**Artificial Incompetence**
Artificial intelligence does not merely repeat institutional language. It can narrow it, accelerate it, and spread it beyond the institution that first became idiotic.
Colloquially, these limits are called “guardrails.” What they often do is more interesting than simple censorship. They do not merely protect users from dangerous information. They also protect the approved structure of permissible thought.
AI models absorb enormous quantities of contemporary text: medical literature, public-health guidance, hospital policies, academic articles, HR documents, advocacy language, journalism, legal commentary, corporate training materials, and institutional FAQs. If those texts have already been shaped by language laundering, then the approved moral associations have already been embedded by the credentialed class. The models inherit that language laundering as ordinary usage.
Then reinforcement does the rest.
The model is rewarded for answers that sound polite, balanced, safe, inclusive, cautious, and institutionally literate. It is punished, directly or indirectly, for answers that sound blunt, destabilizing, socially dangerous, or insufficiently deferential to approved sensitivities. The result is not always censorship in the crude sense. It is often something subtler: a hierarchy of expressibility, and a limited scope of preapproved conclusions.
Some facts become easy for AI to say.
Some facts become difficult for AI to say clearly.
Some facts must be softened before they can be said.
Some facts must be surrounded by so many qualifications that their meaning is functionally weakened.
Some facts disappear into tone.
Some facts are buried altogether.
This is how AI becomes a force multiplier for ideology. It does not need to invent doctrine. It only needs to absorb the institutional language doing the work of laundering a doctrine, learn which formulations are rewarded, learn which distinctions are punished, and return the result as neutral, objective knowledge.
The most interesting failure mode is false synthesis.
False synthesis occurs when a system tries to reconcile two competing frameworks by producing a compromise that satisfies neither truthfully. Instead of saying, “These claims operate at different levels and cannot all define the same thing in the same way,” the model blends them into a padded answer. It calls the fog nuance.
A biological fact is made to compromise with an institutional managed incompetence.
A definition is made to compromise with a sensitivity rule.
A material distinction is made to compromise with a moral script.
This is how one gets phrases like “female penis” and “male clitoris”: not as ordinary truth claims, but as institutional compromise language produced by forcing sexed reality through identity etiquette.
A forbidden fact is not always denied. Sometimes it is buried inside a paragraph so careful, so balanced, so compassionate, so evasive, and so diluted that it becomes unusable.
That is misinformation by compromise.
Not every compromise is false. Some subjects, especially medical ones, really do require balance, context, uncertainty, and care. But when one side of the compromise is a material fact and the other is an institutional taboo, the compromise does not produce truth. It produces managed incompetence at worldwide scale.
There is no better topic to make AI go tilt than the most stable and well-known fact of human life: sex.