ThesisDraft vs. GPTZero: The Number Neither of Us Wants You to Skip

What GPTZero and ThesisDraft Actually Are

GPTZero is the best-known AI detector in the world, and it got there honestly: it was one of the first, it is genuinely well engineered, and it has been iterated on for years. It is a general-purpose detector sold to educators, publishers and individuals, with a free tier, paid plans metered by words, sentence-level highlighting and a writing-process feature that replays how a document was typed.

ThesisDraft’s AI check is narrower on purpose. It does one thing: it tells a student, before they submit, what an AI detector sees in their thesis — and it publishes how often that verdict is wrong about someone who wrote their own work. The first 1,500 words are free with no account; a full document is €4.99 once, with 14 days of unlimited re-checks.

We are not going to tell you our detector is more accurate than GPTZero’s. We have not run a head-to-head benchmark, and claiming a result we have not measured is the exact failure this site was built against. What we can compare is what each tool tells you about its own limits.

Side by Side

The False-Positive Rate

Every detector returns a percentage. Almost none of them return the second number: how often that percentage is wrong about a person who wrote their own work. Those are completely different quantities, and only the second one describes your risk.

Ours, measured on 9 August 2026 against corpora published before ChatGPT existed:

• English, over 400 words — high-probability verdict, roughly 3 false positives in 100 human documents.
• English, under 400 words — roughly 2 in 100.
• German, over 400 words — roughly 4 in 100.
• German, under 400 words — no high-probability verdict is issued at all, because no threshold strong enough to catch AI there costs fewer than about 7 false accusations in 100.

Those figures come with their own caveats, and we publish those too: the long-form weights were fitted on the corpus they were then measured against, so the separation is optimistic; the AI-side corpora are 20–25 documents from two model families; the German long-form corpus is Wikipedia prose, not thesis prose. All of it is on our accuracy page, with the scripts that produced it.

We publish this because of something that happened to us. An early version of this detector scored an AUC of 0.184 against 236 pre-2022 arXiv abstracts — a coin flip is 0.50. It was ranking machine-written text as more human than human-written text, and it had been returning confident percentages the entire time. The failure was invisible from the output and obvious from the corpus. Any detector that has never run that test does not know whether it has the same problem.

What GPTZero Does Better

Several things, and pretending otherwise would undermine the point of this page.

• Scale and maturity. GPTZero has processed orders of magnitude more text than we have, across far more languages and document types, and has been refined against that volume for years.
• Writing-process evidence. Their authorship/replay feature reconstructs how a document was written over time. If you are trying to prove you wrote something, that is stronger evidence than any statistical score, ours included. We do not have an equivalent.
• Institutional integrations. If you are a teacher rather than a student, GPTZero fits into workflows we do not serve at all.
• Breadth of text types. We are tuned for academic prose. GPTZero handles email, journalism, fiction and student work with one product.

If you need to demonstrate authorship of a document you already wrote, use their process-replay feature. If you want to know what a detector will see in your thesis before anyone runs one on it, that is what we built.

Pricing: Credits vs One Payment

GPTZero’s paid plans are monthly subscriptions with word allowances. That model makes sense for someone checking submissions every week of the semester.

A student writing one thesis is not that person. You need a full-document check a handful of times in the final fortnight, and then never again. So the AI check is €4.99, paid once, and it unlocks 14 days of unlimited re-checks — long enough to fix what you found and check again, short enough that we are not selling you a year of something you needed for two weeks. The free tier covers the first 1,500 words with no account, which is enough to see how the tool behaves on your own writing before paying for anything.

If You Are the Student Being Accused

The reason we keep returning to the error rate is that detectors are used against people. A published finding that detectors disproportionately flag non-native English writers is not an edge case in a German university — it describes a large share of every thesis submitted there.

So our verdicts are deliberately conservative. Where the evidence does not support a strong claim, we do not make one: short German text never receives a high-probability verdict from us, and never will on the current evidence. That is a measured limit, not a feature we have yet to build, and we would rather say it than sell a confident number we cannot stand behind.

The same check, the same verdict and the same error rate go to the student before submission and to the examiner afterwards. That symmetry is the product. A tool that only works when your supervisor is not reading it is not one we would ship.

The Verdict

Choose GPTZero if you need breadth across many text types, institutional integration, or documentary evidence of your writing process.

Choose our AI check if you are a student who wants to know, before submitting, what a detector sees in your thesis — and wants the error rate printed next to the verdict rather than left out of it. The first 1,500 words cost nothing and require no account, so the comparison is cheap to make yourself.

Run the free AI check · See our measured error rates