What is AI-generated text detection – and can it be trusted?
AI-generated text detection uses software to estimate whether a text was written by a generative AI tool rather than by a person.
A new question for every marker
Since generative AI became widely available, markers have faced a question they never had to ask before: did the student write this at all? AI-generated text detectors promise an answer. They analyse a submission and return a score or a highlighted report suggesting how much of it may have been produced by AI.
The demand is understandable. Take-home assignments, essays and reports remain central to higher education, and institutions want to know whether the work they assess reflects the student’s own learning. But AI detection works very differently from traditional similarity checking – and the difference matters for how its results should be used.
How does it work?
The details vary between tools, but AI-generated text detection usually relies on one or more of the following:
- Statistical patterns. AI-generated text tends to be more predictable than human writing. Detectors measure how expected each word is, and how much this varies across the text.
- Trained classifiers. Models trained on large collections of human and AI-written text learn to tell them apart.
- Watermarks. Some AI providers can embed hidden patterns in generated text, but this only works if the provider does so and the text is not changed afterwards.
- A probability, not a match. Unlike a similarity report, a detector cannot point to a source. It can only estimate how likely it is that text was generated by AI.
Why does it matter how detection is used?
The main reason is fairness to students. A false accusation of misconduct can have serious consequences, and AI detectors do make mistakes. OpenAI withdrew its own AI text classifier in 2023 because of its low accuracy, and research has found that detectors can misclassify writing by non-native English speakers as AI-generated.
Even the best tools work with probabilities. A small false positive rate becomes many wrongly flagged students when applied across thousands of submissions. That is why universities that use AI detection typically stress that a score is never evidence on its own, and must be supported by other evidence and a conversation with the student.
What AI-generated text detection cannot do
AI detection cannot prove who wrote a text. Unlike plagiarism, there is no source to compare with – only a statistical judgement. Texts that are edited, paraphrased or mixed with the student’s own writing are much harder to classify, and paraphrasing tools are designed precisely to avoid detection. Nor can a detector tell permitted from unpermitted use – for example when a student has been allowed to use AI for research, brainstorming or shaping a problem statement, but not for writing the final text.
It also cannot answer the question that matters most: does the student understand and stand behind the work? As AI improves, the gap between human and machine writing narrows – and an arms race between generators and detectors is not one detection is likely to win.
How does WISEflow approach it?
Our view is that assessment is moving from detection to verification – from guessing whether AI was used to confirming that students can demonstrate their own learning. WISEflow supports that shift in several ways:
- Secure formats where it matters. The lockdown browser and layers of supervision – from device monitoring to live proctoring – remove access to AI during the exam.
- Process, not just product. In closed-book exams, every save and revision is logged, so the development of the text can be documented.
- Similarity that goes beyond words. Originality uses semantic similarity detection to find sources that have been paraphrased, translated or AI-washed – evidence that can be traced back to an original.
- Verification through dialogue. Written and oral elements can be combined, so students can be asked to explain and defend their own work.
- People decide. No tool replaces academic judgement and a fair process.
What should institutions consider?
- Clear rules on AI: state for each assessment what use of AI is allowed, and how it should be declared.
- Careful use of detectors: if used at all, treat a score as a prompt for a conversation – never as proof.
- Assessment design: use formats where students show their understanding, such as oral elements, drafts and in-class work.
- Fair process: make sure any suspicion is investigated with evidence and the student’s right to respond.
AI-generated text detection is not a reliable shortcut to integrity. The more robust answer lies in assessment design and verification – making sure that every grade reflects what the student can actually do.
