Humanising AI text

Why model text goes unread even when it is correct

The signal of machine text is not in the vocabulary, it is in the uniformity. Sentences converge on one length, paragraphs come out the same shape, every list has three items, and clauses find a balanced pair of the "not only this, but that" kind. None of these is wrong. The problem is that none of them was chosen either, and the reader feels it before being able to explain it.

Human prose is lumpy. A thirty word sentence arrives and a five word one follows, because the writer was in a hurry in one place and not in another. That unevenness is what tells a reader somebody was sitting on the other side.

Now the most important sentence in this lesson, and it belongs up front: the goal here is not to fool a detector. For two reasons. First, text detectors err in both directions and the question they answer is not the question that concerns you; Google's own position is not about the tool either but about bulk production of low value content, which we went through against the policy text in the lesson on AI content and SEO. Second and more important, a text edited only to get past a tool is just as worthless after the edit. One score changed, and the reader still leaves at the second paragraph.

So the measure in this lesson is one thing and it stays the same through every decision: will somebody read this to the end and take something away. All four moves that follow came out of that question, not out of the feature list of any detector.

Rhythm: the move with the largest effect

This move can be measured, which makes it the easiest of the four. Take one paragraph and count the words in each sentence. If the numbers sit inside a narrow band, your paragraph is uniform, however good its content is.

The remedy is mechanical too. Break a long sentence in the middle. Join two short ones that follow each other. Somewhere in the text, at the place where the important point lands, leave a short fragment. That is all.

Our own content law puts the rule in one line: vary sentence length hard, mix a five word sentence with a thirty word one inside the same paragraph, and a fragment is fine. Occasionally.

There is a trap that gets mentioned less often: unevenness works when it comes from meaning, not from scissors. The short sentence goes where the point is important or a conclusion has landed. Break sentences at random and the text becomes uneven and also incoherent, and the reader notices the second one first.

And one exception: lists. If you are counting four things one after another, it is natural for all four to share a shape and a length. Uniformity in a list is a feature; uniformity in a paragraph is a signal.

Delete the opening that could belong to any article

It has a one sentence test and it almost always works: read the first paragraph on its own. If it could sit at the top of any article on any subject and nobody would notice, your first paragraph carries no information and it has to go.

In Persian these openings follow a fixed pattern, and because everybody copied everybody the list is short. This site's content law bans 22 Persian phrases outright, and our command line tool flags 12 of them automatically. You have read all the examples: "in the world today", "it should be noted", "stay with us", "without a doubt".

Deleting them is easy. The harder part is filling the hole with something that actually carries information. The simple rule is that the first sentence has to say something arguable: an answer, a position, or a number. If your first sentence is only scene setting, you still have a preamble and only its words changed.

The model leans towards these openings too, and the reason is understandable: in the data it saw, texts often start that way. So this is not a model error, it is a collective habit that the model reproduces faithfully.

Replace the generic claim with something specific

"Site speed affects user experience" is a true sentence that says nothing. Anyone can write it, nobody can refute it, and for that reason the reader takes nothing from it.

The move is this: replace every generic claim with one specific case. What thing, on what, how much changed, and above all what did not change. That last part is where the text becomes believable, because somebody who reports everything improving is selling something.

And now the boundary that separates this move from a serious error: if you do not have the specific, do not invent it. A fabricated number does not humanise a text, it only makes it refutable. The right answer when you have no specific case is to write exactly that: say you have not measured it, or say in which conditions you saw it and in which you do not know. "I do not know" is a human sentence and no model writes it without a reason.

A small recipe for when you are stuck: instead of raising the claim, lower it. Go from "speed matters" to "on this one page we saw this one thing". A small checkable claim always reads better than a large uncheckable one.

The thing only you know

The three earlier moves take a text from bad to readable. This one takes it from readable to quotable, and it is the only move no model can make for you.

It comes in four kinds and all four are real: the step that usually goes wrong, the setting nobody documented, the number that surprised you, and the boundary you hit yourself. One sentence of any of them is enough.

The practical test: find a sentence in the text that nobody could have written without having done the work. If there is none, the text is still a summary of what everybody knows, and more editing will not fix it. That sentence belongs in the brief, and if it was left empty there, go back to the brief.

And the question that always comes up here: what if I have no first-hand experience of this topic at all? The honest answer has only two routes. Either shrink the topic until it reaches the part you have actually done, which almost always makes a better piece anyway, or spend half an hour with somebody who has done it and quote them by name. The third route, writing a sentence as though you had lived it, is the one that gets found out once and takes everything else with it.

And a warning that belongs to this move exactly: do not ask the model to write it for you. Say "add a personal experience" and the model will invent one, invent it well, and you will publish a lie under your own name. This is the one place in the text where filling it with a machine is worse than leaving it empty.

What does not humanise a text

The list of things not to do matters as much here as the list of right moves, because most Persian guides recommend exactly these.

Do not insert deliberate typos. It is neither human nor free; a reader who sees one error assumes the rest of the text was written with the same care. Do not swap plain words for rare ones; heavy prose is old, not human. Do not add similes and adjectives; that is the same decoration the model itself proposes when you ask it to make the text "warmer". And do not stretch the text: lengthening an empty piece only spreads the emptiness thinner.

There is one move that looks harmless and is not: handing the text to a model and saying "humanise this". What comes back is the same skeleton in new skin, and worse, in that pass the model adds sentences nobody has checked. If you keep one sentence from this lesson, make it this one: human editing is deleting, not adding.

And the last thing, which in practice works better than all of it: read the text out loud. Wherever you run out of breath there is a long sentence, wherever your tongue catches there is a bureaucratic phrase, and wherever you get bored yourself, the reader left long before.