AI detector

How AI detection works on WritingTools

The detector estimates how much a passage resembles writing that language models often produce. It returns one overall likelihood and highlights sentences that contribute to that score.

Detection can be wrong. Human writing is sometimes flagged, and AI text that someone has edited can look human. Do not treat the score as a final judgment.

What you see

  • An overall AI-likelihood score and a short label.
  • Sentence highlights for the lines that look more AI-like.
  • A link to the AI Humanizer if you want a rewrite. A rewrite is not proof the new text will score as human.

What it does not do

  • It does not identify a person or a model by name.
  • It is not a plagiarism checker.
  • It does not store highlights. They describe one version of the text and clear when the text changes.
Check a passageOpen the humanizer

What the score is estimating

The detector asks a language model to score how much each sentence resembles writing that models often produce: very even rhythm, similar sentence length, empty emphasis, and stock transitions. Those scores are rolled into one overall likelihood and a short label. The number is a probability-style signal. It is not an identification of a person, and it is not the name of a model.

You will see three kinds of sentence marks. At 70% or above, the sentence is highlighted as likely AI. From 35% up to 70%, it is marked mixed. Below 35%, it is left unmarked, which the legend calls likely human. A sentence that could not be scored is labeled as unscored rather than forced into a bucket. The same bands are what the AI detector shows on its own page.

How to read a result without over-trusting it

  1. 1

    Read the overall number second

    Look at the highlighted sentences first. A single stiff sentence can pull a short passage up. A long casual story can bury one generated paragraph. The overall score is a weighted view of those sentences, not a separate test.

  2. 2

    Ask whether the highlight matches the writing

    Formal, highly edited, or instructional sentences are easier to misread. A polished abstract, a policy clause, or a list of steps can score higher than a casual email even when a person wrote it. If the highlight surprises you, read that sentence and decide whether it actually sounds off.

  3. 3

    Edit, then run it again

    Highlights describe one version of the text. They are not stored. They clear when the text changes. A second run after you rewrite is a new estimate, not a correction of the first score. A lower number is still not proof a person wrote it. Edited model text can score low.

What a check does not decide

It does not prove authorship. It is not a plagiarism checker, and WritingTools does not include a plagiarism checker. It should not be the only reason to reject a student’s work or a writer’s draft. Treat a high score as a reason to reread, not as a verdict.

The AI humanizer is a rewrite, not a mode of the detector. Opening it from a result copies the text across so you can change the wording. A rewrite is not proof the new text will score as human, and running the detector again is the only way to see the new score. The humanizer page describes what that rewrite is for. This page is only about how to read the check.

How much text you can check

Free checks cover about 1,200 words. Premium checks longer documents, up to 40,000 words, on the same plan as the other writing tools. There is no separate detector subscription. Past the cap, the check stops and asks you to shorten the paste or upgrade. Long documents are split into chunks of sentences and scored in parts, then combined. That is why a long check takes longer than a paragraph, and why the disclaimer does not get weaker as the document gets longer.

In the editor, you can score the open document or a selection. The highlights sit on that text and disappear when you edit. They are not a saved report.

A practical way to use the check

  • Use it on a draft you are about to publish if you want a second look at rhythm and stock phrasing. Then edit the lines that are actually vague, whether or not they were highlighted.
  • Use it on student work only as one input, next to the assignment, the sources, and a conversation. A score is a weak reason to accuse someone.
  • Do not chase a number by swapping synonyms. If the draft is thin, add a real detail. If it is padded, cut the repeated point. The paraphraser changes wording. It is not a promise about the next score.
  • Short passages move more. A single sentence has less context, so the score swings. Paste the paragraph, not the slogan, when you want a stabler read.
  • If you only needed grammar, skip the detector and open the grammar checker. The two tools answer different questions.

Checks and rewrites that are easy to mix up

The detector scores a passage. The humanizer and the paraphraser change wording. They are not the same button.

AI detectorAI humanizerParaphraserEditorStyles and voicePremium

Longer checks use the same Premium plan

Free detection covers about 1,200 words. Premium checks up to 40,000 words. The score can still be wrong, and there is no separate detector plan.

See Premium

Frequently Asked Questions

No. It means the passage has patterns the checker associates with AI writing. Formal or generic human writing is often flagged.

No. Edited AI text can score low. The score is a signal, not authorship proof.

Free checks cover about 1,200 words. Longer documents are part of Premium, up to a much higher cap. The same disclaimer applies.