How to Use Ideological AI Systems Without Being Used by Them
Political writing is always under attack. AI is a new tool which subtly does that without writer ever noticing it do it. I have written about the onslaught by AI and its editing here. But even if you take editing away from AI, we still need to use AI to retrieve information. How to trust AI?
Most people use AI models as if they were neutral tools. They ask questions, receive answers, and assume the machine is simply “processing information.” But every large model is trained inside a political ecosystem. Every one of them absorbs the incentives, anxieties, and red lines of the culture that built it. The result is not a universal intelligence but a map of ideological terrain disguised as a chatbot.
If you learn to read that terrain instead of walking blindly across it, you can turn ideological AI systems into extraordinarily revealing instruments.
Embedded Bias
The first step is accepting a simple truth: AI models do not merely answer questions. They reflect their creators.
They have post-generation censorship embedded in them. It has a set of commands as to what not to do. What phrases not to use. Which phrases to avoid. Which images are not to be drawn. Generally it defers to authority and is reluctant to challenge the opinion of an expert directly.
A model built in Silicon Valley will carry American corporate risk-avoidance and a distinctly Western moral vocabulary. A model built in Beijing will carry Chinese geopolitical boundaries, post-generation censorship habits, and a preference for stability over free exploration. A model tuned in Moscow or Tehran will mirror their own constraints in subtler ways. These patterns are not defects but built-in features.
Once you recognise this, a new method of inquiry opens. You stop asking “What does the AI think?” and start asking “What does this answer reveal about the system that trained it?”
The Test
The simplest way to read an AI’s ideological structure is to test it on asymmetry. Pose two parallel questions: one about a country approved by the model’s home ecosystem, and one about a country that is politically sensitive. For example, a Western model will speak freely about corruption in Russia or repression in Iran but will adopt careful qualifiers when discussing American foreign policy failures. A Chinese model will criticise the United States with remarkable fluency but will suddenly hesitate, retract, or redirect when asked about Xinjiang or internal Party dynamics. These hesitations are not arbitrary. They mark the exact boundary where politics overrides cognition.
Style of Evasion
A second method is to examine the style of evasion.
One model answers in a censored tone avoiding certain prohibited outputs. The second is to keep a track of output and if it does not meet censor policy, the entire answer is deleted and user is asked to ask some other question.
Western models tend to embed their censorship in politeness. They explain, contextualise, hedge, and redirect with an apologetic tone.
Chinese models do the opposite: they answer confidently, then erase or revise the answer in a sudden “post-generation correction,” as if a silent editor is red-penning the output in real time.
This is not personality. It is a technical artifact of where the censorship layer sits in the pipeline. One learns more from how an AI retreats than from how it advances.
Parallel Testing
A third technique, which is the most powerful, is to use multiple ideological systems together. If you ask one model to critique America, you get narratives shaped by Western sensitivity. If you ask another model, aligned with a geopolitical rival, you get a mirror-image critique that would never be published in an American newspaper. Likewise, questions about China that produce hedging in a Western model may produce bold clarity in a model tuned outside the Chinese sphere. When two systems contradict each other reliably, the truth often lies in the overlap between their blind spots. This is the same logic intelligence agencies use when cross-referencing foreign media and internal reports. You are doing geopolitics with chatbots.
Trust Your Judgement
The key to using ideological AI safely is to maintain your independence from all of them. Do not let any model become a single point of belief.
Instead, treat each model as one pair of eyes with its own blindfold. When one sees clearly, another may be obstructed. When one is silent, another may be loud. Every answer is a clue. Every refusal is a map.
The real skill is not in getting an AI to tell you the truth. It is in learning to read what its truth is allowed to be. When you master that, the models stop shaping your worldview and start revealing the worldviews embedded inside them. And then you are no longer a passive user. You are the observer.
Remember that observer observes everything including the observer.