Are AI detectors accurate? What a Turnitin AI score really means for your kid

What an AI detector actually measures

No tool can look at a sentence and see where it came from. There is no watermark in ordinary AI text and no signature left behind. What detectors like Turnitin's do instead is score how predictable the writing is — how often each word is the statistically likeliest next word given everything before it. Machine-generated text tends to sit in that predictable band. The detector then converts that score into a percentage that looks a great deal more like evidence than it is.

That distinction matters enormously if you are the parent on the receiving end of an email. "87% AI" does not mean 87% of the essay was written by AI. It means the model's confidence that the text falls in the predictable band. Those are different claims.

So how accurate are they?

Accurate enough to be useful as a flag, nowhere near accurate enough to be used as proof. Vendors publish low false-positive rates measured on clean test sets — long, unedited, single-author documents. Real student work is none of those things. It is short, edited, sometimes co-written with a parent or a tutor, often produced in an unfamiliar formal register because a teacher asked for one.

Two failure directions matter to families:

• **False positives** — honest work that reads as predictable. Formal five-paragraph structures, template-driven lab reports, and the writing of students learning English get flagged disproportionately. Research has repeatedly found non-native English speakers penalized at higher rates.
• **False negatives** — AI-generated text that a student has lightly rewritten, reordered, or run through a paraphraser passes routinely. The kids most deliberately cutting corners are the least likely to be caught.

The combination is the uncomfortable part: the tool is most likely to catch the honest kid with a stiff writing voice and least likely to catch the one gaming it.

Why your kid can get flagged for work they wrote

The most common causes we hear from parents are mundane. Your child wrote to a rubric, so every paragraph has the same shape. They used a graphic organizer the teacher provided. They were told their last essay was too casual, so this one is stripped of personality. They used Grammarly's rewrite suggestions. They researched with AI legitimately, absorbed its phrasing, and reproduced the cadence from memory.

None of those are cheating. All of them raise a predictability score.

What to do if your child is accused

Stay off the detector's turf. Arguing about the percentage puts you in a debate you cannot win and implicitly accepts the number as the thing being judged. Argue about process instead.

• **Pull the version history.** Google Docs and Word both keep it. A document that grew over three sittings with revisions, deletions and typos looks nothing like one pasted in whole.
• **Collect the process artifacts.** Outlines, notes, browser history, the photo of the sticky-note plan on the kitchen table.
• **Ask for the school's policy in writing.** Specifically: is a detector score sufficient on its own to substantiate an academic integrity finding? Most policies say it is not.
• **Have your child explain the argument out loud.** Not defensively — for you first. A child who can explain why paragraph three follows paragraph two wrote paragraph three.
• **Request a conversation, not an appeal, first.** Most teachers are looking for reassurance, not a conviction.

The habit that prevents all of this

Keep the process visible before anyone asks. Drafting in a document with version history on, keeping the outline, and writing a one-line note of what AI was used for — "asked it to explain the causes of the war, wrote the essay myself" — costs nothing and settles almost every dispute before it starts. That disclosure habit is the core of our AI homework rules for families.

For the wider question of where the line between help and cheating actually sits by assignment type, read is using AI for homework cheating? and our parent's guide to AI homework help.

How the main detectors differ

• **Turnitin AI writing indicator** — used by most schools, bundled with plagiarism checking. Reports a percentage of sentences it reads as AI. Turnitin's own guidance says the score is an indicator for review, not evidence.
• **GPTZero** — free and widely used by individual teachers. Reports perplexity and burstiness. More sensitive, and correspondingly more false positives on formal student writing.
• **Copyleaks** — sentence-level highlighting, often used in higher education. Highlighting looks precise, but it is the same predictability scoring underneath.
• **Originality.ai** — aimed at publishers rather than schools; tuned to catch AI aggressively, which is the wrong trade-off for a 13-year-old's essay.
• **"Humanizer" tools** — the paraphrasers students use to defeat all of the above. They work, which is the clearest evidence that detector scores measure style rather than authorship.

The email to send the teacher

Short, unemotional, and asking a procedural question rather than defending your child:

Hi [Teacher] — thanks for flagging this. [Child] and I have gone through the assignment together and they've walked me through how they built the argument. We have the Google Docs version history showing the draft as it developed, plus their outline. Before we go further, could you tell me what the school's academic integrity policy requires as evidence in a case like this — specifically whether a detector score on its own is sufficient? Happy to come in and have [Child] talk through the essay with you.

It works because it concedes nothing, offers evidence, and puts the policy question in writing. The full sequence, including what to do if it escalates, is in flagged for AI: what to do.

What the research actually says

Stanford researchers found that detectors classified more than half of TOEFL essays written by non-native English speakers as AI-generated, while near-perfectly clearing essays by US-born eighth graders. Vanderbilt University disabled Turnitin's AI detector for its instructors, citing false positives and the absence of any way to audit a score. OpenAI withdrew its own AI text classifier in 2023 for low accuracy. Those three facts, together, are the honest state of the field: the vendors closest to the models are the least confident in detection.

A faster read than a detector

If you just want to know whether your own child understood the work in front of you, a predictability score is the wrong instrument. Paste the piece into the Homework Check on our homepage and it will tell you, in plain language, what stood out and — more usefully — the two or three questions to ask your kid tonight. It never assigns a percentage, because a percentage was never the point.