A student submits a brilliant paper. But it’s too well-written, too structured, too “perfect.” The suspicion immediately arises: perhaps it was done by artificial intelligence. The first impulse is to use an AI-generated text detector. We apply it to the paper, and the tool tells us that there’s an 87% probability it was generated by a machine. Then we give free rein to the automation bias that makes us uncritically accept automated decisions. Case closed. Verdict reached.
Or is it? The case is not closed. This computer verdict is not only unreliable; it is, above all, unfair.
Artificial intelligence detectors seem like a logical solution, but they have two fundamental problems. The first is technical: they don’t work well . The second problem is more relevant: even if they worked perfectly, they wouldn’t solve the real problem.
A technically fragile solution
Unlike traditional plagiarism, where a text is compared to existing sources, here there is no original to compare it against. The aim is to distinguish between statistically human text and text statistically generated by a machine to appear human. It’s a dividing line that is difficult to draw and increasingly blurred.
Furthermore, there are reasons to believe this boundary will disappear. The better the generative models become, the more indistinguishable their output will be from human output. Detecting the use of artificial intelligence will be like trying to differentiate between two equally plausible texts, a task that, taken to its extreme, resembles flipping a coin . Pure chance.
The cost of making a mistake
We could accept that detectors make mistakes in some cases. But in education, those particular cases matter a great deal. Like all classifiers, AI-powered text detectors will make two types of errors: false positives and false negatives.
A false positive, that is, accusing a student who has done the work of fraud, has serious consequences: anxiety, helplessness and, in many cases, an accusation impossible to refute.
On the other hand, a false negative, failing to detect those who have used AI, has a more diffuse but equally damaging effect by rewarding those who did not fulfill their academic commitment: it erodes trust in the education system itself, and students perceive that the effort does not pay off, and motivation deteriorates.
Systems can be adjusted to minimize false negatives or false positives, but not both at the same time. (For example: either we adjust the system that detects breast cancer in X-rays so that it doesn’t miss any potential cases, at the cost of overdiagnosing, or we let cases slip through.)
Thus, using these systems will always involve accepting one of two types of injustice. If we minimize false negatives, we are opting for an evaluation based on control: we prioritize preventing any false negatives from “slipping through,” even if some of those detected are not actually texts written by AI.
Conversely, if we care more about avoiding false positives, we will be advocating for an evaluation that prioritizes learning and minimizes the penalty for error to a student who has made the effort to write their work.
A poorly framed problem
However, even if we solved the technical and ethical problems (for example, by opting to let some artificial text slip in so as not to penalize unfairly), we would still not address the essential issue.
Many academic tasks are meaningful because they involve cognitive effort : writing an essay, preparing a report, or solving an exercise requires time and work. And that effort is precisely what generates learning .
Artificial intelligence may not only be causing unfair grades, but it has also broken the link between those tasks and the cognitive effort they required. This completely changes the meaning of assessment. When AI tools are used, learning may not be taking place .
The illusion of detection
The detectors offer something very tempting: a sense of control. They allow us to think that the problem is contained, that it’s enough to identify those who commit fraud by breaking the rules. But that feeling is deceptive.
As the joke goes, we’re looking for the keys under the lamppost, not because we lost them there, but because that’s where the light is. In other words, we try to detect learning in the places we know to look, without worrying whether this necessarily implies that learning is actually happening.
The reliance on final products (a text, a report, a solution) as evidence of learning was already questionable: do they truly guarantee that a student understands a topic? Now, they are simply insufficient. Therefore, investing effort in improving assessment is, at best, irrelevant. And at worst, a distraction.
When the solution makes the problem worse.
The systematic use of detectors shifts the educational relationship toward suspicion. Instead of fostering student co-responsibility in their learning, it introduces a logic of surveillance in which the student becomes a potential offender, ignoring the presumption of innocence, and the teacher, a monitor.
This has more than just ethical implications. It also affects learning. Trust, autonomy, and responsibility are difficult to develop in an environment where the priority is avoiding accusations. Thus, paradoxically, in trying to protect academic integrity, we may be eroding the very conditions that make it possible.
Change direction
Instead of asking ourselves “How do I detect if a student has used AI?”, we could ask ourselves “How do I design an assessment in which using AI without learning is pointless?”.
This means, for example, designing tasks where the value lies not only in the final result, but also in the process followed. Or proposing activities that require interaction, context, or decision-making that cannot be easily delegated.
It’s not a simple or immediate solution. But, unlike detection, it addresses the core of the problem : a rethinking of assessment methods. And this, while uncomfortable, can be an opportunity.
Author Bios: Pharaoh Llorens Largo is Professor of Computer Science and Artificial Intelligence at the University of Alicante and Marc Alier Forment is Professor of Information Systems and Social Aspects of Information Technologies at Universitat Politècnica de Catalunya – BarcelonaTech
