AI and writing
How do AI detectors work, and are they accurate?
By the EssayMate team · Updated November 4, 2026
AI detectors estimate whether a piece of text was written by a language model. Schools use them, students worry about them, and they are often treated as more certain than they are.
This post explains how they work, what the evidence says about their accuracy, and what to do if your own writing is flagged.
How AI detectors work
Most detectors are classifiers: models trained on large collections of human-written and AI-written text to learn the statistical differences between them. Given a new text, they output a probability that it was generated, often per sentence, combined into a document score.
- Perplexity: how predictable the words are to a language model. AI text tends to choose likely words, so it scores as more predictable.
- Burstiness: how much sentence length and structure vary. Human writing usually mixes short and long sentences more than AI text does.
- Learned patterns: frequent phrases, transitions and structures typical of particular models.
None of these is proof. They describe how text usually looks, and plenty of human writing is predictable and evenly structured, especially formal academic prose.
How accurate are they?
Accuracy varies by tool, text length and how the text was produced, and published results are mixed. Several findings are well established:
- False positives happen. Human writing is sometimes flagged as AI, which is the most serious error for students.
- Non-native English writers are flagged more often. A Stanford study found popular detectors misclassified a large share of essays by non-native English speakers as AI-generated, because simpler, more predictable phrasing reads as "AI-like" (Liang et al., 2023).
- Editing changes the score. Paraphrasing, mixing human and AI text, or revising a draft can lower detection sharply, so scores are not reliable evidence either way.
- Even AI developers found it hard. OpenAI retired its own AI text classifier in 2023 because of its low rate of accuracy.
For these reasons several universities, including Vanderbilt in 2023, turned off Turnitin's AI detection, and most guidance to instructors says a detector score should never be the sole basis for an academic misconduct case.
Does Turnitin detect AI?
Turnitin has offered AI writing detection alongside its similarity check since 2023, and many institutions have it switched on. It reports an estimated percentage of AI-written text, which instructors see; Turnitin itself describes the score as an indicator to start a conversation, not a verdict. Whether your institution uses it, and how, is set locally.
What to do if your work is flagged
- Stay calm and ask what evidence there is beyond the detector score.
- Show your process: notes, outlines, drafts and version history (Google Docs and Word keep it).
- Explain your sources and talk through your argument; being able to discuss your work is strong evidence you wrote it.
- Know the policy: read your institution's rules on AI and on appeals.
Keeping drafts and notes as you work is the best protection, whether or not you use AI tools at all.
If you use AI tools
Follow your course's rules and disclose AI help where required. Use AI for what it is good at, research help, structure, feedback, and keep the argument and the final wording your own. Edited, sourced, genuinely revised work is both better writing and easier to stand behind.
Frequently asked questions
Can AI detectors be wrong?
Yes. They produce false positives (human text flagged as AI) and false negatives, and their scores change with editing. A score is an estimate, not proof.
Why was my own writing flagged as AI?
Formal, predictable phrasing and very even sentence structure can look AI-like to detectors, and writers working in a second language are flagged more often. Keep your drafts and notes to show your process.
Which AI detector is the most accurate?
None is reliable enough to prove authorship on its own. Results vary by tool, text length and editing, and independent tests disagree.
Do humanizers beat AI detectors?
They reduce the patterns detectors look for, but no tool can guarantee a result, and the most reliable approach is to edit the writing substantially yourself.
Keep going