A Free AI Detector Gives a Student a Data Point, Not a Verdict

One thing: how that particular model reads the text pasted into it, at that moment. Grammarly says in its own support documentation that a score from its detector is not a clear indication that a professor will see the same number from a different tool, though it expects the two to point the same way ¹. Two detectors are two models, and they don't answer the same draft the same way.

The free tier narrows the reading further.

Pangram runs a free scan alongside its paid plans, and said in 2025 that free users see lightly AI-assisted text as Human and moderately AI-assisted text as AI-generated ². The free result rounds off the paid one. A Human verdict from a free scan means only that the free display did not show the draft as moderately or fully machine-written, which is a smaller claim than the word on the screen implies.

What a free reading does not carry:

• the vendor-set threshold and word floor of the school's own detector ³
• standing as evidence on its own; the drafts and the writing process are what a student can put in front of a panel
• a record of when the draft was written, which is what a version log holds

The student's route to a Turnitin check hits the same wall from the other side: the tool a school licenses is not the tool in a browser tab.

Because a detector reads patterns spread across a stretch of text, and a short passage carries fewer of them. Turnitin raised its own minimum from 150 words to 300 in 2023, saying a document needs that much before its AI writing detector will evaluate it ³. Pangram sets a lower bar. Its model card says the detector is intended for prose of at least 50 words in complete sentences ⁴.

A 200-word discussion post sits under Turnitin's floor and well over Pangram's.

Length is not the only thing moving the number. Nature reported in 2026 that Pangram and GPTZero score text in segments, and Pangram's founder said a paragraph judged human on its own can switch to AI once it sits inside a longer piece ⁵. Two paragraphs from one document, pasted separately, can come back with two different verdicts, and neither verdict describes the file that gets uploaded.

Checking a piece is not checking the paper.

Human sets a low floor of its own: it will not score anything under 50 words, and a reading on 50 to 149 words is weak by its own account. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Less text carries less signal, whichever model is reading it.

Reread them against the earlier drafts, and check when each one was written. Turnitin puts its sentence-level false-positive rate at around 4%, which means roughly one highlighted sentence in 25 is human-written ⁶. A highlight points at a paragraph, and stops there.

Turnitin's own advice to instructors runs the same way, telling them to use the information to start a conversation rather than to draw a conclusion ⁶. That advice holds for a student reading a self-check, with one difference. The student already knows the answer.

The record is the part worth building before submission.

For each flagged paragraph the useful record is small: whatever version history the document kept, plus the sources behind the claims in it. A student holding only a percentage has a number and an assertion. The one holding the trail has a paragraph with a history, and that is what an instructor can actually check.

Rewriting a flagged sentence until the number falls is the move to skip. It tunes a draft toward one model's idea of human writing. The next detector reads that sentence differently, and what comes out the far end belongs to the tool. That loop has a page of its own: what happens when a detector and a humanizer take turns. The honest version of the same instinct is making an essay read like its writer.

Grammarly draws the line in its own documentation: red and blue underline corrections do not usually move an AI detection score, while its generative writing agents do raise it ¹. A student who accepted a suggestion to make a paragraph sound more academic has changed the text a detector reads, even though the argument and the evidence stayed their own.

The size of that gap has been measured. Asking Grammarly to improve an essay, or to make it sound academic, turned work that read as fully human into work that read as fully machine-written in a 2024 Bloomberg Businessweek test, while spelling and grammar fixes barely moved the reading ⁷.

Same document, same student, two readings.

Marley Stevens, a student at the University of North Georgia, says she was put on academic probation for a year after Turnitin flagged her essay and a disciplinary hearing found she had cheated; she says she used only Grammarly's standard spell-check and grammar features ⁷. The line Grammarly draws is the line her account rests on.

Before a free checker gets blamed for a strange result, the editing history is worth reconstructing. Which tool touched which paragraph, and did it correct a word or write a sentence? That question comes up again after a flag lands, in a draft accused of AI after Grammarly, and it is cheaper to answer now.

The evidence points at non-native English writers. A 2023 study published in Patterns tested seven detectors on 91 TOEFL essays written by non-native English speakers and found an average false-positive rate of 61.3%, with more than half the essays wrongly labeled machine-written ⁸. The students, academics and AI developers Bloomberg Businessweek spoke to said the writers most exposed are likely those with a plainer style, including neurodivergent students and second-language writers ⁷.

A free scan doesn't shrink that risk.

Scoring low on a free checker is not a defense, since the instructor's tool is a different model. Scoring high is not an admission either. The reading describes text. The person who wrote it holds the drafts, the notes and the timestamps a detector never sees. For a writer already carrying that risk, a non-native speaker flagged by a detector covers what the research says about it.

Saving that record costs a few minutes before a paper goes in. Rebuilding it after a flag lands costs a great deal more, and some of it cannot be rebuilt at all.