Anyone handing in a term paper sometimes fears a checking tool that labels the text as machine-written. AI detectors promise a quick answer, but their results are only estimates. This guide explains how such tools work, why they do not provide proof and how you can document your own work if the question ever arises.
Short answer: AI detectors estimate from writing style whether a text seems machine-generated. They give a probability, not proof for or against the use of AI. OpenAI withdrew its own detection tool in July 2023 because of its low accuracy. Last updated: October 2026. What counts at your university is set by its examination regulations and procedures.
The key points
- AI detectors measure style features, for example how predictable a text is. Human-written texts can look machine-like too.
- Providers themselves speak of probabilities, not certainty. Results differ clearly between programs.
- A result is an indication, not proof. The university must justify any accusation in the procedure.
- The strongest protection is a traceable history of your work: version history, notes and source lists.
- Whether and how detectors are used is up to your university. If in doubt, ask before you hand in.
How do AI detectors work?
Most methods calculate how surprising each word is for a language model and look for smoothly predictable patterns. Language models often produce exactly such smooth wording. The problem is that humans write like this too, especially in clear technical prose, in formal language or when they follow a guideline closely. The detector therefore recognises a property of the text, not its origin.
Why are the results so unreliable?
OpenAI switched off its own AI Classifier on 20 July 2023, as The Register reported, quoting the company. The reason given was the tool’s low accuracy. A 2023 study by Stanford University also showed that detectors frequently flagged texts by people who write in English as a second language as AI-generated. For students with another first language, this is a serious risk.
In addition, simple rewording can change the results. A text can therefore look more or less suspicious just because it has been rephrased, without anything changing in the work itself.
Why does this affect students with another first language in particular?
Anyone writing in a foreign language often uses a clear, even vocabulary and prefers standard phrasing. Exactly these features can lead a detector to classify a text as machine-made. This does not mean the risk is smaller, but that a result must be read with particular care for this group. If you are in such a situation, collect documents early and inform your supervisor.
Which alternatives to detectors exist?
More reliable than a checking score are methods that make the writing process visible. These include a short conversation about the content of the work, in which you explain your arguments, and a comparison with earlier texts by the same person, such as seminar papers or presentations. Questions about sources and methods also quickly show who really wrote a piece of work. For you, this means that the best protection for your work is being able to explain it.
What does a result mean, and what does it not mean?
The table shows how to read typical results correctly.
| Detector result | What it means | What it does not mean |
|---|---|---|
| High AI score | The text appears statistically smooth and predictable | Proof that an AI tool was used |
| Low AI score | According to the model, the text reads more like human writing | Proof that no AI was used |
| Different values across programs | Each method has its own standards and thresholds | A reliable statement about the text |
| Version history shows how the text grew | The work developed over weeks | A formal proof, but the strongest practical evidence of your own work |
What applies in examination procedures?
A detector result on its own is generally not enough to establish deception. The examiners must justify the allegation, and you usually have the chance to respond. How this works in detail is set by your university’s examination regulations. What matters is that you stay calm, gather the facts and submit documents early that show how you worked. Ask which program was used, in which version and with which result. You need these details for a well-founded response, and without them a score can hardly be interpreted. Guidance on dealing with AI at universities is offered by the recommendations of the University of Kassel (in German).
How do you prove your own work?
The proof does not have to be perfect, but it should hang together. A file with twenty versions that develops over several weeks says more than a document created afterwards. Make sure your records match in dates and file names. Contradictions weigh more heavily than gaps in documentation.
- Use a word processor with a version history, so that every change is saved with a date.
- Keep a research log with search terms, databases, dates and places where you found each source.
- Keep outlines, rough drafts and handwritten notes.
- Share the file and the log on request and prepare to explain the content orally.
The rules for labelling AI use are described in our guide AI in Theses: Labelling and Declaration of Authorship (in German).
Frequently asked questions
Can an AI detector alone prove that I cheated?
Based on what universities and experts say so far, that is not the intended role. A detector result is an indication, which the examiners weigh together with other circumstances. A procedure cannot rest on a number that a tool outputs alone.
Does rewording a text until the detector no longer flags it help?
It is not advisable. Rewording may change the display, but it does not make the text more independent. It can also be seen as an attempt to deceive if you disguise other people’s thoughts. Improve the content, not only the style.
Should I test my own text with a detector first?
That is not necessary and says little. An inconspicuous value does not mean that your work is flawless, and a suspicious value is not an admission of guilt. Your version history and your sources are the better precaution.
When should I contact my supervisor?
When you are unsure which form of AI use is allowed, or when you notice that checking software has flagged your work. An early message with specific questions shows professionalism and clears things up before the matter becomes formal.
What to do now
- Check the examination regulations and the guidelines of your department for rules on AI and checking software.
- Set up a version history and a research log for your current work.
- Clarify with your supervisor whether and how you should disclose AI use.
- Keep your documents until the grade has been issued.
What happens after your studies?
The question of how to prove your own work remains after your studies, for example with publications, expert reports or application processes. If you document your working steps, you can answer follow-up questions with evidence. This applies regardless of which software you use.
Related guides
- AI Use in Studies: Rules, Disclosure and Tools
- AI in Term Papers: What Is Allowed and What You Must Disclose
- AI in Theses: Labelling and Declaration of Authorship
- AI Rules at Universities: Why Every Examination Regulation Differs
- Learning with AI: Opportunities and Risks at University (in German)
Sources: Liang et al., GPT detectors are biased against non-native English writers (arXiv, 2023) · The Register, report on the shutdown of the OpenAI classifier (26 July 2023) · University of Kassel, Artificial Intelligence in Research (accessed October 2026)
Last updated: October 2026. This guide is general information, not individual advice.
