How to humanize an AI-written essay

Before and after

Four sentences of near-identical length became three of varying length. Abstract nouns became a named group and a concrete finding. The paragraph now advances an argument rather than announcing that an argument exists.

Why essay drafts read as machine-written

It is rarely the vocabulary. Model output uses ordinary words; what gives it away is the shape of the prose, and essays show it more clearly than any other format because an essay is long enough for a pattern to establish itself.

Three things recur. Sentences converge on a similar length, usually somewhere between eighteen and twenty-five words, so the paragraph develops a metre that never breaks. Nouns stay at the top of their abstraction ladder — researchers, studies, society, impact — where a person writing from knowledge would have reached for a name, a year or a number. And the joins between ideas get handled by a small stock of transitions, so furthermore, moreover and in conclusion do work that the argument should be doing.

None of this is bad writing exactly. It is unmarked writing: prose with the fingerprints buffed off. That is also why the fix is not to make the essay sound casual. An essay is meant to sound like an essay. It is meant to sound like a particular person arguing something specific, which is a different property from being informal.

Rewrite the paragraph, not the sentence

The common mistake is to work sentence by sentence, swapping words for synonyms. That preserves the rhythm problem exactly — you end up with the same uniform metre in slightly different vocabulary, and it reads worse, because now the word choices are odd as well as flat.

Work at paragraph level instead. Read the paragraph and ask what it is actually claiming, then write that claim in a short sentence. Follow it with the evidence in a longer one. If a third sentence is not adding anything the first two did not, cut it. The variation in length comes out of the structure of the argument rather than being applied to the surface afterwards.

The worked example above went from four sentences of near-identical length to three of markedly different lengths, and it got shorter. That is the usual direction. Drafts of this kind are padded by construction, and the fastest route to prose that sounds human is to delete the sentences that were only announcing that an argument was coming.

Citations are the part that breaks

This is the failure worth being careful about, and it is specific to academic work. A rewrite changes the words around a citation while leaving the citation itself alone — which sounds safe, and is not, because the claim the citation was attached to may no longer be the claim you are making.

A concrete version: your draft says a study found a correlation, you tighten the sentence into something more assertive, and now the reference is supporting a stronger statement than the paper made. Nothing flagged it. The citation is still formatted correctly and still points at a real paper. You have simply misrepresented it, which is a more serious problem than an awkward sentence.

So after any rewrite, read each cited sentence against the source it came from and check that the claim still matches. Do this before you check anything else. A tool cannot do it for you — ours cannot — because it requires knowing what the paper said.

What a detector score on an essay tells you

Less than the number implies. Detection tools disagree with each other on identical text, change without notice, and are measurably harder on some writers than others: seven detectors misclassified 61.22% of essays by non-native English speakers as AI-generated in one peer-reviewed study, and OpenAI withdrew its own classifier after it identified 26% of AI text correctly.

That cuts in an uncomfortable direction for a page like this one. A lower score after rewriting does not prove the essay will be treated as human, and a high score does not prove it was written by a machine. What a sentence-level view is genuinely good for is showing you where a reader would find the writing flat, because the passages that read as machine-written to a detector are usually the passages that read as lifeless to a person.

Use it as an editing aid and not as a verdict. If you want the detail on how these tools work and where they fail, we keep a set of pages on exactly that, including the ones that undercut our own category.

Start with how accurate AI detectors actually are, then why false positives happen. If you have already been accused, this is the page for that.

Formats with the same problem

Dissertation chapters — the same argument problem at ten times the length, where the joins between studies are what fail.

Discussion posts — 250 words with nowhere to hide, so one specific detail fixes the whole thing.

Questions

Will rewriting an AI essay get it past Turnitin?

We do not make that claim and would not trust anyone who does. Detectors change without notice and disagree with each other on identical text, so nobody selling a rewriting tool — us included — can observe or control what one will say about your document. Rewriting for rhythm and specificity tends to lower detection scores, because that is what these tools read as human. It is not a guarantee.

Does rewriting fix plagiarism?

No, and the two problems are unrelated. Plagiarism is using someone's work without attribution; changing the wording of an unattributed passage is still plagiarism, and paraphrasing to avoid a similarity match is what most academic integrity policies specifically name as an offence. If material came from a source, cite the source.

Do I still have to disclose that I used AI?

If your institution requires it, yes. A rewrite changes the text, not the fact that a model was involved in producing it, and no tool can discharge a disclosure obligation on your behalf. Policies vary widely between institutions and often between departments, so read the one that applies to your course rather than a general summary.

How long an essay can I run through in one go?

The free plan takes 1,500 words in a single scan, which covers a standard undergraduate essay in one pass, with 2,000 words a month. The $18 Writer plan takes 5,000 words per request and 25,000 a month. Splitting a long document across several requests is workable but worth knowing about: each piece is rewritten without sight of the others, so read the joins afterwards.

Will my professor be able to tell?

Possibly, and not usually because of a detector. Experienced markers notice when an essay makes no specific claims, cites nothing precisely, and maintains one register throughout — which are the same properties detection tools respond to. The durable answer is to have something particular to say, which is also the thing no rewriting tool can supply.

What to watch for

• •Keep every citation exactly as it was — rewriting must not touch a reference.
• •Check that a rewritten claim still matches the source you cited it from.
• •If your institution requires disclosure of AI assistance, disclose it. A rewrite does not change that obligation.

Other use cases