“AI is here to stay.” “Banning is not the solution.” “Technology is not the problem, but teaching people how to use it well.” “What matters is learning, and AI can help.” “We cannot become AI police.” The above statements are becoming common mantras in the debate about artificial intelligence (AI) in education, especially in public university teaching. But none of them helps address the underlying problem that AI poses to us academics: how to continue guaranteeing, in its presence, that the credentials issued by the university actually capture the competencies and knowledge they aim to capture and are, therefore, reliable for the rest of society, which is the final recipient of the public goods that the university allegedly produces. Because, love it or not by all faculty, that is one of the basic functions of a university, especially if it is public: we have the responsibility before society to ensure as much as possible that graduates have acquired the knowledge and capacities they are supposed to have to practice their profession, and have earned the qualifications recorded in their academic transcript. Who, if not them, should guarantee such a thing?
Today, AI allows any student to simulate performance and real learning, even in traditional exams, without the teacher being able to detect it or, if suspected, without being able to prove it in case of claims: cheap and accessible technologies such as glasses, cameras and AI-connected earpieces already operate as invisible digital “cheat sheets”. The imperfections that evaluation methods used so far undoubtedly have do not make the question of AI’s emergence less urgent nor can they serve as an excuse for inaction. There is no perfect test or evaluation system, but that has not stopped universities from always establishing controls and sanctions for those who attempt to cheat, simulating a level of learning they have not actually acquired: it has happened with other technologies, such as calculators, personal computers or mobile phones.
“We have the responsibility before society to ensure as much as possible that graduates have acquired the knowledge and abilities they are supposed to have”
Of course the debate is not about whether to “ban” AI (nor any other technology), since this can be a useful tool in many senses, but how to guarantee that it is not used to cheat, falsifying actual performance and deceiving the university and society. And that guarantee concerns the university, which, in this sense, performs a role akin to driving-license examiners, notaries, tax or labor inspectors, or any public official who must verify the truth and authenticity of a document or application. We all know that the issue is not the technology itself, but its use: the problem is what that legitimate use of AI is and how to guarantee it.
But AI has upended all traditional methods of control and detection of academic fraud, to the virtual impossibility of guaranteeing the validity of tests and to avoiding a massive level of false positives. Currently, if we use traditional methods (essays, papers, exams or written exercises), the uncomfortable truth is that we cannot know how reliable our evaluations are, or we know even less than before. Ignoring that the problem exists and looking the other way does not seem acceptable.
The incentives that fuel AI-based academic fraud
The most obvious temptation, in this situation, is to rely entirely on values and voluntary compliance of students. But social science already knows enough about under what conditions it is more or less likely to expect that voluntary compliance, so as not to trap ourselves in stereotyped and speculative debates between Rousseau and Hobbes. To summarize: voluntary compliance can be expected when there is little at stake or when there are incentives that reward it; conversely, it is necessary to control and sanction non-compliance when there is a lot at stake and the potential incentives do not offset the cost of complying. It would be naïve to pretend that paying taxes is voluntary.
“AI has upended all traditional methods of control and detection of academic fraud”
Unfortunately, the arrival of AI presents universities with a devilish combination in this respect. It is hard to ignore that students have a great deal at stake when attempting to pass courses and obtain the degree with minimal time and effort; real learning may be a sufficient compensation for the cost of compliance only in vocational cases that in many degrees tend to be minority, and add that AI makes it possible to simulate competencies indefinitely over time. If, in addition, universities themselves openly reiterate that they find it impossible to detect and demonstrate AI use, this reinforces a predominant strategy in that direction by students.
By contrast, universities stand to lose little: if students pass and graduate without having learned enough or with a low-quality education, the costs fall on the rest of society, not on them, while the costs of policing and enforcement, which were assumed fairly easily before AI, are prohibitive in the current context, where claims could grow exponentially. Raising the bar, increasing control, and detecting fraudulent uses of AI carries high costs. Universities lose little or nothing by being permissive about AI use: their students graduate and they, in turn, enroll. Compare this permissiveness with the bureaucratic paranoia applied to every euro spent by a researcher on their competitive projects: there is money and potential penalties from the Ministry, a lot at stake.
“Universities lose little or nothing by being permissive with AI use: their students graduate and they enroll”
It could be argued that universities are also a reputational quasi-market, and if their students have faked their learning, that will hurt them in that market. But in many countries, like ours, student and resource allocation does not depend so much on reputation. Moreover, professors tend to gain reputations more from their research and knowledge transfer than from teaching. On top of that, the simulation of competencies can continue indefinitely in many professions, so reputational damage is contained.
In this situation, the equilibrium that will tend to emerge is straightforward: total freedom in practice, for both students and teachers to use and permit the use of AI, and this even if there are recommendations or declarative, non-binding prohibitions that no one effectively enforces. This is roughly the situation we are currently experiencing on numerous university campuses. The obvious question is whether this is sustainable without a colossal (collective self-deception) and without unleashing a spiral toward absurdity: why couldn’t, then, teachers delegate evaluation and grade assignment to AI? (In fact, they are already mass-producing scientific articles largely drafted with AI). It is inevitable to question what would be the university’s added value in this context.
How can the university assess real learning in the AI era?
In my view, there are two possible and compatible alternatives to this neolibertarian university stance on AI: either deploy a technology capable of reliably detecting fraudulent AI use, or devise a trustworthy learning test that is immune to such use. Until recently it was claimed that the first did not exist, but Pangram may be changing that. As for the second, the only option available to us is the face-to-face oral interaction between student and instructor, but one can doubt its viability as a generalized method: firstly, because AI-enabled technologies can quickly adapt to fraudulently address that option as well; secondly, because its implementation requires many more resources in the public university, enabling smaller groups and personalized attention; finally, the test’s validity would be conditioned by personality biases, such as oral eloquence or the student’s timidity.
“In this situation, the equilibrium toward which the practice will tend is clear: total freedom in practice, for both students and teachers to use and permit the use of AI”
In short, the university needs to urgently and decisively explore the two alternatives mentioned: AI-detection technologies and tests immune to AI, but, instead, I fear what we will witness is a trend toward the strategic equilibrium I described: the student simulates learning while using AI, and the university pretends to believe it, approves it and graduates them. Will we regret it?
