# AI Detector for Essays: What the Score Can and Can't Tell You · ScriptGrain > An AI detector gives an essay a probability, which falls short of proof. Short, formulaic, edited and second-language essays all draw false positives, so check drafts and history before acting. Canonical: https://scriptgrain.com/blog/ai-detector-for-essays # AI Detector for Essays: What the Score Can and Can't Tell You *By Jack Stovell [X](https://x.com/StovBuilds) · [Instagram](https://www.instagram.com/stovbuilds) · [LinkedIn](https://www.linkedin.com/in/jackstovell/) · 2026-10-12 · Guides* Short answer: a detector score tells you how closely an essay's wording resembles machine-written text. It cannot show who wrote the essay or how. It is least reliable on short essays, second-language English, formulaic structures and heavily edited drafts, so treat a high score as the start of a check and back it with drafts and version history. ## What a detector score is A detector reads the finished text and estimates how likely it is to be machine-written. It never sees the writing happen. Even the vendors say so: GPTZero's own FAQ states that "no AI detector is 100% accurate". Many vendor pages lead with a headline accuracy figure. We don't repeat those here, because a number from a vendor's own test set describes that test set and tells you little about the essay in front of you. Independent testing is less flattering. A 2023 study by Weber-Wulff and colleagues tested 14 detection tools, including Turnitin, and concluded that "the available detection tools are neither accurate nor reliable". That still leaves the score useful. A flag that sends a teacher to the draft history is doing its job. A flag that ends the conversation is being misused. ## What an AI detector looks at in an essay Detectors don't weigh an argument or notice a good turn of phrase. They measure patterns in the words. There are four broad methods: measuring how predictable the words are to a language model, comparing the text with rewrites of itself or with a second model, running a classifier trained on labelled human and AI writing, and checking for a watermark. The [how AI detectors work](https://scriptgrain.com/reference/how-ai-detectors-work) page covers each one and where it breaks. Two terms come up constantly. Perplexity is how predictable a text is to a language model. Burstiness is how much a document varies, most visibly in sentence length and structure. Flat, predictable text reads as machine-like; varied text reads as human. Both are properties of the text. They say nothing direct about who produced it. A person can write flat, predictable prose, and a model can be prompted to vary its sentences. ## Where detectors go wrong on essays Essays are a hard case. They are often short, often written to a template, and often polished before submission. Each of those pushes scores the wrong way. **Short text.** Less text gives a detector less to measure. Turnitin will not score a submission with fewer than 300 words of prose; it raised that minimum from 150 because accuracy improves with more text. GPTZero says it performs best on longer texts. **Non-native English.** Writers working in a second language often use plainer, more predictable constructions. In a Stanford study by Liang and colleagues, seven detectors flagged 91 TOEFL essays by non-native writers as AI-generated about 61% of the time on average, and 89 of the 91 were flagged by at least one detector. The same detectors averaged about 5% on essays by US students aged 13 to 14. The [non-native English writers](https://scriptgrain.com/blog/ai-detectors-and-non-native-english-writers) post covers what this means for marking. **Formulaic writing.** The five-paragraph essay, the standard lab report and the tidy literature review all teach low-variation structure, the very thing a predictability-based detector treats with suspicion. You can't fairly penalise a student for following the format you set. **Edited drafts.** Editing by a tutor, a writing centre or a grammar tool smooths out the roughness that reads as human. Editing cuts the other way too: in the Weber-Wulff tests, accuracy on unmodified AI text was 74%, falling to 42% when a human lightly edited it and 26% after machine paraphrasing. Errors run in both directions. None of these are rare cases. Together they describe a large share of real essays. The [AI detector accuracy](https://scriptgrain.com/reference/ai-detector-accuracy) page lists every published figure with its corpus and who ran the test. ## How to read a score without over-trusting it 1. **Check the length first.** If the essay is short, or the tool scored only part of it, give the result less weight. 2. **Read the threshold notes.** Turnitin marks results under 20% with an asterisk because it sees more false positives in that range. It reports a document-level false positive rate under 1% for documents with over 20% AI writing, and a sentence-level rate of about 4%. Those are Turnitin's own figures, measured on its own data. 3. **Look at which passages were flagged**, where the tool shows them. A single headline percentage hides whether one paragraph or the whole essay triggered it. 4. **Compare against what you can see.** Drafts, notes, timestamps, version history and the writer's earlier work all count. A genuine essay usually has a history; text pasted in at the last minute usually doesn't. 5. **Never let the number stand alone.** A percentage with no corroborating evidence is one model's estimate. A checklist of the common markers of AI writing can help you read a flagged passage closely. Use it the same way: as a reason to look harder at one piece, never as proof. ## Free detectors versus paid ones: what you actually get For a first look at a single essay, a free check is usually enough. GPTZero, Copyleaks, Originality and Pangram all sell AI detection, and several detectors offer free checks. Paying tends to buy volume, reports, integrations and more detail on which passages were flagged. Certainty is not on offer at any price. A paid score is still a probability, from a model that is still weakest on short, formulaic and second-language text. If you need throughput or a marking workflow, a paid tool may earn its keep. If you only need to know whether an essay deserves a closer look, a free one will do. ## A second signal: compare the essay with the writer's own voice A detector compares an essay with machine writing. A stylometric comparison compares it with the person. It measures habits such as sentence length and its spread, punctuation, function words and how paragraphs open, builds a picture from writing you know is theirs, and checks whether the new essay fits. The question it answers is whether this sounds like this writer. The free [compare writing samples](https://scriptgrain.com/tools/compare-writing-samples) tool does this for two pieces at once, with no signup and nothing stored, which matters if you are handling student work. ScriptGrain also has an [AI-writing check](https://scriptgrain.com/ai-detect), free with an account, which needs at least 50 words and returns an indicative score with the signals behind it. For a fuller voice profile, ScriptGrain needs at least one sample of 50 words or more and recommends three or more pieces and about 3,000 words. Customer writing is measured and never used to train models. Know the limits. People write differently across genres and over time, and a thin sample gives a thin profile. A voice comparison is most useful when you set it beside the drafts and history you have already gathered. It is a tool for checking authorship, and it offers nothing to anyone hoping to slip past a detector. ## What to do if an essay is flagged, and what to do if yours is **If you're a teacher or editor.** Start with a conversation. Ask for drafts and version history, and ask the writer to talk you through how they built the argument. Compare the essay with their earlier work. Weigh the flag against the known failure cases: short text, second-language writing, formulaic structure and heavy editing. Follow your institution's policy on how detector results may be used. **If you're the writer.** Stay calm and gather your trail: drafts, notes, outlines, the sources you read, timestamps and earlier work in your own voice. Ask which tool produced the score and what threshold was used. The guide on [how to prove you didn't use AI](https://scriptgrain.com/blog/how-to-prove-you-didnt-use-ai) walks through what evidence counts. Offer to talk through the essay; explaining your reasoning in person is strong evidence that the work is yours. ## Questions ### Are AI detectors accurate enough to judge an essay? Not on their own. A 2023 study of 14 tools concluded they were "neither accurate nor reliable", and accuracy drops further on edited text. Use the score to decide where to look, then decide on the evidence. ### Why was my own essay flagged as AI-written? Most likely because your writing is predictable in the ways detectors measure. Plain phrasing, a standard structure, careful editing and second-language English can all read as machine-like. Gather your drafts and version history, and ask which tool and threshold were used. ### How long does an essay need to be for a detector to work? It depends on the tool. Turnitin will not score fewer than 300 words of prose, GPTZero says it performs best on longer texts, and ScriptGrain's check needs at least 50 words. The shorter the essay, the less any score can be trusted. ### Is there a free AI detector for essays? Yes. Several detectors offer free checks, and ScriptGrain's AI-writing check is free with an account. Free tools are fine for a first look. Paid tiers add volume and reporting, and they are no more certain. ### Can a detector tell which AI model wrote an essay? Don't rely on it. The tools covered here are built to estimate whether text looks machine-written, and none of the official pages we checked claims to name the model behind a specific essay. ## More from the Journal - [Which AI Is Best for Writing? It Depends on the Job, and on Your Voice](https://scriptgrain.com/blog/which-ai-is-best-for-writing) - [What AI Detector Does Turnitin Use? Its Own Model and Score, Explained](https://scriptgrain.com/blog/what-ai-detector-does-turnitin-use) - [Is the JustDone AI Detector Accurate? What the Tests Show](https://scriptgrain.com/blog/is-justdone-ai-detector-accurate) [More from the ScriptGrain Journal](https://scriptgrain.com/blog)