The honest answer to "how do I write essays that pass AI detectors" is that working around them is the wrong frame. An essay written in your own voice from the first sentence reads human by default because there are fewer patterns for a detector to flag. The patterns detectors learn from AI output (uniform rhythm, furniture transitions, abstract claims, hedged conclusions) are the same patterns generic academic instruction tends to push students toward, which is why even student-written essays sometimes score high. This guide is the five-step workflow that builds an essay structurally resistant to detection, the rhythm exercises that fix the most common false-positive trigger, and an ESL-aware framing for non-native writers facing a documented bias in the average commercial detector.
Five steps in order. The first three are how voice enters the essay, the fourth is how it stays in once the draft is done, and the fifth is verification rather than chasing a number. Skipping straight to an AI rewriter at the end is the single most common mistake; the patterns were already baked in upstream.
Open a blank document and build the outline by hand. Pick the angle, list the three or four points in the order that makes sense to you, and write one short line for each point in your own words. If the essay is on the role of memory in education, your four points might be the meeting fatigue we measured, the assigned reading position, a counter-argument from a classmate, and your own take. Those four lines carry your voice into every paragraph below them. Outlines built by a model produce essays that read AI even after heavy editing, because the structural choices were never yours.
Spend ten minutes on a single sentence. Read it aloud. Cut hedging words ("It could be argued that", "In many ways"). Commit to a position specific enough that another student could disagree with it. A thesis like "memorisation matters more than current pedagogy admits, because skill transfer depends on it" sets the voice for every body paragraph that follows. A hedged thesis like "memorisation has both advantages and disadvantages in modern education" lets the rest of the essay drift toward the assistant default register, where the patterns live.
Each body paragraph needs at least one specific: a named source by page number, a concrete number, an example from a recent class discussion, or a one-sentence personal anecdote tied to the claim. Three specifics in a 1,200-word essay usually move the AI score 15 to 25 points down because models cannot invent them; they synthesise from training data. Specifics also break sentence rhythm, shift vocabulary away from model defaults, and force the prose to read like it came from a writer who knows something.
After the first draft, count words per sentence in each paragraph. If most sentences land between 18 and 24 words, you have an AI tell, regardless of who wrote the essay. Detectors call this low burstiness, and it is the single strongest signal modern models weight. Rewrite half the sentences to be much shorter (five to eight words) or much longer (28 to 40 words with a deliberate clause structure). The contrast is the voice. Real student writing has wide variance; AI-assisted writing flattens it, and that flatness is what the detector scores.
Paste the revised draft into TextSight. Read the sentence-level highlights rather than the overall score. Aim for above 70 on graded essays, above 80 if you want margin. Each red sentence is a candidate for one more manual pass in your own voice. Reserve the AI rewriter in Light mode for the one or two sentences that stay stubborn after a manual rewrite, never for whole paragraphs. The scan is the second opinion; your own voice is the answer.
Detectors score statistical patterns rather than authorship. An essay written entirely by you can still flag if the prose sits in the same neighbourhood as model output. Three structural causes account for most false positives, and each one has a concrete fix you can apply on the next draft.
Modern academic instruction pushes students toward sentences in the 18 to 24 word range with consistent clause structure. That range is also where models default. The overlap is enough for a detector to assign a high probability score even when the prose is original. Variance in sentence length is the cheapest fix; two short sentences and one long sentence in every paragraph moves most essays out of the danger band by 10 to 15 points.
"Furthermore", "Moreover", "It is worth noting that", "In conclusion" are scaffold phrases that RLHF training rewarded models for using. Students who learned formal essay structure reach for the same phrases, and the result reads AI to the detector. Delete furniture transitions on every pass; let paragraph breaks carry the structure. Replace one hedge per paragraph with a direct statement of position you can defend.
"This phenomenon affects many groups" and "Such policies have significant consequences" are model-default sentences because they synthesise from training data without committing to anything. Replace one abstract claim per paragraph with a specific named example, a number, or a personal detail. Three swaps in a five-paragraph essay usually move the score 10 to 20 points, because specificity is the one thing models cannot synthesise reliably.
Each common essay structure leaks a different AI tell. Knowing which signal your format defaults to tells you which sentences to look at first when the highlights come back from the scan.
Introduction, three body paragraphs, conclusion. The skeleton itself is the strongest tell because ChatGPT writes this format perfectly and detectors learn that pattern. Break the symmetry by merging two body paragraphs, opening with a specific claim instead of a thesis preview, or putting the strongest point at the end of the essay rather than the middle. Drop the transition words at the start of body paragraphs entirely; trust the paragraph break.
Claim, counterclaim, rebuttal. The tell is uniform hedging; the assistant softens every claim and every rebuttal with the same register ("It could be argued that", "However, it is also worth considering"). Sharpen the claim sentences. Pick a side and let the prose show that you have. Hedge in your own voice (a personal observation, a specific limit, a named context) rather than the assistant polite default. One paragraph of sharp position-taking shifts the whole essay's voice.
The tell is perfect symmetry; the assistant mirrors every point on subject A with an exactly equivalent point on subject B. Real student writing is asymmetric because you know one side better. Break the mirror by adding one extra paragraph on the side you actually know, skipping a symmetric point that the comparison does not need, or weighting your conclusion toward one side. Symmetry reads AI; asymmetry reads human and reads honest.
All three AI rewriter modes available on every paid plan. Students with a verified .edu email get Pro at $13.99 per month instead of $19.99. Full details on the pricing page.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing →
If your first language is not English, the detector risk is structurally higher because non-native prose patterns overlap with the patterns detectors learned from AI output. The workflow on this page is gentler for you, not stricter, and the scoring is calibrated against a separate ESL sample.
Public detectors are trained mostly on formal academic English written by native speakers and on generated text produced by models that lean toward the same register. ESL prose patterns overlap with that register in ways that produce false positives, and several university studies have measured the average commercial detector at an 18 to 22 percent false-positive rate on human-written ESL essays. TextSight returns roughly 11 to 13 percent on the same sample in our internal evals, which is around 40 percent fewer false positives. Not zero, but a meaningful gap for students whose first language is not English.
The five steps stay the same; the emphasis moves toward steps 3 and 4. Specific evidence and rhythm variance fix the bulk of ESL false positives, because the underlying signal is structural rather than vocabulary. Read aloud is especially useful for ESL writers because the spoken voice catches sentences where the formal academic register collided with non-native phrasing in a way the detector reads as AI. Spend less time worrying about idiomatic vocabulary and more time varying sentence length.
Score-reduction tricks treat a detector as the opponent; voice-first writing treats the detector as a verifier of something true. The second framing is faster, more durable, and produces essays you can defend in a follow-up question. We would rather build the skill than ship the trick.
Essays you wrote yourself. The thinking is yours, the argument is yours, the structural choices are yours. The workflow helps you keep the writing's voice intact through the parts where formal academic register tends to flatten it. The five steps build a skill: the ability to read your own writing the way a detector reads it, then revise. That skill transfers to every essay you write for the rest of your education.
Any detector workaround that works today probably stops working in a year as detectors update. The arms race is asymmetric because detectors only need one signal to flag; AI rewriter tools need to defeat every signal at once. A piece written in your own voice has fewer signals to defeat in the first place. The detector cannot find a pattern that was never there. That is the durable defense, and it does not depend on what models or detectors do next year.
The companion guide for essays you wrote yourself but that still flag because of register or rhythm.
Open the guide →The original-first workflow for any kind of writing, with six voice exercises and the AI rewriter line.
Open the guide →The detector page focused on college coursework, Turnitin context, and the ESL false-positive bias.
Open the detector →The full student workflow, .edu discount, and how to use TextSight inside your academic policy.
Open student page →Sentence-level highlights, ESL-aware calibration, citations preserved. .edu Pro at $13.99 per month.