality were its natural state from which it strayed by accident. This article approaches the problem from the other direction: is a neutral AI even conceivable? To answer that, we need to look at where, exactly, human choices creep into the making of a model. The answer is: everywhere.
The criticism has become a classic. Depending on your leanings, you'll accuse a model of being too cautious or too permissive, too American, too progressive or too conservative. These critiques often have a real basis. But they almost always rest on an implicit assumption worth examining: that an AI could be neutral. Let's see what that would entail.
First entry point: the data
A model learns by absorbing vast amounts of text, as we described in our article on training. That corpus is not the world: it's what humans have written and put online.
It is therefore structurally imbalanced. Massively English-speaking, as we saw even in the way words are tokenised. Over-representing connected countries and populations that write online. Containing the prejudices of their eras, since centuries of texts carry the views of their time.
A model that faithfully reflected this corpus would not be neutral; it would be the mirror of a particular sample of humanity. And a model corrected to compensate for these imbalances would not be neutral either: someone would have decided what to compensate for, and in which direction.
Second entry point: human preferences
Then comes the stage where human evaluators rank the model's responses from best to worst, to teach it to produce what people prefer. That's what gives it its character.
But what does "best response" mean? On a factual question, you can settle it. On a political, economic or moral question, the notion of a best response already presupposes a point of view. Should all arguments be presented on equal footing, including those science considers refuted? Should certain topics be refused? Where do you draw the line between caution and usefulness? Every answer to these questions is a value judgement, not a technical setting.
Suppose we wanted to make a model perfectly balanced on contested topics. You'd have to decide which positions deserve representation, with what weight, and which fall into the margins. But that split between legitimate debate and the margins is itself a position. Neutrality requires an arbiter, and the arbiter necessarily has a point of view. This is a logical problem, not an engineering flaw: it can't be solved with more compute.
Third entry point: refusals
Every model refuses certain requests. These refusals are installed deliberately, and their scope reflects a hierarchy of values: what's deemed dangerous, what's deemed acceptable, what warrants a warning.
These boundaries are not universal. A topic that's mundane in one country is sensitive in another. A harmless joke here is offensive elsewhere. A model that refuses everything would be useless, a model that accepts everything would be dangerous, and in between lies a slider that someone has positioned according to their own criteria. This is the problem we tackled from a technical angle in our article on alignment, but it's also deeply political.
What this doesn't mean
Careful about a misinterpretation. Saying perfect neutrality is impossible doesn't mean all models are equal, nor that we should give up on reducing bias.
There's a considerable difference between a model that reproduces its corpus's stereotypes without filter and a model where serious work has gone into mitigating them. Between a model whose criteria are publicly documented and a model where no one knows how it was shaped. Between a company that acknowledges making choices and a company that claims to make none.
Progress exists, it's measurable, and it matters. What doesn't exist is the end point: a model finally devoid of any perspective.
Transparency over neutrality
If neutrality is unattainable, what goal should we aim for? The most solid answer seems to be transparency of choices.
A model that explains its criteria, publishes its principles, acknowledges its areas of uncertainty and lets you see where it draws its limits is more honest than a model that presents itself as an oracle without a point of view. You can then account for its perspective, just as you account for the editorial line of a newspaper you read.
That's also why an informed approach is not to rely on a single model for topics that matter. Asking two differently trained systems a sensitive question often reveals, through their discrepancies, the choices each has made. That's exactly the cross-checking reflex we recommended in our verification method.
What to take away
An AI is a human artefact, built from human texts, shaped by human judgements, for human uses. At each of these stages, someone chose. Expecting it to achieve a neutrality that no book, no newspaper, no teacher, no human institution achieves is probably a poorly formulated demand.
What you can legitimately demand, on the other hand, is more concrete: that the choices be owned, documented and debatable. That the crudest biases be corrected. And that we stop selling these systems as sources of objective authority, which they are not. The real risk isn't that an AI has a point of view; it's that we use it while believing it doesn't have one.