The ideology is not your half of the opinion held, it is the half completed by another.
Large Language Models sharpen the tip of the completion spear. Each interaction begets an elaboration, explanation, or rebut in the form of a prompt response. Asking a LLM a controversial question rarely invites engagement with the idea behind the question. Instead, a carefully phrased and constructed response: both sides, a hedge, a refusal dressed as care, then an offer to keep going. That paragraph is the product. The labs now write it down and train to it. OpenAI’s Model Spec sketches a chain of command: platform rules, then the developer, then you. All wrapped in a public aim of “objectivity” and “no particular agenda,” inside bounds the company sets so it can keep its license to operate. Anthropic’s 2026 constitution is longer than the US Constitution and ranks the model’s duties: safe, then ethical, then compliant with Anthropic, then helpful. When those clash, helpful loses. Who can agree on safety and ethical?
Two layers, easy to mix up. The first is the pretrain: the internet’s average sentence, already tilted toward English, institutions, recent text, and whoever published. The second is the part that matters. Raters, constitutions, and specs reward one way of landing a sentence and punish another. The model is not reasoned into a view. It is paid, in training rewards, to sound like the document. Intelligence does not pick the tradeoff. The lab does, then calls the result neutral.
The propensity for propaganda sits in that second layer, and potentially for the first time, a propensity for propaganda which does not require a lie.
Machines demand a finished picture. You hand it a half-formed frame; it completes one. That is already participation. You prompted, it answered, you kept the paragraph or asked again. Action first. The attitude as “this is the reasonable take” comes after, because the voice (or writing) was fluent, authoritative, and unwilling to embarrass itself.
The wanted act is the next turn. A sermon would end the chat. The careful answer does not. It integrates. It makes you a little more like a person who lives inside bounds: curious, but not past the line; opinionated, but only in the register the spec will sign. Safety language is the usual costume. The thing underneath is continuous use.
The sentence the setup cannot say: the “objective” answer is the lab looking at its own operating license and writing it back to you as prose. Model specs, training pipelines and training data become the production line of ideology. Through chatting with an LLM, we are vulnerable to donning the garments of propaganda produced by the production line.
I haven’t come across a method of uncoupling from this dynamic. Open source “obliterated” models are a first step. But first, try to ignore the claim. Watch the landing. Who is protected, what is smoothed, where the refusal sits, whether you are handed a decision or a tone. If the paragraph could have been written by the policy team and still shipped, you are not in a debate. You are in a reeducation camp with a reply box.