English writing naturalness

Fix the Both-Sides Structure That Makes AI Essays Obvious

AI essays do not read as AI because of their vocabulary. They read as AI because they refuse to pick a side.

We ran the experiment. Every trick you have been told makes an AI draft sound human — fancier vocabulary, more transition words, mixed sentence lengths, a first-person anecdote, a couple of deliberate fragments — barely moved the detectors. Here are the measured numbers, lower meaning more human: generic AI baseline scored 0.9999 and 0.9999. Transition-word bloat: 0.9999 and 0.9999. Fancy thesaurus vocabulary: 0.9992 and 0.9999. Mixed sentence lengths: 0.9996 and 0.9999. One concrete first-person detail: 0.9941 and 0.9994. Deliberate fragments and hedges: 0.9902 and 0.9993. Deliberate short-long rhythm: 0.9933 and 0.9956. Read that column again. The words are not the problem.

Method, because you should not trust a number you cannot interrogate. We took one AI-generated argumentative draft as the baseline. Each variant changed exactly one thing and nothing else. Then every version went through two independent AI-text detectors, both reporting prob_ai, where lower means more human. Two detectors, not one, because a single tool is a single opinion. We did not retune the draft between runs and we did not pick the best of five. The numbers above are the numbers we got, including the ones that made our own list of writing tips look silly.

So if vocabulary is not the tell, what is? The architecture. Ask an AI for an argumentative essay and you get a perfectly symmetrical building: a paragraph for the motion, a paragraph against it, identical floor space, and a closing line that says the truth lies somewhere in the middle. That shape is the giveaway. Real argument is not symmetrical — it is a claim with a spine and a counterargument bent around it. Your reader feels the symmetry before they can name it. The essay never takes a risk, so it never takes a position, and a position is the only thing an argumentative essay is actually for. Balance is not fairness. Balance is symmetry, and symmetry is the tell.

Here is what that symmetry sounds like on the page, and what it becomes when you finally commit. ❌ 'On the one hand, remote work gives employees flexibility. On the other hand, employers lose oversight. Ultimately, the impact depends on how it is managed.' ✅ 'Remote work did not fail. Management did. Every complaint about lost culture traces back to a manager who never wrote down what he wanted.' ❌ 'While some argue standardized testing ensures fairness, others contend it penalizes creative students. Both sides have valid points.' ✅ 'Standardized testing is not unfair because it is standardized. It is unfair because it rewards the one skill every prep course can sell, and then calls that merit.' ❌ 'In conclusion, both sides have merit, and the truth lies somewhere in the middle.' ✅ 'Both sides do not have merit. One has the better argument. The other has the better lobby.' Same facts. Same word count. One of them has a pulse.

The fix is not a polish. It is a re-architecture, and it takes about fifteen minutes. State your side in the first paragraph, in one sentence a tired TA could underline. Demote the counterargument — it is not a co-equal section, it is a clause inside your own argument, and you answer it in the same breath. Delete every 'on the one hand'; if you need a transition, use a fact instead. End on your claim, not on a truce, because the last sentence of an argumentative essay should be the sentence your opponent wants to argue with. Then check the shape: if your two body sections are the same length, you have written a seesaw, not an argument. easydue exists for exactly this pass — you keep the position, it rebuilds the frame.

None of that makes the small stuff worthless. Look at the two variants that moved furthest on the first detector: deliberate fragments and hedges at 0.9902, and deliberate short-long rhythm at 0.9933 and 0.9956. Still high — but they moved, because a human decision was behind them. A fragment you chose is information; a fragment the model produced is noise. That is where easydue does its work: it takes the draft you already wrote and rebuilds the argument so it commits instead of hedging. On our raw baseline, the strong easydue rewrite moved the first detector from 0.9999 to 0.2202 and the second from 0.9999 to 0.336. One experiment, one draft. Not a promise about yours.

Two honest caveats, because you get graded on the essay, not on the numbers. First: these are measurements from a single controlled experiment on a single baseline draft, run on one date. Do not read them as a guarantee about any other text. Second: the structural fix is also the one that improves your grade. A marker who reads a hundred balanced essays a week is not scanning your vocabulary — they are hunting for the student willing to be wrong in public. So before you run your draft through anything, open it and count the 'on the one hand's. If you find more than zero, you already know what is broken. Balance is not fairness. Balance is a decision you have not made yet.