I believe in humans.
Three years ago, a lot of people stopped believing in humans. Content Cucumber lost 90% of its clients to ChatGPT. They decided a model could do the writing, and they left.
Humans are back. Taking humans back into content is the work now, and what follows is the argument for it.
Here are the rest of the facts, so you can weigh what follows against them.
I use AI every day, and I use it on data. It reads search data at a volume nobody is going to sit through by hand. It finds the queries where a page is sitting one place off the edge of breaking through, the two pages competing for the same term, the vertical that looks flat in aggregate and is climbing underneath. Finding an opportunity inside data that already exists is backward-looking work, and that is what a model is built for.
I do not let it write. That is the line I hold, and the rest of this is why.
I am not predicting what happens when a business hands its content to a model. It already happened to mine.
AI gets you Sideways, not forward
1. A model builds from what already exists
A language model produces text by predicting what comes next based on everything it was trained on. That is the whole mechanism, and it is a good one. What it means in practice is that the output is a rearrangement of what has already been written somewhere.
That direction has a name. It is sideways. The model looks backward across its training and reassembles, which is useful when the job is summarising, researching, or explaining something already understood. It is the wrong tool when the job is saying something that has not been said, because there is nothing in the mechanism that points forward. A person pushes content forward. A model moves it sideways, and no amount of prompting changes the direction of travel.
2. Why does AI content all sound the same?
Because you are all using the same few models, there are a handful of frontier systems, trained on overlapping corpora, and your competitors are prompting them with the sentences you are prompting them with. There are only so many ways to ask for a blog post about your category.
Run that at scale and the whole field converges. Ten companies in one industry publish ten versions of the same page, each of them fine, none of them distinguishable. That is the race to mediocrity, and it is worth being precise about the word. Mediocre does not mean bad. It means the middle. A model returns the most probable next words, and publishing the most probable words is a working definition of the middle of the field.
3. Nothing breaks, which is why you will not notice
For a while it looked unique. People decided ChatGPT was the answer to content creation, and early on you could not tell. Three years in, the patterns have surfaced, and it all looks the same. Em dashes everywhere. If X then Y, over and over. None of it arrives as a failure, though. Your content gets more consistent, typos vanish, every dashboard says it worked, and what leaves is the reason to pick you. Content nobody read does not start getting read because you published more of it.
It leaves fingerprints
4. The tells are detectable, and not only by people
When I see AI-created content, I stop reading. There is no analysis and no second guess about whether the tells are conclusive. Why read something that someone has already written?
The habits are not subtle once you have them. The two-beat reversal: this is not X; it is Y, which sounds like insight and is formatting. The inflated descriptor standing in where a company name belongs. The sentence that announces a point is important instead of making it. We ran our own page through six external detection engines and two of them flagged five instances of the same pattern, with no access to our style guide and no idea what they were looking for. If a machine with no context spots the habit, a reader who has seen four other pages that week spots it faster, and that reader does what I do.
5. Does passing an AI detector mean the content is good?
If it reads well, I will read it. The detector does not get a vote.
An AI detector is only as good as whoever designed it, and plenty of them are not looking for emotion or humanness at all. They count patterns. A pattern count can tell you a machine produced the words. It cannot tell you whether a person is anywhere in them. Ours has the same limit. It missed a pattern that four external engines caught, because the check was hunting for a phrase instead of a behavior.
So a clean score settles less than people think. We scored our own Industries hub. Clean on every machine we test for. Five out of ten on human uniqueness. Competent, readable, and replaceable. No detector would have flagged that page, and any competitor could have published it.
Let’s face it. This shit ain’t going to move the score on an AI tool (Or will it?)
It has nothing of its own
6. Can ChatGPT have an opinion?
Ask a model what it thinks, and you get the balanced summary of what has been thought. That is the design rather than evasion. It has no position to defend and nothing at risk in being wrong, so it hands you the consensus with the edges filed off.
Which means every opinion in your content has to come from a person, and that is where the do-it-yourself plan falls over. The pitch is that one marketing person can generate at scale with a model and still have time to research it, verify it, and edit it. Generation became free. Reading, checking, and deciding did not, and none of them compress. When we scored our own hub, counter-consensus came back at zero out of two until a person supplied a position, and the score went from 5 to 9. Same page, same length. The variable was somebody sitting down and having an opinion.
7. It has never been in the room
I used to believe volume mattered. I ran a content company on that assumption for years, and I was wrong. What matters is the quality of the writing and how human it is, and those are the two things a model cannot hand you.
No model could have corrected me on that. It took years of shipping work and finding out which of it earned anything. That kind of knowledge was never written down anywhere a model could read it, because it sits in one place, inside the head of somebody who was in the room.
8. Why does AI content not get cited?
Because it is new to you and not new to anybody else. You prompted a model, you had not seen that output before, and it read as something. It came from everyone, which makes it new to no one.
Citation goes to what is new. An engine names a source because that source had something the others did not have. Another variation on a thing that has already been cited is not a candidate, it is the fortieth version of a page that earned one once. The 23 scored citation factors put specificity and self-contained passages above 8, and both measure the same underlying question, which is whether there is anything here worth attributing to you.
What it costs downstream
9. Can AI content still rank?
It can rank and still earn you nothing, which is the outcome nobody plans for. A buyer who reads a page that could have come from anywhere closes the tab. An engine looking for a passage worth citing skips it for the same reason.
Ranking is one measurement in the middle of a longer loop. We discover, measure, ask, publish, and measure again, and a ranking number is only useful as an input to the next decision. When search was ten blue links, average content still collected something by sitting in position seven. When an engine picks a handful of passages and names the sources, the middle of the field gets nothing, because being one of forty pages saying the same thing gives an engine no reason to pick yours.
10. Someone still owns it in eighteen months
Generation was the cheap part. The expensive part is the page still being right two years from now, when the rules changed, your pricing moved, and the thing you claimed became untrue.
A model will regenerate the page on request. It will not notice that it needs regenerating, and it will not know which of your claims went stale. That is someone’s job for as long as the page is up, and the person doing it needs to know the field well enough to spot what changed. If you were going to need that person anyway, the question stops being whether to hire one and becomes when.
What we take from all of this
We use data to find the content and people to write it. Discover, measure, ask, publish, measure again. It is a loop rather than a production line, and the data finds what is worth writing while a human makes the final call on whether it gets written.
This article came out of that loop. The search data said the questions people ask about ChatGPT and content carry ten times the demand of the next subject we were considering, so this is the piece that got written. The data found it. A person decided it was worth saying. Then a person said it.
If you want to know where your own content sits, we ran the scoring on ourselves, found a page that was clean and forgettable, and fixed it with one human opinion. We will run it on yours and tell you the number.
What’s in your content?
