AI detectors: how reliably can AI-written text be spotted?

This article is an overview based on vendor information, documentation and publicly available reporting, not a scored test of our own. Prices and features change quickly; the vendor's own terms always take precedence. We only score what we have used ourselves for weeks: those tools are listed under AI tools compared.

How an AI detector works

An AI detector does not read a text the way a person does. It measures statistical features: how predictable the choice of words is and how uniform sentence length and structure are. Language models prefer likely words and write evenly; people jump around more. Many tools combine this measurement with a classifier trained on examples of AI-written and human-written text.

The result is a probability, not a finding. "87 per cent AI" does not mean that 87 per cent of the text was written by an AI, but that the text resembles what the tool has learned to regard as typical of AI. Reading the figure as proof misunderstands the instrument.

The well-known tools

GPTZero is the best-known tool and checks German, French, Spanish and Portuguese as well as English; a free monthly allowance is enough for occasional tests. QuillBot's AI detector checks up to 1,200 words per scan for free and, according to the vendor, handles more than 20 languages, German among them. Copyleaks lists 30 languages, while Originality.ai and ZeroGPT are aimed mainly at English-language text.

Scribbr's free version is limited to English and 500 words; German texts are only checked as part of the paid plagiarism check, from 17.95 euros. The German vendor PlagAware itself states that its score is not proof of AI use. Universities mostly use Turnitin, which has supported German since autumn 2025 and is licensed only to institutions.

What stands out is how cautiously the vendors themselves phrase things. Scribbr speaks of "no 100 per cent accuracy", and Turnitin now shows scores between 1 and 19 per cent only with an asterisk because they are too uncertain.

What research says

The most thorough study comes from Berlin. In 2023 a team led by Debora Weber-Wulff of HTW Berlin tested fourteen detectors, Turnitin among them, and concluded that they were "neither accurate nor reliable". The tools tended to classify texts as human, and even light editing caused their hit rate to collapse.

OpenAI withdrew its own detector in July 2023 after six months: it recognised only 26 per cent of AI texts and wrongly labelled 9 per cent of human texts as AI. The RAID benchmark, published in 2024 with more than six million texts, showed that detectors are easily fooled by other models, other settings or targeted changes.

Turnitin states that its own software produces false alarms in fewer than one per cent of documents with more than 20 per cent AI content, and around four per cent at sentence level. At a university with ten thousand submissions a semester, even one per cent means a three-digit number of students wrongly suspected.

False alarms: who is hit hardest

A 2023 Stanford study gave seven detectors essays from the TOEFL English test, written by people whose first language is not English. On average the tools classified 61 per cent of them as AI text, while they were almost always right about essays by native speakers. The reason: people writing in a foreign language more often choose the obvious words, just like a language model.

Formal and highly standardised texts are at similar risk: job applications, minutes, legal texts. ZeroGPT reportedly even judged the American Declaration of Independence to be 98 per cent machine-written. And anyone who polishes their own text with DeepL Write or a grammar checker makes it statistically more uniform; how far such tools intervene is shown in our article Rewriting text with AI.

Humanizers: the arms race

The detectors were followed by "humanizers": services that rephrase an AI text so that it passes the check. They vary sentence length, swap likely words for rarer ones and scatter small irregularities. A study from June 2025 showed that targeted paraphrasing reduced the detection rate across all the detectors tested by almost 88 per cent on average, watermarks included.

That makes detectors worthless when it matters, but it does not make humanizers harmless. The texts often get worse, sometimes to the point of distorting the meaning, and the legal position does not change: anyone who conceals prohibited AI use is cheating, whether a program notices or not.

Watermarks instead of guesswork

More reliable than guessing from the finished text is a watermark that the provider builds in while writing: an invisible shift in word choice that can be proven with the right key. Google released its SynthID Text method as open source in October 2024 and uses it to mark text from Gemini; it presented the SynthID Detector checking portal in May 2025.

OpenAI did not introduce a finished method in 2024, arguing among other things that it could be bypassed by translation. In August 2026 Anthropic announced that it would watermark text from Claude; for now the detection interface is open only to selected organisations. Anthropic itself says that a complete rewrite removes the mark.

Brussels is adding momentum: since 2 August 2026, Article 50 of the AI Act has required providers to label AI content in a machine-readable way; systems already on the market before then have a grace period until 2 December 2026. That still does not add up to proof in every case: open models on your own computer mark nothing, and every mark fades with editing.

What German courts require as evidence

In 2023 the Munich Administrative Court ruled on an applicant for a master's programme at TU Munich whose essay had been rated 45 per cent AI by software. The court sided with the university but relied mainly on the assessors: the comparison with a sample text from ChatGPT, the striking difference from an essay submitted the year before and the examiners' judgement.

The administrative courts in Kassel (February 2026) and Cologne (June 2026) have also upheld findings of cheating without a detector playing any part: in Kassel, contradictions between the written work and the oral examination were decisive, in Cologne invented references. Since 2025 the Digitale Lehre Hub Niedersachsen has explicitly advised universities against using detectors in examinations, not least because uploading students' texts to a service in the United States requires a legal basis under the GDPR.

What you can do with a detector

For teachers and editors, a detector is at most a reason to look more closely, never a verdict. Conversations about the text, a short oral follow-up, drafts and checking the sources say more: invented references give an AI away more reliably than any percentage.

If you write yourself and fear suspicion, keep drafts, notes and the version history in Word or Google Docs. And if you use AI where it is allowed, say so: an open note about the tool saves any debate about percentages. The rules companies should set for this are in our guide to the GDPR and the AI Act.

In short: AI detectors deliver probabilities, not proof. They miss rephrased AI text and suspect people, especially non-native speakers. Watermarks from Google and Anthropic only help with unedited text. German courts base cheating findings on conversations, comparison texts and invented sources, not on percentages.