AI-text detectors are not proof, and institutions propose assessing the process instead
Programs that promise to reveal whether a piece of work was written by artificial intelligence get it wrong, including for students who have not used it, and even their own makers advise against punishing anyone on their say-so. This report explains what the studies, the European Regulation, data protection authorities and education guidelines say, and sets out a no-cost plan for assessing fairly in any subject: state which uses are allowed, ask for evidence of the process and listen to students.
7 minute read
A percentage that looks like a verdict
An AI-text detector does not check who wrote a piece of work: from the way it is written, it estimates the probability that it resembles what a language model usually produces. The result arrives as a percentage or a label, and that numerical form lends it an appearance of objectivity it does not have. A correct, predictable text can resemble machine writing without ever having been near a machine, and a generated text that is edited afterwards can pass as human. That is why a detector offers, at most, an indication to be checked, never proof.
The problem has already left the classroom. The diary reported the debate over the detectors’ “witch hunt” in publishing, published on 2 November 2026: publishers running manuscripts through detectors, accused authors with no way of showing they had not used AI, and two novels from 1946 and 1969 that the same detector first flagged as showing traces of AI and then as human. In the ChatGPT-IA-edu group, the conversation moved on to student assessment. One position argued for checking authorship by knowing the person and cross-checking with other formats on the same topic; another, for setting out a clear framework beforehand on which uses of AI are allowed and what evidence an accused student can provide. They agreed on the second point: clarify what is allowed and what for.
What the studies measure
The most cited study was signed by Weixin Liang, James Zou and other researchers at Stanford University in the journal Patterns in July 2023: “GPT detectors are biased against non-native English writers”. They ran 91 TOEFL essays, written by people who speak English as a second language, through seven detectors. On average, the detectors classed more than 61% of those human essays as AI-generated; 18 of the 91 (19.8%) were flagged by all seven at once, and 89 (97.8%) by at least one. By contrast, 88 essays by US eighth-grade students were classified almost without error. The authors strongly advise against using these detectors in educational or assessment settings.
That same year, Debora Weber-Wulff and seven co-authors tested fourteen tools, including the commercial Turnitin and PlagiarismCheck, in the study “Testing of Detection Tools for AI-Generated Text”, published in the International Journal for Educational Integrity. Their conclusion is that the available tools “are neither accurate nor reliable”: they tend to classify generated text as human, and it takes only a little editing or disguising of the text for them to perform considerably worse.
For Spanish there are more recent data. Sheila Queralt, Beatriz Esparcia, Marco R. Lessi, Lucía Sánchez-Vecina and Laura Úbeda Cuspinera published, in September 2025 in the journal Intelética, “¿IA, Humano o híbrido?: Fiabilidad de herramientas de detección de IA en textos multiautoría” (in Spanish). Using 180 Spanish-language texts (human, generated with ChatGPT, Gemini and Grok, and mixed), they compared Originality.ai, GPTZero and Copyleaks. Originality.ai had the best overall performance and GPTZero the lowest false-positive rate, but none of the three reached acceptable reliability with hybrid texts, and the authors observed wrong classifications made with high confidence.
Essays written by people that the detectors took for AI text
OpenAI, the company behind ChatGPT, launched its own text classifier in January 2023 and withdrew it on 20 July that year. According to the note added to its announcement, it did so because of its low accuracy: in its tests, it correctly identified only 26% of AI-written texts as likely generated and wrongly labelled 9% of human texts in the same way.
Turnitin, the most widely used detector in universities, admits on its blog that the risk of error is not zero. In March 2023, its chief product officer, Annie Chechitelli, wrote that, because the false-positive rate is not zero, teachers “will need to apply your professional judgment”. In June 2023 she gave the company’s own figures: less than 1% false positives per document for those with at least 20% AI writing, and about 4% per sentence, and she asked for the result to be used “to initiate a conversation, not to draw a conclusion”. These are the company’s own data.
GPTZero is even more explicit on its limitations page: “These results should not be used to punish students”. It adds that its accuracy is higher with texts similar to those in its training set, mostly English prose written by adults, and that heavily edited text is harder to classify.
The European and Spanish framework
The European Artificial Intelligence Act classes as high-risk, in point 3 of its Annex III, systems intended to evaluate learning outcomes and those intended to monitor and detect prohibited behaviour by students during tests. High-risk does not mean banned, but subject to strict obligations, including human oversight. According to the European Commission, on its page updated on 19 October 2026, those obligations have been postponed to 2 December 2027, while the AI literacy obligation has applied since 2 February 2025.
Data protection adds two cautions. Without referring to detectors, the Spanish Data Protection Agency (AEPD) reminds readers on its blog (in Spanish) that Article 22 of the GDPR recognises the right not to be subject to a decision based solely on automated processing, that it is not enough to “invent” human involvement, and that whoever supervises may fall into automation bias, that is, accepting without question what the machine says. In addition, on a related matter, online classes and exams, the AEPD indicates that the tools are chosen by the school or the education authority (in Spanish), with the advice of its data protection officer, and that teachers should use those made available to them.
Among education authorities, the one that speaks most clearly is Catalonia. Its Orientacions per a l’ús de la intel·ligència artificial als centres educatius (in Catalan), published by the Department of Education and Vocational Training in September 2026 and reviewed in the paper’s Library, justify their caution, among other reasons, by the fact that detectors offer insufficient guarantees, especially outside English, and conclude that they are inadvisable: “els millors detectors continuen sent les persones” (the best detectors are still people). They also establish that AI should not grade or determine the level of achievement, as this is a high-risk use under the European Regulation.
At national level, the INTEF’s Guide to the use of artificial intelligence in education (in Spanish), in its version 2.0 published on 8 September 2026, deals with academic integrity and includes a model declaration of AI use and a questionnaire for students to reflect on their relationship with it. The diary has not found in other regional guides a pronouncement on detectors as explicit as the Catalan one.
From product to process
The alternative most often repeated by institutions is not better surveillance but a different way of assessing. The MIT committee’s report on AI in teaching, dated 13 August 2026 and explained in the paper’s Library, asks teachers to review their learning objectives first and only then design AI-aware assessments. It proposes oral exams, semester portfolios and out-of-class assignments accompanied by a conversation in class, and that every course have a clear AI policy, communicated in a standard format.
On surveillance, the report recommends not relying on detectors, recalls that they may mistake the writing of people whose first language is not English for AI, and notes that MIT’s own disciplinary committee does not consider a detector result sufficient on its own. It prefers other evidence: platforms that keep a version history, partial submissions with dates, and teacher comments at several stages, not only on the final product.
To decide which use is allowed in each task there is a widely used tool, the AI Assessment Scale, created by Mike Perkins, Jasper Roe, Leon Furze and Jason MacVaugh and revised in December 2024 in a paper signed, among others, by Perkins, Roe and Furze. It has five levels: no AI, AI planning, AI collaboration, full AI and AI exploration. Its authors present it as a tool for deciding what role AI plays in a task and for redesigning it, and it has been translated into more than thirty languages.
A plan for any subject
The guidelines do not present AI only as a threat to assessment. The Catalan Orientacions accept that it can help create assessment instruments, mark unambiguous answers or detect common errors, always with pseudonymised data, and ask for activities that require interpretation and a personal position. The decision on a grade or a sanction must be taken by a person who can change it, not by a program.
With that, a teacher can start without paying anything. First, decide and write down for each task which uses of AI are allowed and why, with a simple scale such as the five-level one, and ask students to declare how they used it; the INTEF model is a good starting point. Second, ask for evidence of the process: a preliminary outline, drafts, the document’s version history or submissions in stages with comments.
Third, add a short explanation, oral or written in class, in which students say in their own words what they did, why they made their choices and what they learned; it works equally well in primary, secondary or vocational training. Fourth, redesign some task so that it depends on what was done in class, on the students’ own data or on a reasoned opinion. And, if something does not add up, talk before accusing.
What it is best not to do is just as clear: grading with AI, sanctioning on the strength of a detector’s percentage, or uploading students’ work to services the school has not authorised. When in doubt about a tool, the education authority’s data protection officer can advise.
AI-text detectors are not proof, and institutions propose assessing the process instead
Programs that promise to reveal whether a piece of work was written by artificial intelligence get it wrong, including for students who have not used it, and even their own makers advise against punishing anyone on their say-so. This report explains what the studies, the European Regulation, data protection authorities and education guidelines say, and sets out a no-cost plan for assessing fairly in any subject: state which uses are allowed, ask for evidence of the process and listen to students.
A percentage that looks like a verdict
An AI-text detector does not check who wrote a piece of work: from the way it is written, it estimates the probability that it resembles what a language model usually produces. The result arrives as a percentage or a label, and that numerical form lends it an appearance of objectivity it does not have. A correct, predictable text can resemble machine writing without ever having been near a machine, and a generated text that is edited afterwards can pass as human. That is why a detector offers, at most, an indication to be checked, never proof.
The problem has already left the classroom. The diary reported the debate over the detectors’ “witch hunt” in publishing, published on 2 November 2026: publishers running manuscripts through detectors, accused authors with no way of showing they had not used AI, and two novels from 1946 and 1969 that the same detector first flagged as showing traces of AI and then as human. In the ChatGPT-IA-edu group, the conversation moved on to student assessment. One position argued for checking authorship by knowing the person and cross-checking with other formats on the same topic; another, for setting out a clear framework beforehand on which uses of AI are allowed and what evidence an accused student can provide. They agreed on the second point: clarify what is allowed and what for.
What the studies measure
The most cited study was signed by Weixin Liang, James Zou and other researchers at Stanford University in the journal Patterns in July 2023: “GPT detectors are biased against non-native English writers”. They ran 91 TOEFL essays, written by people who speak English as a second language, through seven detectors. On average, the detectors classed more than 61% of those human essays as AI-generated; 18 of the 91 (19.8%) were flagged by all seven at once, and 89 (97.8%) by at least one. By contrast, 88 essays by US eighth-grade students were classified almost without error. The authors strongly advise against using these detectors in educational or assessment settings.
That same year, Debora Weber-Wulff and seven co-authors tested fourteen tools, including the commercial Turnitin and PlagiarismCheck, in the study “Testing of Detection Tools for AI-Generated Text”, published in the International Journal for Educational Integrity. Their conclusion is that the available tools “are neither accurate nor reliable”: they tend to classify generated text as human, and it takes only a little editing or disguising of the text for them to perform considerably worse.
For Spanish there are more recent data. Sheila Queralt, Beatriz Esparcia, Marco R. Lessi, Lucía Sánchez-Vecina and Laura Úbeda Cuspinera published, in September 2025 in the journal Intelética, “¿IA, Humano o híbrido?: Fiabilidad de herramientas de detección de IA en textos multiautoría” (in Spanish). Using 180 Spanish-language texts (human, generated with ChatGPT, Gemini and Grok, and mixed), they compared Originality.ai, GPTZero and Copyleaks. Originality.ai had the best overall performance and GPTZero the lowest false-positive rate, but none of the three reached acceptable reliability with hybrid texts, and the authors observed wrong classifications made with high confidence.
Essays written by people that the detectors took for AI text
OpenAI, the company behind ChatGPT, launched its own text classifier in January 2023 and withdrew it on 20 July that year. According to the note added to its announcement, it did so because of its low accuracy: in its tests, it correctly identified only 26% of AI-written texts as likely generated and wrongly labelled 9% of human texts in the same way.
Turnitin, the most widely used detector in universities, admits on its blog that the risk of error is not zero. In March 2023, its chief product officer, Annie Chechitelli, wrote that, because the false-positive rate is not zero, teachers “will need to apply your professional judgment”. In June 2023 she gave the company’s own figures: less than 1% false positives per document for those with at least 20% AI writing, and about 4% per sentence, and she asked for the result to be used “to initiate a conversation, not to draw a conclusion”. These are the company’s own data.
GPTZero is even more explicit on its limitations page: “These results should not be used to punish students”. It adds that its accuracy is higher with texts similar to those in its training set, mostly English prose written by adults, and that heavily edited text is harder to classify.
The European and Spanish framework
The European Artificial Intelligence Act classes as high-risk, in point 3 of its Annex III, systems intended to evaluate learning outcomes and those intended to monitor and detect prohibited behaviour by students during tests. High-risk does not mean banned, but subject to strict obligations, including human oversight. According to the European Commission, on its page updated on 19 October 2026, those obligations have been postponed to 2 December 2027, while the AI literacy obligation has applied since 2 February 2025.
Data protection adds two cautions. Without referring to detectors, the Spanish Data Protection Agency (AEPD) reminds readers on its blog (in Spanish) that Article 22 of the GDPR recognises the right not to be subject to a decision based solely on automated processing, that it is not enough to “invent” human involvement, and that whoever supervises may fall into automation bias, that is, accepting without question what the machine says. In addition, on a related matter, online classes and exams, the AEPD indicates that the tools are chosen by the school or the education authority (in Spanish), with the advice of its data protection officer, and that teachers should use those made available to them.
Among education authorities, the one that speaks most clearly is Catalonia. Its Orientacions per a l’ús de la intel·ligència artificial als centres educatius (in Catalan), published by the Department of Education and Vocational Training in September 2026 and reviewed in the paper’s Library, justify their caution, among other reasons, by the fact that detectors offer insufficient guarantees, especially outside English, and conclude that they are inadvisable: “els millors detectors continuen sent les persones” (the best detectors are still people). They also establish that AI should not grade or determine the level of achievement, as this is a high-risk use under the European Regulation.
At national level, the INTEF’s Guide to the use of artificial intelligence in education (in Spanish), in its version 2.0 published on 8 September 2026, deals with academic integrity and includes a model declaration of AI use and a questionnaire for students to reflect on their relationship with it. The diary has not found in other regional guides a pronouncement on detectors as explicit as the Catalan one.
From product to process
The alternative most often repeated by institutions is not better surveillance but a different way of assessing. The MIT committee’s report on AI in teaching, dated 13 August 2026 and explained in the paper’s Library, asks teachers to review their learning objectives first and only then design AI-aware assessments. It proposes oral exams, semester portfolios and out-of-class assignments accompanied by a conversation in class, and that every course have a clear AI policy, communicated in a standard format.
On surveillance, the report recommends not relying on detectors, recalls that they may mistake the writing of people whose first language is not English for AI, and notes that MIT’s own disciplinary committee does not consider a detector result sufficient on its own. It prefers other evidence: platforms that keep a version history, partial submissions with dates, and teacher comments at several stages, not only on the final product.
To decide which use is allowed in each task there is a widely used tool, the AI Assessment Scale, created by Mike Perkins, Jasper Roe, Leon Furze and Jason MacVaugh and revised in December 2024 in a paper signed, among others, by Perkins, Roe and Furze. It has five levels: no AI, AI planning, AI collaboration, full AI and AI exploration. Its authors present it as a tool for deciding what role AI plays in a task and for redesigning it, and it has been translated into more than thirty languages.
A plan for any subject
The guidelines do not present AI only as a threat to assessment. The Catalan Orientacions accept that it can help create assessment instruments, mark unambiguous answers or detect common errors, always with pseudonymised data, and ask for activities that require interpretation and a personal position. The decision on a grade or a sanction must be taken by a person who can change it, not by a program.
With that, a teacher can start without paying anything. First, decide and write down for each task which uses of AI are allowed and why, with a simple scale such as the five-level one, and ask students to declare how they used it; the INTEF model is a good starting point. Second, ask for evidence of the process: a preliminary outline, drafts, the document’s version history or submissions in stages with comments.
Third, add a short explanation, oral or written in class, in which students say in their own words what they did, why they made their choices and what they learned; it works equally well in primary, secondary or vocational training. Fourth, redesign some task so that it depends on what was done in class, on the students’ own data or on a reasoned opinion. And, if something does not add up, talk before accusing.
What it is best not to do is just as clear: grading with AI, sanctioning on the strength of a detector’s percentage, or uploading students’ work to services the school has not authorised. When in doubt about a tool, the education authority’s data protection officer can advise.
Key points
An AI-text detector offers, at most, an indication: independent studies find it unreliable and especially unfair to people writing in a second language.
In Spanish, none of the three detectors tested by Queralt and her team in 2025 was reliable with texts mixing human and AI writing.
The makers themselves warn: OpenAI withdrew its classifier in 2023 over low accuracy, Turnitin asks for its result to be used to start a conversation and GPTZero, not to punish.
The European Regulation classes AI that assesses learning as high-risk, and Catalonia advises against detectors and against grading with AI.
The alternative is to assess the process: AI uses declared for each task, drafts and version history, submissions in stages and a short oral explanation.