AI Detector Comparison: What Each Tool Reports

No AI detector comparison gives you a verdict on authorship. Each major tool reports a probability estimate for the text you paste, labelled in its own way, and each one will disagree with the others on the same paragraph. Here is what each actually outputs, where the limits sit, and how to combine signals without overreacting to any single score.

What does each major detector actually report?

The useful part of an AI detector comparison is not which one “wins”. It is what lands on your screen when the scan finishes, and whether that output is granular enough to act on.

Tool	What it reports	Granularity	What it cannot tell you	Good for
GPTZero	An overall AI-probability score, plus per-sentence highlighting on supported plans	Document and sentence	Whether a human wrote the sentence with AI help, or a model wrote it cleanly	Spotting which lines carry the signal
Turnitin	An AI-writing percentage for the submitted document, surfaced to instructors inside the similarity report	Document-level segment	Nothing about intent, drafting history or whether AI was used for editing only	Institutional submissions and policy workflows
Copyleaks	An AI-probability score with sentence-level highlighting and a human/AI split view	Document and sentence	Why a particular sentence scored high; it reports the signal, not the cause	Editors checking long documents line by line
Originality.ai	An AI-probability score, often paired with plagiarism and readability checks	Document-level, with some section detail	Author identity, and it cannot separate AI-assisted editing from full generation	Content teams running bulk checks
Sapling	An AI-probability score with sentence-level highlighting in its writing tools	Document and sentence	Nothing about the drafting process behind the text	Writers checking their own drafts before submitting
DeAIze	A document score plus every sentence highlighted by AI probability, with a cross-check against configured third-party detector APIs	Document and sentence	Still a probability estimate, not proof of authorship	Seeing which lines carry the signal and re-checking after edits

Two things stand out from that table. First, almost every tool gives you a number, and only some give you the sentence-level view that tells you where the number comes from. Second, none of them report intent, drafting history or whether the ideas are yours. A score is a reading of surface patterns, not a statement about how the text was made.

Why do two detectors disagree on the same paragraph?

They are measuring different things. Each detector runs its own model over its own features, with its own threshold for calling a sentence AI-generated. One tool may weight perplexity and burstiness; another may lean on token predictability, sentence-length variance or punctuation rhythm. Change the threshold and the label flips.

We ran our own detector across two small corpora on 2026-09-13, using the local ensemble at a 30% flag threshold. Twenty-three hand-written DeAIze guides scored a mean of 21% AI, with a median of 22% and a range of 0 to 46%. Five of those 23 human-written pieces were flagged at that threshold. Twelve unedited model passages scored a mean of 33%, median 34%, range 4 to 46%.

Read those two ranges again. They overlap. Human writing reached 46% and unedited model output dropped as low as 4%. That overlap is the whole story of why detectors disagree: the signal is real but noisy, and a single reading near the middle of the range tells you very little on its own. The same measurement found AI tells running at 4 per 100 words in the human sample against 1 per 100 words in the unedited model output, which is a useful reminder that the surface patterns people associate with AI are not reliably present in either group.

If you want the mechanics behind those features, our breakdown of how AI detection works walks through the signals each tool is scoring.

If you are comparing tools and want to see the sentence-level view for yourself, DeAIze scores a document and highlights every sentence by AI probability, so you can judge the flagged lines rather than the headline number. Free scans accept up to 800 words, subject to rate/capacity limits. Longer scans require an account and credits, with a 3,000-word maximum per check. The separate rewrite trial gives you 200 words once after email confirmation with no credit card, and pricing is pay-as-you-go with credits that never expire.

What are the known limits of AI detectors?

Three limits matter more than any accuracy claim.

False positives. In our 2026-09-13 run, 5 of 23 human-written articles were flagged at a 30% threshold. Those were pieces written by people, about their own work, with no model involvement in the drafting. A flag is a reading, not a finding. If you are on the receiving end of one, our guide to a false positive AI detector result covers what to document and how to respond.

Bias against non-native English writers. Predictable vocabulary, repeated sentence shapes and conservative phrasing all push a score up, and those patterns show up more often in writing produced by people working in a second language. The tool is not detecting AI; it is detecting regularity, and regularity has many causes. There is a dedicated piece on why an AI detector flags non-native English writers if that is your situation.

Short texts. A two-sentence abstract or a 40-word product blurb gives the model almost nothing to work with. Scores swing wildly on short inputs, and a high reading on a paragraph is close to meaningless. Judge detectors on how they behave across a full document, not on a one-line test.

How should you read a score without overreacting?

Treat the number as one reading from one instrument. A few habits keep it in proportion.

1. Look at where the signal sits. A document score of 30% built from one flagged paragraph is a different problem from the same score spread evenly across every sentence. Sentence-level highlighting is what makes this visible. DeAIze reports both the document score and a per-sentence AI probability for exactly this reason.
2. Compare against your own baseline. Run a piece you wrote before AI tools existed through the same detector. If it also scores high, you have learned something about the tool, not about your writing.
3. Ignore short-text scores. Under a few hundred words, the reading is noise.
4. Ask what decision the score is actually driving. If nobody is going to act on it, stop looking at it.

What should you do when two detectors disagree?

Disagreement is normal, and it is information. Work through it in this order.

• Check the granularity. If one tool highlights sentences and the other only gives a document number, the sentence-level result is more actionable. Start there.
• Re-run on a longer sample. If the disagreement is on a short passage, extend the sample and see whether the gap narrows. It usually does.
• Edit the flagged lines, not the whole document. Rewriting untouched sentences to chase a score risks flattening your voice for no gain. Fix only what the detector points at, and keep the meaning intact. Our mixed human AI writing guide covers finding the seams.
• Re-score after editing and compare before and after. DeAIze runs this as a closed loop: score the text, rewrite only the flagged sentences, keep a rewrite only when semantic similarity says the meaning survived, then re-score. The site reports an average AI-score reduction of 60% or more after humanizing, with a best case of 71%, and on the example paragraph used on the site a 92% AI score becomes 4%. Those are the product’s own predicted scores, not a verdict from any third-party detector.
• Cross-check against a third party where it matters. Where configured, DeAIze runs the output against real third-party detector APIs — GPTZero, Turnitin, Copyleaks, Originality.ai and Sapling — rather than trusting an internal guess. If your institution uses Turnitin, that cross-check is the number you care about.

For a deeper dive into why two well-regarded tools split on the same essay, see our GPTZero vs Turnitin comparison.

How do you choose between detectors responsibly?

Stop looking for the most reliable detector and start matching the tool to the decision.

Your situation	What to prioritise	What to ignore
Submitting to an institution	Whatever your institution runs, plus a sentence-level check of your own draft	Marketing claims about accuracy
Editing a long document	Sentence-level highlighting and a before/after re-score	Any single document-level number
Running a content team	Consistent thresholds across the whole team and a written policy for what a flag triggers	One-off spot checks on short texts
Writing in a second language	A tool that shows which sentences carry the signal, so you can judge whether the flag is about regularity rather than AI	Tools that only return a headline score
Comparing tools before buying	Transparent methodology and the ability to re-check after edits	Percentage-accuracy claims from any vendor

Two practical rules. First, decide your threshold before you scan anything, so you are not negotiating with yourself after the result arrives. Second, keep a record of what you ran and when. If a flag ever becomes a formal question, a dated log of your own drafts and edits is worth more than any score.

And keep the scope honest. Detection tells you something about surface patterns in a piece of text. It cannot tell you whose ideas are on the page. The ideas, data and conclusions have to be yours, whoever typed the sentences.

Our own methodology page sets out how we run and report these measurements, including the small-sample caveats that apply to the figures above.

---

Measurement note: figures in this article come from our own detector run on 2026-09-13 — n = 23 hand-written guides versus n = 12 unedited model outputs, median AI score 22% and 34% respectively. Method: /methodology/.