I Skimmed Sixteen AI Articles and Published Every One

Notes from the field

In February I set up a system to write articles for this site. Over six weeks it published sixteen of them, just over thirty one thousand words. I looked at every one before it went live.

Not properly. I skimmed them. They looked fine.

They were not fine.

What a skim actually checks

A model that writes badly is easy to catch. You read two sentences and bin it. That was not the problem here. The sentences were clean. The structure was sensible. The tone was roughly right. Headings in the right places, paragraphs the right length, the occasional decent turn of phrase.

A skim measures fluency. Fluency was never the thing that was wrong.

Every statistic in those articles sat exactly where a real statistic goes, inside a sentence built exactly the way a sourced sentence is built. The shape of authority, without any of the substance. You cannot see the difference at reading speed. That is the entire problem, and it is why I missed it sixteen times.

I spot patterns for a living. Thirty five years of noticing when something is off in a system, and a brain that is uncomfortably good at it. I still missed this, because I was checking the wrong layer.

What was actually in them

I went back through all sixteen properly. Here is the count.

Seventy three numbers. Not one of them attributable to anything. In thirty one thousand words about cyber security and regulatory compliance, published by a firm that asks people to trust its judgement on exactly those subjects.

The duplicates are their own kind of funny. Two articles called The Invisible Apprentice, three weeks apart. Two called The Executive Guide, published a day apart. Nobody noticed, including me, because nobody was reading them. They were competing with each other in search results for an audience that did not exist.

The zero for internal links is the one that tells you what these really were. Thirty one thousand words that never once referred to anything else we do, anything we have built, or any other article on the site. Not content. Filler in the shape of content.

Why I am telling you

Because your people are doing this right now, and they are checking it the same way I checked mine.

Not on a publishing schedule. In a quote that goes to a client. In a board paper. In a policy document. In a summary of a contract that somebody then acts on. In code that gets merged. The output arrives fluent, well organised and confident, and the person who asked for it reads it the way I read mine, which is to say quickly, looking for whether it seems right.

It always seems right. That is what the technology is good at.

The failure mode is not that AI writes badly. It is that AI writes plausibly, and plausible is precisely what defeats a skim.

What we changed

We did not stop using AI. We use more of it now than we did in February, and the business runs on it.

What changed is what review means. Checking whether something reads well is not review, because reading well is the one thing you can rely on. Review means checking the layer the model cannot vouch for: every number has a source or it comes out, every claim about the law is read against the legislation, every recommendation is one we would give on the phone.

That is slower. It is slower in exactly the place where the time was supposedly saved, which is the honest trade and the part nobody selling AI tools wants to put on a slide. The drudgery genuinely goes. The judgement does not, and it cannot be delegated to the thing that produced the work.

It also means somebody owns the output. Those sixteen articles went out under a byline that was not a person. They are now under mine, because I approved them and that is where responsibility sits.

What happened to them

All sixteen are gone. I deleted them rather than leaving them up with a correction at the top, because a correction at the top of a bad article is still a bad article. There is no version of that six weeks worth keeping.

If you followed a link to one of them and landed on the blog, that is why.

Questions I have been asked since