Detectors estimate
They assess how predictable word choice and sentence structure are and output a probability.
Limit: Plain, clear writing looks suspicious. Lightly edited AI text, on the other hand, often slips through[1].
Guide & generator · University and work
Detecting AI-generated text: proof over detectors.
AI detectors deliver probabilities, not proof. If you can show how a text came about, you are in a far stronger position when suspicion arises, in coursework just as much as in a job application, a report or client work.
Free · no sign-up · runs only in your browser
What detection can do
Three ways to detect AI. None of them proves cheating.
Detectors guess, watermarks only work with models that apply them. The only reliable evidence is what shows how a text came about.
Watermarks mark
Gemini, Claude and, in the EU, ChatGPT hide a statistical pattern in their word choice as they write[5][7][8]. Since August 2026, the AI Act has required machine-readable marking[6].
Limit: Only the provider can check, so far only for selected organisations. A match shows the model helped write the text, including when translating. No match does not mean “written by a human”.
Evidence shows the path
Version history, notes, source work and, where relevant, AI chat histories document how a text developed over days.
Limit: The evidence has to be created while you work. It can hardly be produced convincingly after the fact.
Watermarks live
How SynthID works. Explained right in the chat.
Five questions about AI watermarks, answered step by step. Pick one or just watch.
Is it true that AI text now carries a watermark?
Avi
For the big models, yes. The watermark sits right in the word choice:
- ModelMethodsince
- GeminiSynthID2024
- ClaudeSynthID variantAug 2026, worldwide
- ChatGPTits own methodOct 2026, in the EU
With SynthID you cannot see any of it: no special characters, no metadata, just a slightly shifted word choice.
Since August 2026, the AI Act has required machine-readable marking.
How does SynthID work?
Avi
For every next word the model has several fitting candidates. A secret key and the preceding words give each candidate a point or none:
Simplified exampleSecret key
The survey shows a clear rise.
Round 1Round 2Chosen
clear1
strong0
sharp0
marked1
clear1
marked0
clear
Every round awards new points. A tie is decided by chance.
Across a whole text, more words with a point slip in than chance would predict. Reading the text, you notice nothing.
Gemini responses in the live test≈ 20 mUser ratingsjust as good
Google DeepMind published the method in Nature in 2024.
And how is it detected?
Avi
With the same key, the provider recounts the points of every word:
Example valuesSame key
Thesurveyshowsaclearriseindailyuse,especiallyamongyoungerrespondents.
Words13with a point10Average77 %without watermark≈ 50 %
77 %
0 %without watermark ≈ 50 %100 %
Without a watermark, roughly every second word lands on a point by chance. If the average is clearly higher, a watermark is likely. The longer the text, the more certain the result.
Only the provider holds the key. Public AI detectors cannot check the watermark.
Can my university check this?
Avi
So far only selected organisations, because only those the provider admits can check:
- ClaudeDetection API in private preview, including educational institutions
- ChatGPTDetector only for approved researchers and expert bodies
Public AI detectors do not know the key. They can still only estimate.
I only had Claude translate it.
Avi
Then the translation is marked, because Claude chose every word in it. Light proofreading, by contrast, often leaves too little signal, according to Anthropic.
Does a match mean I cheated?
Avi
No. A watermark only shows that a model helped write the text:
MatchA model was involvedno proof of cheating
No matchNothing provenno proof that a person wrote it
What is allowed is set by exam regulations, lecturers or employers. Your best move is to disclose AI use and document how the text came about.
Create AI use statementEvidence folder
Avi
The generator is further down this page: free, no sign-up, and it runs only in your browser.
Simulated conversation. Words, points and the 77 % value are simplified example values, not a measurement. Sources:[5][7][8][9]
What counts when suspicion arises
A detector score alone proves nothing.
Anyone who alleges cheating has to prove it. TU Darmstadt states explicitly that detectors “do not provide proof”[3]. Examination offices and courts therefore work with the overall picture.
Unauthorised AI use still has consequences: at the end of February 2026, the Administrative Court of Kassel upheld two failed final theses. It granted leave to appeal[3].
Typical indications[4]
- Invented or incorrectly cited sources
- Wrong facts and contradictions typical of AI
- A clear gap between the written work and the oral examination
- Basic questions about your own work remain unanswered
- Phrasing typical of AI, but only in combination with other indications
The same precaution helps against each of these: you know your sources and can explain your work.
Precaution
The evidence folder: five records that arise as you work anyway.
Create a folder at the start and keep what accumulates anyway. In a company, this is the project file: briefing, drafts, approvals and a review note.
01
Version history
Word, Google Docs, Overleaf or Git save intermediate versions. Work in one file instead of loose copies.
02
Notes & outline
Excerpts, mind maps and discarded outlines. Photos of handwritten notes are enough.
03
Source work
Reference managers such as Zotero or Citavi, annotated PDFs, access dates. You opened every source yourself.
04
AI chat histories
If you are allowed to use AI and do: export chats with dates and disclose the use.
05
Being able to explain
Research question, method and result in your own words. The best preparation for any conversation.
Generator
Disclose AI use before anyone asks.
Many universities require an AI statement or a list of aids, and many clients ask for one. The generator creates a clean draft that you adapt to your requirements.
01
Choose the purpose
Coursework, report, publication or job application.
02
Tick honestly
Only what you actually used. “No AI” is a valid answer too.
03
Copy or download
The text is created in your browser, nothing is transmitted.
A template, not legal advice. If your university or client provides its own form, use that instead.
Accused of using AI?
Stay factual, reply in writing.
A conversation about AI use is not yet a verdict. These four steps help you keep track.
Important
This page summarises studies and public regulations and is no substitute for legal advice. If you receive a formal decision, watch the deadlines.
01
Ask for the reasons
Ask in writing what the suspicion is based on. A detector score alone is not proof.
02
Look up the rules
Examination regulations, course requirements and the declaration of independent work you signed.
03
Present your evidence
Version history, notes, sources and, where relevant, AI chat histories. Offer to explain your work orally.
04
Get advice
Student unions often offer free legal advice. For formal decisions, a lawyer specialising in examination law can help, at work the works council.
For businesses
The same question arises in every office.
Applications, agency copy, reports and press releases are routinely written with AI. The same means make this fair and reliable: rules set in advance, documented review, no detectors as proof.
Assess applications fairly
Detectors demonstrably disadvantage people who are not writing in their native language[2]. A clear AI rule in the job advert and a conversation about concrete work samples are fairer.
Agree on AI use up front
Agree in advance what AI may be used for, and deliver a short AI use statement with every result. This replaces mistrust with traceability.
Record the human review
For AI text on matters of public interest, the disclosure obligation does not apply if a human reviews it and takes editorial responsibility[6]. Keep a record of this review.
AI labelling obligations
Quick answers
Frequently asked questions
Can lecturers reliably detect AI-generated text?
No. In a comparison of 14 detectors, none reached an accuracy of 80 percent. Detectors deliver probabilities, not proof. Universities therefore base an accusation of cheating on the overall picture, such as invented sources or clear differences between the written work and the oral examination.
Does text from ChatGPT, Claude or Gemini carry a watermark?
Increasingly, yes. Gemini marks text with SynthID, Claude has done so worldwide since August 2026 with a variant of it, and since October 2026 OpenAI has been rolling out its own method for ChatGPT in the EU. So far only selected organisations can check, and for Claude this includes educational institutions. A match shows that the model was involved in the word choice, for example when translating. According to Anthropic, pure proofreading often leaves too little signal. A match alone does not prove cheating.
Is AI allowed in coursework and theses?
Each university decides for itself, often even each course. Many allow certain uses if they are disclosed. What counts are the examination regulations, the lecturer's requirements and the declaration of independent work that you sign.
What belongs in an AI use statement?
Which tools you used, for what and where, and that you checked the results and take responsibility for them. If your university or client provides its own form or list of aids, use that instead.
Should I run my text through an AI detector beforehand?
That does not create certainty. An unremarkable result proves nothing, a conspicuous one may be a false alarm. You would also be uploading unpublished work to a third-party service. Your time is better spent on a clean evidence folder.
Should companies check job applications with AI detectors?
We advise against it. False alarms particularly affect people who are not writing in their native language. A clear rule on AI use in the job advert and a conversation about concrete work samples are more meaningful.
Do companies have to label AI-generated text?
Not across the board. For published text on matters of public interest, Article 50 of the EU AI Act requires disclosure. This does not apply if a human reviews the text and takes editorial responsibility. You should document this review.
Are my entries in the generator stored?
No. The generator works entirely in your browser. Your entries are neither stored nor transmitted to Avolang or third parties.
Sources
- Weber-Wulff, D. et al. (2023): Testing of detection tools for AI-generated text. International Journal for Educational Integrity 19, Art. 26: link.springer.com
- Liang, W. et al. (2023): GPT detectors are biased against non-native English writers. Patterns 4(7): arxiv.org
- hessenschau (4 March 2026), report on how universities in Hesse deal with AI in final theses (in German): hessenschau.de
- Düsseldorf University of Applied Sciences, Legal & Compliance department: guidance on cheating, as of May 2026 (in German): hs-duesseldorf.de
- Google AI for Developers: SynthID Text: ai.google.dev
- Regulation (EU) 2024/1689 (AI Act), Article 50: eur-lex.europa.eu
- Anthropic (14 August 2026, updated 1 September 2026): How Claude’s text watermark works: anthropic.com
- TechCrunch (5 October 2026): OpenAI will start watermarking ChatGPT’s text in the EU: techcrunch.com
- Dathathri, S. et al. (2024): Scalable watermarking for identifying large language model outputs. Nature 634, 818–823: nature.com
As of 19 October 2026. Each study tests specific tools at a specific point in time. Newer detectors may perform better or worse; none reaches the quality of proof. Quotations from German sources are our translation.
Contact
Where is your team losing time today?
Enter your email, that is all. Zino Lang replies personally within 24 hours. In a free 20-minute fit check we clarify whether AI really solves your bottleneck.