Compare
Two detectors, the same papers
We ran the same 960 documents through our detector and Pangram 4. On writing by people and on fully AI-written text, the result is a tie. We have not run Turnitin on these papers. What published studies found about it is further down this page.
Side by side
Same 960 documents, two detectors
Same detector on every plan. A flagged paper is one given a Mixed or AI verdict.
| What we counted | Our detector | Pangram 4 |
|---|---|---|
| Writing by people, wrongly flagged (600) | 0 of 600 | 0 of 600 |
| AI-written documents missed (220) | 0 of 220 | 0 of 220 |
| AI-edited papers flagged, each at its default line (ours 6%, Pangram 4 about 10%) | 102 of 140 (99 at a 10% line) | 88 of 140 |
| Learner essays wrongly flagged (180 of the 600) | 0 of 180 | 0 of 180 |
Our detector, run v1_8_s2_seed303, standard setting (6%), read on 2026-09-13 and again at 6% on 2026-09-15. Pangram 4 as it answered on 2026-09-08, scored once by us through its web dashboard.
- The first, second and fourth rows are a tie, not a lead. The 220 AI-written documents come from models our detector was trained on.
- Our default line (6%) sits lower than Pangram 4’s (about 10%). At the same 10% line the edited count is 99 to 88. Most of the gap is papers where a model added a paragraph. One test, 140 papers.
- Checked against a record of what the AI model changed, our detector gave the expected verdict on 124 of 140 edited papers, and Pangram 4 on 110. We wrote that labelling rule. Across our five training runs this row is a statistical tie.
- The detector that is live was chosen from five training runs by a rule written before the results. Its default line moved from 10% to 6% on 2026-09-15. Pangram 4 was scored once, on 2026-09-08.
- The 180 learner essays in the table are part of our fairness test, where our detector flagged 0 of 1,195. The 180 are a small set, and a tie. Pangram Labs reports 1 false flag in 24,586 learner writing samples for Pangram 4, a set about 20 times larger than ours that includes spoken-dialogue transcripts. Our ceiling is far looser, so this is not a win for us.
- A flag from either detector is a reason to look closer. It is not proof that anyone cheated.
What the verdicts mean
Human means the part of the paper that reads as machine-written is at or under your setting’s line. Mixed means some of it reads as machine-written or machine-edited. AI means 80% or more of it does. When this site says a paper was flagged, it got Mixed or AI.
What AI-edited means
Each of the 140 papers was written by a person, then partly changed by an AI model in one of four ways. Whether an edit breaks a rule is your school’s policy, not the detector’s.
- A light polish
- A rewrite for clarity
- An added paragraph
- A machine-written continuation
Limits
Where we fall short
Light polish
When an AI model lightly polishes a human paper, both detectors mostly miss it. Of 39 polished papers, ours flagged 3 and Pangram 4 flagged none.
Paid humanizer tools
They can hide machine writing from our detector. At one tool’s most expensive setting, none of 12 papers was called AI and 4 came back human-written. Pangram 4 was not tested on these.
Edited papers called Human
Even at the default setting, our detector called 38 of 140 partly AI-edited papers human-written.
Heavily rewritten text
On heavily rewritten machine text, our detector and Pangram 4 often disagree: same verdict on 31 of 51 documents. That counts disagreement, not who is right.
Short text
Our detector does not answer under 50 words, and readings under 150 words are weak. Machine-written text under 300 words has not been measured.
English academic writing
Our detector reads English academic and student writing. It is not built or tested for other languages or other kinds of writing.
Writing before 2022
Every human-written document we tested was written before 2022, so we know it is human. Writing from after AI tools became common has not been tested.
Our own tests
We run and grade our own tests. No outside party has checked these results yet.
Turnitin
What outside studies found about Turnitin
We have not tested Turnitin ourselves. Everything in this section comes from Turnitin’s own published statements and from independent studies, each named and dated.
Fully AI-written papers
A 2026 study of 160 academic papers found Turnitin scored all 40 fully AI-written papers under its own 20% threshold, missing every one. The 40 were master’s-level papers an AI model wrote for the study.
Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity, published 2026-06-23, Table 4. Read 2026-09-15.
Lightly reworded AI text
A 2024 study found Turnitin’s accuracy on AI-generated text fell from 50% to 7.9% once the text was lightly reworded. It used 15 AI-written texts produced in 2023 and 89 reworded versions of them.
Perkins et al., Simple techniques to bypass GenAI text detectors: implications for inclusive education. International Journal of Educational Technology in Higher Education, published 2024-09-17, Table 8. Read 2026-09-15.
What Turnitin says
Turnitin says its own document-level false positive rate is under 1% for documents it scores at 20% AI or higher, and shows no percentage at all below that line.
Turnitin, Understanding the false positive rate for sentences of our AI writing detection capability, 2023-06-14; and Turnitin’s guide AI writing detection model, read 2026-09-15.
- These studies used their own papers. They are not a test of our detector against Turnitin, and they do not rank the two.
- Turnitin’s AI writing detection is sold only to schools and institutions, under an institution licence. A teacher or student cannot buy it individually. Source: Turnitin help center, Can I purchase an individual Turnitin license?, read 2026-09-15.
Cost
What Human costs
Free
$0
2,000 words a day.
Schools and universities
$0
With a verified school email: 1,000 papers a month, up to 25,000 words a paper, in the web app, the API or the MCP server.
Larger volume
Ask
For a district or a university that needs more than the free plan gives each teacher. Write to us and say how many papers a term.
No plan checks plagiarism or images, and none fetches web addresses.
Get the school planTurnitin is a trademark of Turnitin, LLC. Pangram is a trademark of Pangram Labs, Inc. Human is not affiliated with either.
Check a paper yourself.
Paste one you know was written by a person, and one you know was not. Read what our detector marks in each.