AI policies are repeating every mistake of plagiarism policies

For most of the noughties, plagiarism policies rested on three assumptions: that copying was a moral failure rather than a developmental one; that one standard of “acceptable help” could be applied fairly to students with very different resources; and that software could tell us who was guilty. It took a decade of research to establish that all three assumptions were wrong. We are now writing AI policies as if none of it happened.

Let’s think about what we got wrong the first time. We moralised a developmental problem. Patchwriting – copying and lightly modifying source text – is a stage nearly every novice writer passes through, research shows, while most textual plagiarism by second-language writers involves no intentional deception, other studies reveal. Nevertheless, we processed these students through misconduct panels rather than teaching them.

Research on proofreading shows that the limits of acceptable third-party assistance remain disputed and mostly unregulated, creating a grey area that particularly affects second-language writers. Yet we pretended our standards were neutral.

And we trusted the machines. Text-matching software returned a number and institutions treated it as a verdict. In 2020 I took part in a multi-country evaluation of those tools; they varied enormously in what they detected, and not one could responsibly stand in for human judgement. They were support tools. We used them as juries.

The bill for these mistakes was not paid evenly. International students, most of whom write in an additional language, are reported for academic misconduct at two to three times the rate of domestic students despite comparable attitudes to integrity and comparable actual rates of misconduct. The difference is not in the behaviour. It is in who attracts suspicion.

Now, we are making all the same mistakes again, faster.

And yet, we have a lens now that we did not have then. Hybrid human-AI writing is becoming normal, which many argue places us in a post-plagiarism era in which historical definitions of plagiarism no longer work. If that is right, then the question existing policies are built to answer, whose words are these? is the wrong one. Post-plagiarism thinking does not tell institutions to permit everything. It tells them that policing the surface of a text is a lost cause, and that integrity must be located somewhere else: in judgement, in process, in what a writer can account for.

We are moralising a developmental problem for the second time. AI use is being framed as a binary of permitted or cheating, when for second-language writers it is a question of learning. To be clear, I am not arguing that these students should generate essays with AI, simply that the decision weighs differently on them. While institutions may point to language entry requirements as proof of proficiency, being “academically functional” in a language is a baseline threshold, not the finish line. A student’s interlanguage development continues long after that initial exam is passed.

Their language proficiency is simultaneously the learning goal and part of the assessment criteria, so delegating language production to a machine can undermine both their development and their grade. That is a pedagogical dilemma, not a character flaw, and it deserves a teaching response rather than a disciplinary one.

We are pretending our standards are neutral for the second time. “AI may be used for brainstorming but not for generating text” sounds even-handed. It is not. It assumes a writer for whom language production is not the barrier, which is to say, a native speaker. For a second-language writer, language and content are not separable in the way that rule requires. It silently applies a stricter standard to these students.

Finally, we are trusting the machines for the second time, on worse evidence. Text-matching software at least measured something real: strings of shared words. AI detectors guess authorship from style. In 2023, several of my co-authors from that 2020 study tested the new generation and found them neither accurate nor reliable: not one reached 80 per cent accuracy, and they flagged human writing as machine-made and waved AI-generated text through. Nor do they err at random. GPT detectors misclassified more than half of authentic essays by non-native English writers as AI-generated, while judging native-speaker essays almost perfectly, researchers at Stanford found. The constrained vocabulary and predictable syntax of a non-native speaker were interpreted by the detector as machine-like.

So, we are once again treating a machine’s output as evidence against the students whose honest work it is least equipped to recognise. The result is a trap with no exit. Use AI and risk your learning, your marks and your integrity record. Avoid it and risk being accused of cheating anyway. Yet these students are not in the room where any AI or plagiarism policies are decided. This is the fourth mistake, and the one that produced the others.

We know the fixes, because we spent 10 years finding them. Name what is actually being assessed (language development, content knowledge, or both) and what role AI may play in each, then stress-test the rule against a multilingual student rather than an imagined native speaker. Never let detector output stand alone as evidence. Assess the process, not just the final product: drafts, portfolios and reflections show judgement where a finished text no longer can. And put multilingual students on the committees writing the rules.

What we do not have this time is a decade. Plagiarism policies were allowed to be wrong for years, and a generation of second-language students paid for the delay with misconduct records that followed them for life. AI policies are being written now, in this year's committees, for students who are already sitting in our classrooms. We know how this story ends. The question is whether we make them live through it twice.

Özgür Çelik is an EFL Instructor at Balikesir University, Turkey.