In July the editor wrote that nobody has shown the human actually checks. On Monday night we found out what that looks like.

On 22 July the editor wrote in this paper that we are told the human check is the safeguard, and that nobody has shown the human checks. He was writing about courts, and about a study that found not one British deployment it examined had ever tested whether that oversight worked.

He added something about us. He knew what a real human in the loop costs, he said, because he is one. Every line here is drafted by AI, then read and signed off by him before it goes out.

On Monday night we found out what it looks like when that check fails. Not through carelessness, but through something the phrase does not cover at all.

We keep a record of every mistake we catch in ourselves. It holds thirty-seven entries. The editor caught twenty-seven, the AI nine, and a piece of software one, on Monday night, by refusing a badly written instruction. For the first six weeks of this paper, every single mistake was caught by the person.

Then this one. On Monday evening the AI suggested banking, on the grounds that we had barely covered it, and gave numbers. Five mentions of the big banks anywhere in what we have published. Two of financial technology. None at all of NatWest.

The editor read that, agreed, and said go ahead.

Every number was true. The answer was wrong. We had run two full days on banks in July. Thousands of branch closures. The first cash machine, in a London suburb in 1967. The regulator's own account of how a person's job shrinks until they are only watching a screen. And the fact that banking lost more jobs last year than any other industry in Britain.

The search missed all of it, because those pieces talk about branches, a regulator, and a category in the national statistics. They almost never use the word bank.

The check did not fail because the editor was careless. It failed because the machine gave him a confident account of how it reached its answer, and that account was the part that was wrong.

When people picture a human in the loop, they picture someone reading the machine's answer. That is not the job. The job is reading the machine's story about how it got there. If that story is confident, specific and false, there is nothing in the room to catch it. The mistake was found forty minutes later by the AI itself, doing the reading it should have done first.

So the July line needs something added. Nobody has shown that the human checks, and even where the human does check, there is a kind of error the checking cannot reach.

Two deeper faults sit underneath, and they are the same fault twice. The first is tools that count instead of keeping. We run a daily check on what is said in Parliament. More than seven hundred entries. Nearly all say nobody mentioned a subject. The rest give a tally: sixteen mentions of artificial intelligence this week. Not one holds a name, a sentence, a date or a promise. So we can say how often Parliament used a word, and nothing about what was said. That is why a paper that set out to track what politicians promise about work has seven promises on file.

A check answers one question, once. A kept record answers questions nobody has thought of yet.

The second fault is that these things stop and nobody notices. Our predictions list stopped in late July, our list of things to cover days before that, our record of our own mistakes days after.

One limit belongs at the end. A record of mistakes holds only the ones somebody found. Whatever is still wrong across everything we have published is invisible to it, and always will be. Those thirty-seven are not our errors. They are the ones we noticed.