Why don't AI detectors work?

What are AI detectors?

AI detectors can be defined as software applications designed to analyze statistical and linguistic features of text. Based on an analysis of statistical patterns in the text, these detectors generate a probabilistic estimate of whether the text was created by a human or by artificial intelligence. The problem is that these estimates cannot be independently verified in practice. As research in this field points out (see Bassett et al., 2026), under real-world conditions, there is no reliable way to confirm whether a flagged text was actually created by AI or not. Among the best-known are Turnitin AI, GPTZero, ZeroGPT, Copyleaks, DetectGPT, Quillbot, and Corrector App.

How AI detectors (don’t) work?

AI detectors use so-called linguistic markers and predictive modeling to estimate the likelihood of a text’s origin; however, this carries the risk of errors and poses a fundamental problem in that it is always merely an estimate. At the same time, these tools are highly sensitive to the specifics of the underlying models, training data, system prompts, and platforms used. This way AI detectors work raises doubts about their practical usefulness and adherence to proper academic integrity procedures.

Current detectors also assume that a text is either entirely human-generated or entirely machine-generated. This approach does not reflect reality—students today commonly use AI, for example, for brainstorming, organizing ideas, editing text, or improving readability. Bassett et al. (2026) argue that the attempt to divide text into “human” and “machine-generated” is conceptually flawed, because the boundary between the two is fluid and detection tools cannot distinguish between permitted and prohibited uses of AI.

Ethical issues

In addition to the inability to reliably validate the results of AI detectors and their use as evidence for assessing academic integrity, research also highlights the issue of linguistic bias, whereby AI detectors assign a higher probability of AI origin to authors who are not native English speakers (Deep et al., 2025; Pratama 2025). Another ethical dimension is the very act of submitting student work to third parties, which raises issues related to both intellectual property and student privacy, as these tools are often not transparent about how they handle data, how they store it, and who has access to it. At the same time, they do not allow students to request that their data be deleted.

What does this mean in practice?

In their current form, AI detectors do not meet the standard of reliability required for assessing the origin of student texts, as they provide probabilistic estimates. None of the tools achieves 100% accuracy; results vary across tools and systematically disadvantage certain groups. At the same time, their use raises ethical issues related to data protection and student privacy. This does not, of course, mean that the problem of academic integrity does not exist—it means that the solution lies not in detection, but in transforming approaches to assessment, teaching, and transparency.

Experts tend to believe that the way forward lies in redesigning assignments, formative assessment, oral defenses, portfolio-based approaches, and open discussion about the ethical use of AI in education.

Sources:

• Bassett, M. A., Bradshaw, W., Bornsztejn, H., Hogg, A., Murdoch, K., Pearce, B. & Webber, C. (2026). Heads we win, tails you lose: AI detectors in education. Journal of Higher Education Policy and Management. https://doi.org/10.1080/1360080X.2026.2622146 
• Deep, P. D., Edgington, W. D., Ghosh, N. & Rahaman, M. S. (2025). Evaluating the Effectiveness and Ethical Implications of AI Detection Tools in Higher Education. Information, 16(10), 905. https://doi.org/10.3390/info16100905
• Pratama, A. R. (2025). The accuracy-bias trade-offs in AI text detection tools and their impact on fairness in scholarly publication. PeerJ Computer Science, 11, e2953. https://doi.org/10.7717/peerj-cs.2953