expulp Verdict

Verdict blog, Research, explained, week of 2 November 2026

Can AI-generated text be detected? What three studies found

Detection tools leaned toward calling AI text human, teachers missed it while feeling sure, and one 100% claim covers only half the test.

Can AI-generated text be detected? The short answer

Not reliably, on the evidence of these three studies: published tools and trained teachers both failed to separate ChatGPT text from human writing. A third study reports 100% accuracy for its own classifier, but that figure applies to human essays only, and the rest of this piece shows why that matters.

Everything here comes from the published abstracts of three papers, all set in education and all about ChatGPT. Abstracts leave out much of what a careful reader wants, such as per-tool scores, sample sizes and tables. Where a number is missing, this piece says so rather than filling the gap.

What the tool and teacher studies measured

Weber-Wulff et al., 2023 tested 12 publicly available tools and two commercial systems, Turnitin and PlagiarismCheck, asking whether they could reliably tell human-written from ChatGPT-generated text, and whether machine translation and content obfuscation changed that. They conclude the tools are neither accurate nor reliable, with a main bias towards classifying output as human-written, and that obfuscation significantly worsens performance. The abstract does not give accuracy for individual tools or the size of the obfuscation drop, so those figures cannot be reported here.

Fleckenstein et al., 2024 ran two experiments with 89 novice and 200 experienced teachers. Neither group could identify ChatGPT texts among student essays, though there were some indications that experienced teachers judged more accurately, and both groups were overconfident. The abstract does not say how many essays were misjudged or whether teachers tended to call AI texts human or human texts AI.

Those gaps matter because the two failures are not the same shape. The tools have a stated direction: they lean towards "human", so their typical error is letting AI text through. For teachers, the abstract gives no direction, only that judgments were wrong and held with too much confidence. A cautious tool that misses AI text and a confident reader who may err either way create different risks, and the second is the one more likely to end in an accusation.

How to read a 100% AI detection accuracy claim

Cingillioglu, 2023 built a support vector machine classifier on an n-gram bag-of-words discrepancy language model, designed above all not to classify genuinely human essays as AI-generated. The headline sentence in the abstract reads: "Via support vector machine (SVM) algorithm 100% accuracy was recorded for identifying human generated essays." Note the last four words. The figure is about the human class.

A worked example shows what that can hide. The numbers below are invented for illustration and do not come from the paper. Picture a test set of 100 human essays and 100 AI essays, and a lazy classifier that labels every essay human without reading it. It gets all 100 human essays right, so it scores 100% on the human class. It gets all 100 AI essays wrong, so it catches nothing. Its overall accuracy is 100 out of 200, or 50%, the same as a coin toss. Its human-class score is perfect and it is useless as a detector.

This is not a claim that Cingillioglu's classifier behaves like the lazy one. The point is narrower: the 100% figure cannot tell them apart. To do that you need the AI-class result, and the abstract does not report one. It says that AI essays "can be detected by teachers and educational designers using the proposed language model and machine learning (ML) classifier at a high accuracy", but gives no number for that. It also says human essays must be identified with 100% accuracy "even if the overall classification accuracy performance is slightly reduced", without stating the overall figure. The abstract gives no sample size and does not describe a test set. It mentions comparisons with OpenAI, GPTZero and Copyleaks software, but not their scores. A missing AI-class number is a different charge from a low one: the abstract does not show the classifier failing, it simply does not show it succeeding. The full paper may contain these figures; this reading is limited to the abstract.

You can reuse a short test on any detector claim. Does it report both errors, human text wrongly flagged and AI text missed? Does it give the sample size? Was it scored on text the classifier never saw during training? Does it say which model produced the AI text? Applied to this abstract, the answers are: the first error only, no, not stated, and ChatGPT. The abstract does stress recall and the F2 score as measures, which is a sound instinct, but the values are not given. Read this way, the 100% is a design commitment to protecting human writers, not yet evidence that AI text gets caught.

Who pays when AI detection gets it wrong

The two errors land on different people. A missed AI essay costs something diffuse: the fairness of a grade, the value of an assessment. A human essay wrongly flagged costs one named student, who may face a disputed mark or an integrity case for work they wrote. That imbalance is exactly why Cingillioglu puts zero false accusations first, and it is why Weber-Wulff and colleagues discuss the drawbacks of using these tools in academic settings.

Several questions a student or institution would ask cannot be answered from these abstracts. None describes what an appeals process looks like or how a student could contest a detector score. None examines whether some groups, such as non-native English writers, are flagged more often. These are open questions in this evidence, not settled ones, and any school relying on detection would need answers from elsewhere before acting on a score.

The teacher finding adds a specific danger. A wrong judgment held with confidence is more likely to be acted on than one held loosely. Fleckenstein and colleagues found overconfidence in both novice and experienced groups, so experience alone did not remove it.

Process-based assessment and its limits

If origin cannot be read off a finished text, the obvious alternative is to look at how it was made: drafts, notes, revision history. Fleckenstein and colleagues frame their results as evidence for the debate on exam strategies, which points towards changing how work is set and assessed rather than spotting AI afterwards.

That advice has real objections. Reviewing drafts takes teacher time that most do not have, and a draft history can itself be produced with AI, step by step. None of the three papers tests process-based assessment, so they cannot say whether it works better. What they do show is that the finished text is a weak signal, which makes a single final submission the least informative thing to grade.

There is a quieter lesson too. ChatGPT produced essays with relatively little prompting that experienced teachers could not tell apart from student work, so surface style says little about who wrote something. What still distinguishes work is whether it adds an argument, a source or a connection the reader did not already have. Asking what a piece contributes, rather than who typed it, is the kind of reading Expulp Verdict is built around, and it stays useful whichever way the authorship question turns out.

Questions people ask

Can AI-generated text be detected by tools like Turnitin?

Weber-Wulff et al. (2023) tested 12 public tools plus Turnitin and PlagiarismCheck and concluded they were neither accurate nor reliable. Their main bias was towards labelling text as human-written, and obfuscation significantly worsened performance.

Can teachers tell if an essay was written by ChatGPT?

In Fleckenstein et al. (2024), neither 89 novice nor 200 experienced teachers could identify ChatGPT texts among student essays. Experienced teachers showed some signs of more accurate judgments, but both groups were overconfident.

Does a detector with 100% accuracy prove AI text can be caught?

Not on its own. Cingillioglu (2023) reports 100% accuracy for identifying human essays, a score a tool labelling everything human would also reach. The abstract claims high accuracy for AI essays but gives no figure for it.

What should a credible AI detection claim report?

It should give both error rates, the sample size, results on unseen test text and the model that generated the AI text. The Cingillioglu abstract reports only the human-side result and names ChatGPT as the source.

Sources