Part six of the series AI Content: 10 Strategies That Work. A new part every Friday.
The most useful way to think about the division of labour between a human and a model is not philosophical. It is operational: what the model does faster and just as well, and what it does worse or not at all.
What a model does quickly and reliably
A first version of a structure. It does not have to be good, it just has to exist, because editing is easier than starting on a blank page. Ten variants of the same idea so you can pick the best. Converting a text from one format to another, from an article into an email, from notes into a summary.
Then summarising a long source document, checking terminological consistency across a document, and finding the places where a text contradicts itself. Those are tasks about working with text as material, and models are reliable at them.
And finally ideas that you then narrow down. The model generates thirty angles, twenty-five of which are usable only as filler. The remaining five contain one you had not thought of, and that one pays for the exercise.
These tasks share one property: you can judge the result immediately. You can see whether the structure is logical or the summary accurate, and you do not have to verify it anywhere.
What a model cannot do
It has no way of knowing what happened in your field last week unless you tell it. It has no opinion anyone will be held responsible for, because responsibility only rests with someone who has something to lose. It does not know the numbers from your own operation.
And above all it cannot decide what to cut. That point is the most underrated of all. Editorial work is ninety per cent deletion, and language models are bad at it, because their default tendency is to add.
Tell a model to „improve this text“ and you get a longer text back. Tell it to „cut this in half“ and it usually obeys, but it cuts precisely the specific details that made the text worth reading and leaves the general framing.
Cutting is therefore the work you have to keep. It also happens to be the part that turns an average text into a good one, so it is not a bad division.
Four ways of working together that hold up
The first is that a person sets out a concept and the model develops it. Suitable when you have a clear thesis and need volume. The risk: if the concept is weak, the model will not fix it, it will only inflate it. A weak idea at three thousand words is worse than a weak idea at three hundred.
The second is that the model generates variants and the person selects. Suitable for headlines, angles and structure. Models are genuinely useful here, because generating twenty options costs a minute and the human brain selects better than it invents from nothing.
The third is the model as an opponent. A brief along the lines of „find the three strongest objections to this argument and do not go easy on me“ gives better results in practice than „improve this text“. A model tends to agree, so you have to tell it explicitly not to.
The fourth is the model as a translator between fields. A brief like „describe this marketing problem the way an epidemiologist would“ produces nonsense some of the time and a genuinely new angle some of the time. The ratio is roughly three to one in favour of nonsense, which is still worth it, because producing that one idea costs a minute.
Why AI text detectors do not work
That is not an opinion, it is documented. OpenAI launched its own classifier in January 2023 and shut it down in July of the same year, citing „its low rate of accuracy“. Its own published numbers: it correctly flagged only 26 % of genuinely AI-written texts and labelled 9 % of human texts as AI.
Worse, the errors detectors make are not random. A study published in Cell Patterns in July 2023 tested TOEFL essays by non-native speakers of English. The average false positive rate was 61.22 %, and at least one of the tested detectors flagged 89 out of 91 essays as AI. For essays by American eighth-graders, meaning native speakers, the false alarm rate was 5.19 %.
A non-native speaker was therefore accused more than eleven times as often. For any company publishing in English as a second language, that is not an academic detail but a concrete risk if someone runs their texts through a detector.
Turnitin promised under 1 % false positives at launch, and in June 2023 its chief product officer conceded that at sentence level it is around 4 %. A more interesting detail comes from their own blog: 54 % of falsely flagged sentences sit directly next to sentences that really were written by AI.
The detector therefore fails most at exactly the working method this whole part describes, a text where a human and a model take turns. The practical consequence is twofold: do not try to get past a detector, and do not use one as a criterion of quality. The criterion of quality is whether the text contains something that is not available elsewhere.
What to add so the text is worth reading
Your own numbers from your own operation, even small ones. You do not need data from thousands of projects, a sentence like „across our eighteen clients last year it worked out like this“ is enough. It is specific, verifiable and nobody else has it.
A specific name, place and date instead of a general situation. A description of what did not work, because that appears in content least often and reads best. And an opinion where it is clear who holds it, because an anonymous opinion commits nobody.
The best test is to look for sentences that could stand in anyone else’s article. Go through the finished text and cut or replace them. What is left is your added value, and if nothing is left, that is useful information about whether to publish the text at all.
Previous part: Visibility in the year people stopped clicking. Next Friday: Images, and how to spot a marketing myth.
The whole series: AI Content
1. Why most advice about AI content no longer holds · 2. A publishing plan that assumes volume does not work · 3. The prompts that make a difference, and the fact-checking · 4. Personalisation that is more than a different salutation · 5. Visibility in the year people stopped clicking · 6. Where AI ends and a human begins · 7. Images, and how to spot a marketing myth · 8. Measurement that tells you something · 9. One piece of content, several channels, without copy-paste · 10. Law and ethics, specifically, as they stand in August 2026 · 11. Build a system, not a pile of articles · 12. What to take from the whole series