Detecting AI-written text is not like scanning for a malware signature. Detectors estimate how statistically predictable a piece of writing is — a metric called perplexity — plus how much that predictability varies from sentence to sentence, called burstiness. Text that reads very smooth and evenly paced scores as "likely AI." The problem is that plenty of strong human writers, particularly non-native English speakers, ESL students, and some neurodivergent students, naturally write in a smooth, evenly paced way. That is the core reason false positives keep happening, and why no serious 2026 policy treats a detector score as a verdict on its own.
How it works
- Perplexity scoring measures how predictable each word choice is according to a reference language model. Lower perplexity, meaning more predictable, skews toward "AI-like."
- Burstiness scoring measures variation in sentence length and structure. Human writing tends to be "bursty," mixing short and long sentences unevenly; AI output, especially unedited, tends to be more uniform.
- Classifier models trained on labeled human-versus-AI text sets add a second layer, spotting patterns statistical measures alone miss — though these classifiers also degrade quickly against paraphrased or lightly edited AI text.
- None of this reads intent or meaning. A detector cannot tell you whether a student understood what they submitted; it can only estimate a statistical resemblance to model output.
- Scores shift by writing task. The same student can score differently on a personal narrative versus a structured five-paragraph essay, since detectors calibrate against typical patterns for each genre rather than against that specific student's baseline.
Why false positives keep happening
Detectors were trained mostly on native-English, neurotypical writing samples as the "human" baseline. Writing that departs from that baseline — because English is a second language, because of a learning difference, or simply because a student writes in short, plain sentences — scores as more "AI-like" even when no AI was involved. Turnitin's own disclosed accuracy figures put false positive rates low in aggregate but non-trivial at the sentence level, which matters a great deal when a single flagged paragraph triggers an academic integrity meeting. Students who use dictation software or grammar-correction tools for a documented accommodation often see the same effect, since those tools also smooth out the natural irregularity detectors are trained to treat as "human." Paraphrasing tools and lightly edited AI drafts cut the other way too, producing false negatives that let genuinely AI-written work pass through clean.
Detector comparison
| Tool |
Approach |
Known weakness |
Typical use |
| Turnitin AI writing detection |
Classifier plus perplexity, LMS-integrated |
Sentence-level false positives |
Higher ed, K-12 districts |
| GPTZero |
Perplexity plus burstiness |
Struggles with heavily edited AI text |
Individual teachers, schools |
| Originality.ai |
Classifier-based, content-team focus |
Built more for publishers than classrooms |
Content and marketing teams |
| Copyleaks |
Combined plagiarism plus AI detection |
Pricing scales fast at large volume |
Schools, institutions |
Common mistakes
Treating a detector score as a final verdict. A percentage score is a probability estimate, not proof. Pair any flag with draft history, in-class writing samples, or a short conversation with the student.
Applying detectors unevenly. Running suspicious-seeming papers through a detector while skipping others introduces bias, often against students who already write differently from the training baseline.
Not disclosing detector use in a syllabus. Students should know upfront that submitted work may be screened, both for fairness and to reduce disputes later.
Ignoring the paraphrase gap. Detectors are noticeably weaker against AI text that has been paraphrased or edited by a second tool. A clean scan does not guarantee no AI was involved.
FAQ
Are AI detectors accurate enough to fail a student on their own?
Most academic integrity offices now say no — a detector flag should prompt a conversation and a look at drafts or process evidence, not stand alone as proof.
Do detectors get fooled by paraphrasing tools?
Yes, regularly. Running AI output through a paraphraser or a second AI pass measurably lowers detection scores without changing the fact that a model wrote the underlying content.
Why do ESL students get flagged more often?
Detectors compare writing against a baseline built mostly from native-English samples. Writing patterns common in second-language English, such as more uniform sentence structure, read statistically similar to AI output.
Should schools drop AI detectors entirely?
Few are going that far. Most are downgrading detectors from a disciplinary tool to a discussion-prompting one, which better matches what the underlying statistics can actually support.
Where to go next
For the classroom policy side of this, see AI for teachers in 2026 and AI for schools in 2026. If plagiarism, not just AI-generated text, is the concern, AI plagiarism checkers compared for 2026 covers the adjacent tool category.