An honest comparison of what each tool does, including the things they do better than we do.
Detectors differ less in accuracy than in the job they do. GPTZero is aimed at schools, ZeroGPT is fast and free, Copyleaks and Originality.ai combine a similarity check with an AI estimate, Turnitin lives inside a university's submission system. None of them lets you verify its accuracy from the outside, because you do not know the texts it learned from — so compare the language, the output and what the tool does with your texts, not the numbers on a pricing page.
The descriptions follow what the vendors state themselves and what the tools return. You will find no accuracy figures here — they cannot be verified.
A US detector aimed at education. It flags individual sentences and offers material and integrations for teachers.
Strength:a long track record on English texts and tooling built for classrooms.
Limit:Czech and Slovak are side languages for it, and it knows nothing about how you write.
A free, fast detector that flags the sentences it reads as AI. We run it as one of our seven detectors, so you see its result next to the others.
Strength:you can try it immediately and without an account.
Limit:one number without context, with nothing to compare your earlier writing against.
Combines a similarity check with an AI estimate in one report. Aimed at institutions and companies, with support for many languages.
Strength:plagiarism and AI in one place, plus integrations with school systems.
Limit:a paid service built for organisations, not for a single author.
A tool for content agencies and SEO teams — credits, team accounts and checks on commissioned copy.
Strength:a workflow for text delivered by outside writers.
Limit:built mainly for English, and there is no free trial run.
Built into the university's submission system. The AI indicator appears in the same report as the similarity check.
Strength:a vast database of student work to compare against.
Limit:the institution buys it; a student cannot run it on their own text.
The detector built by the makers of ChatGPT. OpenAI shut it down in 2023 because of low accuracy.
Strength:worth remembering as a fact: not even the model's author could reliably recognise its output.
Limit:it no longer exists, so nobody can point to it.
And what you will not find here. Both take five minutes to discover by trying, so there is no point hiding either.
Each measures a different property of the text and we never merge them into one score. When they disagree, that is information about the text, not a fault.
The tool is built for them, not for English with Czech bolted on.
Style deviation measures how far a text departs from your own earlier texts of the same type. A general detector cannot ask that question, because it does not know your writing.
We look separately for invisible characters that travel with a copy-paste out of a chat and stay unseen in an editor.
After a detection you can humanise the text and have questions generated from it, so an author can show they know their own work.
No similarity check against a database of papers and no integration with school submission systems. Turnitin and Copyleaks cover plagiarism; we answer a different question.
Five things that tell you more about a tool than any accuracy claim.
Run five texts whose authorship you are sure of — three human, two generated. That test says more than a whole pricing page.
There is nothing you can do with 68 %. With a list of flagged sentences there is, because you can read them and judge for yourself.
Look for whether texts are stored, for how long, and whether they are used for training. With someone else's work or company material that matters more than the score.
A detector that only compares against a general norm answers a different question than one that knows your earlier texts.
If you cannot try a tool on your own texts first, you are buying a number you know nothing about.
The questions people ask most often when choosing between detectors.
It cannot be settled from the outside. Accuracy depends on the texts a tool learned from and the texts you give it, and you see neither as a user. The only test that carries weight is running the tool on your own texts, where you know the author.
Because it would be a number measured on a sample we chose ourselves. Such a number looks good and means nothing. We would rather write down what each detector measures and where it gets things wrong.
Yes as a view from several angles, no as proof. Averaging numbers from tools that measure different properties makes no sense — it is more useful to check whether they point at the same sentences.
No, they are two different questions. A similarity check looks for text that already exists somewhere; a detector estimates whether a model wrote it. Generated text usually matches nothing.
Paste a text whose author you know and see what seven detectors say about it side by side.
Run a detection