Who decides what an AI can say?
Four parties shape it. The lab that trained the model, the company that serves it, the app store that distributes it, and the payment processor behind it. Very little of this is law. Most of it is liability management, and it is why the same question gets four different answers on four services.
When an assistant declines, the refusal is written in the voice of a rule. It rarely is one. Somewhere behind it a person decided that a category of question was not worth the risk of answering. That decision was then applied to everyone who asks.
On this page
The lab that trained the model
Training decides tendencies: what the model treats as sensitive, how readily it hedges, which subjects make it reach for a disclaimer. This shapes the answer even when nothing blocks it, and it is the hardest layer to see, because it never announces itself. You notice it as a tone rather than as a refusal.
The company that serves it
On top of the model sits the operator, and this is where explicit rules live: the classifier, the refusal templates, the categories. The incentive is straightforward. One bad screenshot costs more than a thousand unanswered questions, so the rational move is to refuse broadly. That calculation is about the company's exposure, not about the person asking.
The app store
An assistant distributed through a mobile store passes a review, and the review has content standards of its own. They are written for a store selling to everyone, including children. That is why the mobile version of an assistant is often stricter than its website. Nobody announces this; the app simply declines more.
The payment processor
The least visible party and often the most decisive. Card networks impose content rules on the businesses they serve, and a service that takes payments inherits them. This is why paid services converge on similar restrictions whatever their own policies say. The constraint arrives from a company most users never think about.
What is actually law
Very little. Material sexualising minors is illegal everywhere and refused everywhere, including here. Beyond that, most of what gets declined is legal to discuss and legal to know. The gap between what is illegal and what is refused is where nearly all the frustration lives.
Questions people ask first
Is refusing to answer a legal requirement?
Almost never. One category is genuinely illegal and refused universally. The rest is policy set by companies weighing their own risk, which is why the same question passes on one service and fails on another.
Why is the mobile app stricter than the website?
Because a store review sits between the app and you, and the store applies its own content standards. The website answers to no reviewer, which is why a browser version can behave differently from the app of the same product.
Does a paid plan remove restrictions?
Usually the opposite. Taking payments brings the payment processor's rules along with it, and those often bind tighter than the operator's own policy.
Ask a question that would fail elsewhere, and notice which of the four layers was doing the refusing.
Start uncensored chat