Why your internet filter does not cover AI

If you have a filter installed — TAG, Gentech, one of the others — you have done something real. Those organisations do careful work, they do it for free in many cases, and a well-configured filter genuinely changes what a device can reach. Nothing here is a criticism of them.

But there is a gap between what a filter does and what conversational AI is, and it is a gap most families have not yet been told about.

How a filter decides

Web filtering works at the level of the address. A request goes out for a page; the filter checks the domain, sometimes the path, against a list and a category; it allows or it blocks. Everything downstream of that decision is a matter of trust in the site.

That model works well for the web, where a site is broadly one kind of thing. A news site is a news site. A shopping site is a shopping site. Categorise it once and the decision holds.

Why AI breaks the model

An AI assistant is not one kind of thing. It is a blank page that becomes whatever the conversation makes it. The same address, in the same session, can produce a homework explanation, a business email, and material you would not want in the house — and from the filter's point of view, all of it is the same single approved connection to the same allowed domain.

The filter cannot read the prompt. It cannot read the response. Both are inside an encrypted connection to a site that has already been permitted. There is nothing further for it to decide.

The filter is not failing. It is answering a question nobody is asking any more: not "which site is this?" but "what is being said inside it?"

The two options a filter actually has

Faced with an AI site, a device filter has exactly two moves, and neither is satisfactory:

• Block it entirely. Clean, enforceable, and increasingly costly — AI is becoming part of ordinary schoolwork and ordinary parnassah, and a household that blocks it wholesale is accepting a real disadvantage.
• Allow it entirely. Which is what most households have quietly ended up doing, usually without deciding to, because a child needed it for a school project once.

Most families are in the second position and believe they are in the first, because the filter is installed and the filter has always been the answer.

What closing the gap requires

The check has to move to where the content is, which means it has to happen on a server between the person and the model — before the prompt reaches the model, and again before the answer comes back. That is not something a device filter can do from outside the connection, and it is not something a browser extension can do reliably either, since anything running on the device can be turned off by whoever is holding the device.

Practically, that means the filtering has to be part of the AI tool itself. Which is an uncomfortable answer, because it means trusting a vendor rather than an organisation you already know — so the questions in the next section matter more than usual.

What to ask before you trust one

1. Is the prompt checked, the answer checked, or both?
2. Does the check run on your servers, or on the device?
3. Can I add blocked topics of my own, or is the policy the same for every customer?
4. What happens when the safety check cannot run — does it refuse, or answer unchecked?
5. Can I see what my child asked and what was declined?
6. Does it cover images and video, or only text?

A vendor that answers all six plainly is worth considering. One that answers with "100% safe" is telling you it has not thought about the fifth question, which is the one that matters when something goes wrong.