Students increasingly use LLMs to write essays, term papers, theses — any assignment that isn’t done under an instructor’s continuous supervision. Naturally, the higher education system is unhappy about this and is trying to fight back. And the simplest solution is to detect AI-generated texts and reject them.
This article is about the non-obvious ways AI detectors affect education.
A note on context: I’m writing from inside the Russian-language university system. Some details below are specific to it — for example, every bachelor’s degree ends with a written thesis and a formal defense before a committee, and AI detection is often built into the plagiarism-checking software that universities routinely run every thesis through. But the dynamics I describe are, as far as I can tell, universal.
Students
Today I see a well-established trajectory of LLM habit formation, starting in school, running all the way to the thesis, and continuing into the workplace.
The student expects a well-defined task, which they feed to an LLM, getting a finished, formatted solution as output. Schools, in my view, have let this process run its course entirely; moreover, some teachers admire the use of “new technologies”, seeing it as a sign that their students are advanced.
First-year students keep running the same familiar algorithm in their very first classes. Homework assignments — absolutely all of them — are automatically done by an LLM. Seminars and practical sessions are spent in a continuous chat with AI. Essays and term papers are generated with a one-line prompt; the more sophisticated students also upload the department’s formatting guidelines into the chat.
In practice, the average instructor faces the temptation to let it slide and ask a few perfunctory questions about the paper, which the student can just about manage to answer. If the instructor’s enthusiasm hasn’t burned out yet, a long, nerve-racking war begins:
— You generated this. Redo it.
— I did it all myself. Do you want to see my lecture notes?
(I’ve never understood what lecture notes have to do with the paper.)
At this point the instructor faces a choice. Option one: sit down with the student, ask questions, have them explain a number of points, demonstrate that they don’t understand, give them an individual assignment, go through it together, give another one, and so on — while at least a hundred other students are waiting in line, not to mention family, home and a personal life. Option two: run the text through the first AI detector that comes to hand, point at the screen and say: “This is generated. Redo it.”
AI detectors and reward hacking
Deep down, of course, the instructor hopes the student will repent, delete the generated text, study the topic, form their own opinion and present it in their paper. In reality, in most cases it doesn’t work that way. The student feeds their text into the same AI detector and starts experimenting. Hooray! Shorten this sentence and it turns from red to green. Remove participial phrases, turn a bulleted list into running prose, replace a term with a simpler synonym — and there’s more and more green. Work is in full swing. The text becomes choppy, clumsy, in places outright dumb, but the goal is achieved: the detector shows 100% human.
The instructor checks it and gives a grade that depends on their conscience and the university’s policy.
Did the student understand anything? Did they get to grips with the subject? I very much doubt it. Work like this trains only one thing: the speed of editing text and the ability to produce sentences that mean roughly the same as the samples in front of you. It leads nowhere near an understanding of the subject, and any reasonably literate high-schooler could do it.
Moreover, we are watching a paradox unfold in real time. Universities have spent years teaching students to write coherently, argue their case and use precise terminology. Now, by way of AI detectors, they are indirectly pushing students to imitate clumsy writing (though sometimes it isn’t imitation). Simply because smooth academic prose has low perplexity and is regularly flagged by detectors as machine-generated.
ML engineers will find this situation very familiar: in neural network training it’s called reward hacking. An economist would call it Goodhart’s law — when a measure becomes a target, it ceases to be a good measure. A neural network is no less lazy than a student and, given the chance, will find the easy way to collect the reward. For a student it’s often easier to reshuffle text, training their typing speed, than to make a cognitive effort (something they haven’t done in a long time) and do the work themselves.
The funny part is that in the end you may get a paper in which every single letter was typed by the student — and yet it was effectively written by an LLM, and the student still doesn’t understand what they’re trying to hand in. A Ship of Theseus, rebuilt plank by plank in real time.
This leads to a stronger conclusion: the quality of a student’s work doesn’t depend on the quality of the AI detector. Even a perfectly accurate detector wouldn’t change the outcome, because the rewritten text really is human text. A perfect detector of machine text is still not a detector of human understanding.
The student becomes a mechanical appendage of the LLM whose job is to get past the detector. There is already a whole industry of “humanizer” tools that do this automatically. The student is simply the manual version: the student becomes the humanizer.
Consequences
In theory, the thesis supervisor or the defense committee is obliged to check whether the student understands their own work. But the presence of official AI detectors makes them complacent (although at some universities the thesis defense was a formality before, and still is).
A green light from the AI detector lets everyone speak of the student’s own work: there’s no need to ask awkward questions, and one can simply proceed on the assumption that the student understands what they wrote.
In other words, when the question is whether a student understands their own project, AI detectors are useless and even harmful. They sell a false sense that the student did the work themselves.
If a student’s paper passes the AI detector, that tells you only one thing: it passes the AI detector.
What to do
The only reliable way I see to assess a student properly is an in-person final examination, with the task generated at the moment the exam begins. The original meaning of the word masterpiece is a good reminder here: in medieval guilds, an apprentice proved the right to be called a master by making something in front of other masters.
So it should be now. I’m convinced that theses and take-home projects should become a thing of the past, as should accepting any work done outside the exam room as grounds for assessing a student or pupil. Only an in-person examination of a degree candidate can show whether they are fit to work in their field.
Give a heating engineer input data — generated by that same LLM at the start of the exam — plus reference handbooks and Excel, and let them do the thermal calculation for a boiler or a district heating network in front of the committee. An economist gets a huge spreadsheet of raw company data and analyzes it without internet access. A programmer writes working code; a power engineer draws a power supply diagram.
For fields like philosophy, of course, nothing needs to change. The potential damage from a poorly trained philosopher is highly doubtful.
I know none of this is new. Blue-book exams are making a comeback in American classrooms, and the oral defense is standard in many European systems. But almost everywhere it is treated as a final checkpoint — and that is the problem. All of this is useless if a student first encounters such an approach at the end of their studies: by then there is nothing left to check. Unfortunately, it has to be introduced from school and required from the first year of university. And that takes more teachers, more time and more effort.
It’s much more convenient to run the text through an AI detector and hand out the diploma.
Автор: Kamil_GR
