Welcome to Foxhole Web!

Every complaint about AI I actually agree with is a complaint about how it is used, not about what it is.

The models are extraordinary. They read, they summarise, they explain things patiently at whatever level you need. That part is genuinely new and genuinely useful. Then look at what has been built with them.

What was built instead

Chatbots that answer from memory with total confidence, including the parts they invented, with no way for you to tell which parts those are.

Search results drowning in articles generated to rank rather than to inform, written by nobody, checked by nobody, about a subject the writer never encountered.

Summaries sitting on top of the thing you actually wanted, so you read a machine's impression of an article instead of the article.

Products designed to keep you in the box, answering follow-ups forever, because time spent is the number someone is being paid to raise.

None of that is a limitation of the technology. Every one of those is a decision. Someone chose engagement over usefulness, volume over accuracy, and a confident tone over a checkable one.

The distinction that matters

How it functions is not the problem. How it is used is.

A model that reads six web pages and tells you what they say is doing something useful and verifiable. The same model answering the same question from memory, with no sources, in the same confident voice, is doing something else entirely, and you cannot tell the difference by looking at the output.

That gap between the two is where all the damage happens.

So this is built the other way

Foxhole Web does not let the model answer from memory. Every question runs a live web search first, and those results are the only material the model gets. Every claim in the answer carries a number pointing at the page it came from.

That produces three things I care about:

• You can check it. The source is one click away. If the answer looks wrong, go and find out.
• It cannot go stale. Ask about this morning and you get this morning's pages, not a memory of last year.
• Nothing is hidden. Your question is not sold, not attached to your name, and never used to train anything. There is one labelled advert on a results page, and it has no say in what the answer tells you.

What this does not fix

It does not make the answers true. The model can misread a source. It can cite a page that does not quite say what the answer claims. The sources themselves are sometimes wrong, because people are sometimes wrong.

If you want to help find the places where it falls short, that is what we are looking for.