Skip to main content

AI detectors: what they can tell you and what they can't

An open guide for teachers and school leadership teams. The short answer: today no detector can prove that a student used AI, and the Catalan Department of Education says so too. What works is assessing the process.

By Julien Bellet Ribes

Content reviewed on

CC BY 4.0 licence: you can copy and adapt it, crediting IAescola.

In three sentences

  1. 1Text detectors are wrong too often to accuse anyone: OpenAI withdrew its own in 2023 because it identified only 26% of AI-written texts and flagged 9% of human-written ones as AI.
  2. 2Watermarks such as SynthID only recognise content from one specific company, and a negative result does not prove the text is human.
  3. 3The Department of Education asks schools to assess the process, to agree with students which uses of AI are allowed, and to encourage them to declare their use.

Do the maths for your class

A detector with a low error rate, applied to a whole class, ends up accusing someone who did nothing wrong. Adjust the values.

By default, the 9% that OpenAI published for its detector (2023).

2.7 students wrongly accused, on average

Probability of accusing at least one: 94%

Illustrative: the real error rate depends on the detector, the language and the type of text.

Do AI text detectors work?

Not well enough to base any decision about a student on them. OpenAI withdrew its classifier in July 2023 “due to its low rate of accuracy”: it identified 26% of AI-generated texts and flagged 9% of those written by people as AI.

An evaluation of 14 tools published in 2023 in the International Journal for Educational Integrity concluded that they are “neither accurate nor reliable”. A 2024 study of six detectors and 805 tests measured an accuracy of 39.5%, which dropped by 17.4 points when the text had been edited to evade them.

Turnitin, the most widely used in universities, warns that its result “should not be used as the sole basis” for taking action against a student, and it does not show scores between 1% and 19% because that range has more false positives. Vanderbilt University switched it off in 2023: with a 1% error rate, about 750 of the 75,000 papers submitted each year would have been wrongly flagged.

Sources: [1], [2], [3], [4], [6]

Do they disadvantage some students more?

They can. In a 2023 study, seven detectors flagged more than half of 91 essays by non-native English speakers as AI, while getting almost all of 88 essays by US eighth-graders right. The sample was small, and other studies do not find this bias with detectors built for a specific exam: it depends on the detector, the language and the texts. That is one more reason they cannot be used as evidence.

Sources: [7]

What about Catalan?

We have not found any independent, published evaluation of how reliable these detectors are in Catalan. Websites advertising an “AI detector in Catalan” give their own accuracy percentages without explaining how they were measured. Turnitin documents English, Spanish and Japanese; Catalan is not listed.

Sources: [5]

What is a watermark such as SynthID?

It is an invisible signal that some providers add to what their AI generates. Google applies it to Gemini's images, video, audio and text, and other companies, such as OpenAI, have started to adopt it.

It has limits that Google itself explains: the Gemini app only recognises content created with Google tools; if it does not find the mark, that means Google's AI did not make it, “but it could have been made by another system”; and in text the mark weakens when the text is heavily rewritten or translated. So it is not suitable for checking a student's work.

Sources: [8], [9]

And content credentials (C2PA)?

They are metadata that explain how an image was created, and OpenAI and Google use them. But they are easily lost: most social networks strip them when an image is uploaded, and so does a screenshot. OpenAI sums it up this way: an image without this metadata “may or may not have been generated” with its tools.

Sources: [10], [11]

What does the Department of Education say?

The 2024 guidelines are clear: “Actualment no hi ha cap detector completament fiable de productes creats per aplicacions d'IA” (p. 14). They also ask teachers to rethink learning activities and assessment (p. 21).

The 2025 framework for teachers' digital competence in AI asks them to explore these tools and assess “la fiabilitat, així com el respecte a la privacitat” (indicator 3.A.3.3), to focus assessment on the process (3.B.3.1), and to agree the permitted uses with students and encourage “la declaració d'ús” (3.B.3.2).

Sources: [12], [13]

And the EU AI Act?

Since 2 August 2026, Article 50 requires providers of generative AI to mark what they generate so that it is machine-readable and detectable as artificially generated. Systems already on the market before that date have until 2 December 2026.

The obligation falls on providers (and, for deepfakes and texts of public interest, on whoever publishes them), not on students or schools. There is also a practical point: uploading a student's work to an external detector means processing a minor's personal data.

Sources: [14], [15]

What to do instead of detecting?

  1. 01Agree at the start which uses of AI are allowed in each task, and ask for a short declaration: which tool, what for and what was used from it.
  2. 02Assess the process: drafts, the document's version history (Google Docs and Microsoft 365 keep one) and the decisions the student can explain.
  3. 03Include a part done in class or a short oral defence of the work.
  4. 04Redesign tasks so they require the student's own context: data from the school, the student's experience, a specific learning situation.

Sources: [13], [12]

What if I suspect a piece of work?

  1. 01Do not accuse anyone on the basis of a percentage: not even the most widely used vendor recommends it.
  2. 02Talk to the student: ask them to explain how they did it, define the key concepts or continue a paragraph in front of you.
  3. 03Look at the version history, if the document has one.
  4. 04If necessary, apply the school's rules based on what you have seen and heard, not on what a tool says.

Sources: [4], [6]

Sources

Primary sources, consulted on 5 October 2026.

  1. 1OpenAI, “New AI classifier for indicating AI-written text” (withdrawal notice, 20-07-2023)
  2. 2Weber-Wulff et al. (2023), “Testing of detection tools for AI-generated text”, International Journal for Educational Integrity 19:26
  3. 3Perkins et al. (2024), International Journal of Educational Technology in Higher Education 21:53
  4. 4Turnitin, guide “AI writing detection model”
  5. 5Turnitin, AI writing detection in Spanish
  6. 6Vanderbilt University, “Guidance on AI detection and why we're disabling Turnitin's AI detector” (16-08-2023)
  7. 7Liang et al. (2023), “GPT detectors are biased against non-native English writers”, Patterns 4(7)
  8. 8Google, SynthID: developer documentation
  9. 9Gemini Help: check whether content was generated with Google AI
  10. 10C2PA, frequently asked questions
  11. 11OpenAI, C2PA in ChatGPT images
  12. 12Departament d'Educació (2024), “La intel·ligència artificial en l'educació. Orientacions i recomanacions per al seu ús als centres”
  13. 13Departament d'Educació i FP (2025), “Competència digital docent en intel·ligència artificial”
  14. 14Regulation (EU) 2024/1689, Article 50
  15. 15European Commission, FAQ on the transparency obligations of Article 50

Training for teaching staff

IAescola runs training for teaching staff on how to assess when students have AI within reach, based on your school's real tasks.

See the training for teachers