Technology
What an AI detector measures, why its result is a probability rather than proof, and what to arrange before challenging a pupil or employee about it.
No, not with certainty. An AI detector estimates how likely a text is to have been generated by a language model. It can miss AI text and flag human text. The result is a reason to investigate further. It is too weak to serve as evidence for a decision about a pupil, student or employee.
This is the position on 13 October 2026. The figures below come from sources we opened today. Before choosing a tool yourself, check its current documentation. This article is an explanation rather than legal advice.
A language model chooses, word by word, what is likely to come next. An AI detector reverses this: it examines how predictable a text is and compares it with what it knows about people and models. Predictable, uniform text receives a high score. Our article What is an LLM? explains what a language model does.
A detector therefore does not see origin. It does not know who sat at the keyboard or how much rewriting took place.
A plagiarism checker compares text with existing sources and shows similarities. You can check this by placing the source alongside it. An AI detector has no source to compare. Its score is an estimate you cannot recalculate yourself.
The two are linked: a more strictly configured detector misses less AI text and wrongly flags more people.
OpenAI released its own classifier on 31 January 2023 and stated in the announcement that it was not fully reliable. In its own test on English texts, it identified 26% of AI texts as likely AI-written and incorrectly classified 9% of human texts that way. OpenAI described the classifier as very unreliable below one thousand characters. The company said the tool should not be a primary decision-making tool. The top of the same page now states that the classifier has been unavailable since 20 July 2023 because of low accuracy.
In 2023, a group of European researchers tested fourteen detection tools, including two commercial systems. Their research, published in the International Journal for Educational Integrity, concludes that the tools are neither accurate nor reliable, tend to classify texts as human-written, and perform even worse after rewriting.
Stanford researchers presented seven widely used detectors with 91 essays by people writing English as a second language. According to their paper, published in Patterns, an average of 61.22% of those essays were incorrectly labelled AI-generated. Essays by US eighth-grade pupils passed almost without error. The researchers' explanation: second-language writers use more predictable vocabulary, and predictability is what a detector measures.
These studies are from 2023 and concern English text. We found no independent research of comparable scale on Dutch text.
Turnitin, a supplier used by many educational institutions, wrote in its own explanation from 2023 that less than 1% of fully human-written documents are incorrectly flagged, for documents in which the tool identifies 20% or more AI text. At sentence level, the company gives about 4%, with more errors in texts mixing human and AI writing. These are the maker's own figures, as at 2023. Ask the supplier for current figures.
A calculation using a fictitious school: four hundred pupils submit an assignment, and none uses AI. At a 1% error rate, four pupils are wrongly flagged. At 9%, that becomes thirty-six.
An assessment that someone wrote a text using AI, linked to their name, is personal data. Article 22 GDPR gives people the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them or otherwise significantly affects them. There are exceptions, and the data subject then has at least the right to human intervention, to express their view and to contest the decision.
A grade automatically invalidated above a score threshold, or an application letter rejected unread for that reason, comes close to this. If a human really examines the case, with authority to depart from the score, it is not a solely automated decision. Whether your process falls under Article 22 requires a legal assessment.
Article 5 GDPR requires personal data to be accurate. A note saying “assignment written using AI” in a pupil's or personnel file is a fact about a person. If based solely on a detector, you cannot demonstrate accuracy. The data subject may request rectification under Article 16.
In higher education, the examination board may take measures if a student commits fraud; this appears in Article 7.12b of the Dutch Higher Education and Research Act. The institution's own rules determine what counts as fraud and evidence. Our article on AI in education explains how a school otherwise handles pupil data and AI.
Employer and employee must act as a good employer and good employee, as stated in Article 7:611 of the Dutch Civil Code. A warning or dismissal based on a score you cannot explain fits poorly with that. Start with the arrangement: without an AI policy saying what employees may do with AI, there is little basis for challenging someone. Our article on writing an AI policy provides an outline; AI in HR covers recruitment and appraisal.
It can. Annex III of the EU AI Act lists as high risk, among others, systems evaluating learning outcomes, systems monitoring and detecting prohibited student behaviour during tests, and systems monitoring and evaluating workers' performance and behaviour. Whether a particular AI detector falls within this depends on its provider's intended purpose and your use. The requirements for those high-risk systems apply from 2 December 2027. See the AI Act in autumn 2026.
Anyone pasting an assignment, application letter or report into an online AI detector sends that text to another company. Such texts often contain a name. They may contain much more: an internship reflection, an essay about personal religious beliefs, a letter about a period of illness. Some may be special category data, for which stricter rules apply. Our article on special category data explains what that includes.
In its warning about AI chatbots and data breaches, the Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, wrote that entering personal data against organisational arrangements is a data breach, and that most providers store input. That warning concerns chatbots. The reasoning is the same for a free online detector.
Our page on AI chatbots and the GDPR explains what happens to text pasted into an AI service.
The AI Act opts for marking at the source rather than retrospective detection. Article 50 has applied since 2 August 2026.
Providers of AI systems generating synthetic audio, images, video or text must ensure output is marked in a machine-readable format and detectable as artificially generated or manipulated, as far as technically feasible. This obligation does not apply to systems merely helping with ordinary editing or not substantially changing input. According to the European Commission's timeline, systems already on the market before 2 August 2026 have until 2 December 2026 to meet this requirement.
On 10 June 2026, the Commission published a code of practice on marking and labelling AI content. Participation is voluntary; Article 50's obligations are not. According to the Commission, around 190 organisations had signed by late July 2026. On 13 October 2026, we could not verify whether the Commission's Article 50 guidelines, published as a draft in May 2026, had been finalised.
A marking helps with text coming directly from a marked system and not subsequently edited. It says little about text someone retyped, rewrote or made with another model. An absent marking does not prove a human wrote the text. And the obligation rests on the provider, rather than the pupil or employee.
Realistic images and audio carry a separate disclosure obligation. We discuss that in our article on deepfakes and the law.
| What you want to know | What an AI detector can do | What works better |
|---|---|---|
| Did this pupil write the assignment themselves? | Give a probability, without a source or explanation | Discuss the content, show drafts and notes, have some work done at school |
| Did an employee use AI for this report? | Give a probability; often no signal for rewritten text | An AI policy agreement to disclose AI use, and a colleague checking content |
| Was this application letter written using AI? | Give a probability, with more errors for non-native writers | Determine whether it matters, and assess skills in an interview or task |
| Does this text come directly from an AI system? | Estimate based on style | Read the marking the provider must apply under the AI Act, where present |
PrivaSearch is not an AI detector. It does not assess texts or people. It relates to another part of this subject: what happens to personal data when a document reaches an AI service.
PrivaSearch sits between your documents and the AI model. Personal data is detected and replaced with pseudonyms, and a colleague approves the result before a document becomes searchable. A connected AI assistant receives pseudonymised passages with a page reference, rather than original files. The steps appear on How PrivaSearch works. You can connect Claude, ChatGPT Business and Mistral Vibe (formerly Le Chat).
Two boundaries apply. Detection is never complete, so human review is mandatory. And pseudonymised data remains personal data for your organisation. Anything you type yourself into a connected assistant's window goes directly to its provider.
The page on the GDPR and the AI Act explains how the measures relate to the rules. The page on the risks explains what happens when employees work with AI tools without arrangements. To see how it works, book a demo.
Not reliable enough to base a decision on. OpenAI discontinued its own detector in July 2023 because of low accuracy. Independent research into fourteen tools in 2023 described them as neither accurate nor reliable. The result is a probability, and the error margin varies by tool, language and type of text.
Yes. This is called a false positive. It happens more often with short texts, formal or uniform texts and non-native writers. In a Stanford study, on average more than 61% of essays by writers using English as a second language were wrongly flagged. A high score therefore says something about a text's style rather than necessarily its origin.
A school would be wise not to do so based on the score alone. The result is a probability rather than proof, and a decision with consequences for a pupil requires assessment by a human examining the case themselves. A conversation, drafts and an oral explanation provide more support than a percentage.
Not without arrangements. The text contains personal data and goes to another company. You need a lawful basis, a data processing agreement with the provider, clarity about retention and reuse, and must inform the people concerned. Someone independently using a free tool falls outside those arrangements. That may be a data breach.
Since 2 August 2026, generative AI providers must mark their output in a machine-readable way as artificially generated, as far as technically feasible. Existing systems have a transitional period until 2 December 2026. The obligation rests on the provider. An absent marking does not prove a human wrote the text, for example because text was rewritten or retyped.
We found no independent research demonstrating this for Dutch text. The familiar studies concern English. Ask any supplier claiming Dutch support about the test behind that claim: which texts, how many, and how often human text was wrongly flagged.
How retrieval-augmented generation works, how it differs from uploading and fine-tuning, why it reduces errors and which privacy questions to ask.
What the Model Context Protocol is, how a connection differs from uploading a file, what an assistant receives and which risks you should check.
What a language model does, what tokens and a context window are, why a model invents things and what it does and does not remember from your input.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →