A Trap for the Advanced Student: How to Break the Habit of Blindly Trusting Neural Networks

© HSE University

Andrei Ternikov, Associate Professor at the St Petersburg School of Economics and Management at HSE University–St Petersburg, has developed a method for conducting online exams that significantly limits students’ ability to use ChatGPT and other AI models to obtain correct answers. Andrei Ternikov spoke to the HSE News Service about his approach—which won the HSE University Autumn Educational Innovation Competition, received an Alfa Future grant, and was presented at an international conference in Japan.

Not to Forbid, but to Complicate

Over the past three years, generative artificial intelligence (programmes such as ChatGPT, YandexGPT and their equivalents) has become accessible to virtually everyone. The way such systems work is simple: you enter a question in plain text, and within seconds they produce a detailed, well-formulated answer. In everyday life, this is a convenient assistant; for education, however, it poses a serious challenge, as students have begun using neural networks during exams.

A typical scenario looks like this: a student opens a test, copies the question, pastes it into ChatGPT or a similar tool, receives the answer, and transfers it back. The entire process takes 10–15 seconds. Meanwhile, the lecturer has no way of determining whether the student actually understands the material or is simply adept at using the clipboard. The problem is particularly acute in online courses, where it is impossible to place an invigilator next to every candidate.

Of course, one could take the route of prohibitions: blocking access to neural networks, introducing screen-monitoring systems, requiring students to sit exams on camera, and so on. In my view, however, it is far more productive not to ban AI, but to make mindless copying useless—or at least more difficult.

‘You May Use AI, but You May Not Paste Its Answers’

At HSE University in St Petersburg, I teach an introductory course on AI for non-technical students—future economists and managers. Banning them from using neural networks in a course devoted to neural networks would be absurd. Therefore, the exam rule was formulated as follows: ‘You may use AI. You may not mindlessly copy questions and paste its answers.’ Unfortunately, some students interpreted this as an invitation to cheat.

The final exam takes the form of a multiple-choice test: students must answer 14 questions in 20 minutes, without the ability to return to previous ones. Visually, the test appears ordinary, but ‘under the hood’ each question contains hidden code. When a student copies a question, certain letters and words are imperceptibly altered. Although the question looks normal on the screen, the clipboard—and subsequently ChatGPT—receives a distorted version of the text. As a result, the neural network produces an incorrect answer.

Andrei Ternikov

Photo courtesy of Andrei Ternikov

In a similar way, sentences with the opposite meaning are automatically mixed into the copied text. For example, a question about the advantages of a technology acquires additional phrases about its disadvantages when copied. In some cases, so-called ‘forced hallucination’ is used: the neural network confidently produces a plausible but entirely incorrect answer, because the hidden code in the question deliberately steers it towards faulty logic during the interaction.

Other techniques were also implemented—for instance, blocking the copying of text on the page, intercepting the clipboard (when a student attempts to paste copied content, a substituted text or image appears), concealing the content of a question when screenshot keys are pressed, and introducing interactive elements that cannot be fed into a neural network through simple copying.

All these traps are triggered only when there is an attempt to mechanically transfer the question text into an external application. A student who has understood the course material and answers independently can complete the test without any difficulty.

The Most Valuable Skill

The method was tested over two academic years on a sample of more than 900 students, and the results proved convincing. The distribution of marks in the online test approached a normal curve (without a clear bias towards either very high or very low scores). By contrast, in standard online tests without protection against AI, marks are typically skewed towards higher scores, as most students simply obtain ready-made answers from neural networks.

The analysis showed that around 70% of students attempted to use AI for direct copying of questions. It was precisely in this group that marks were noticeably lower: the hidden traps had worked. Students who relied on their own knowledge, or who used neural networks thoughtfully (as a prompt for reflection rather than as a crib sheet), consistently achieved high results.

For students, such exams became not merely a test of knowledge but also a valuable learning experience. They saw first-hand that blindly trusting a neural network can lead to mistakes. In my view, this is one of the most important skills that an introductory course in artificial intelligence can provide.

Relevance and Originality

Experts from the HSE Fund for Educational Innovation highlighted several strengths of this methodology. In addition to its clear relevance in the context of the rapid spread of generative AI, they noted the originality of the approach (AI is used not only as a subject of study but also as a tool for assessing knowledge), its strong technical design, and its scalability: the method can be adapted to a wide range of disciplines and platforms.

Experts from the Alfa Future programme for lecturers—under which I received a grant as an innovative educator with achievements in teaching and academic work—regarded my developments as a systemic contribution to the advancement of higher education. Thanks to this grant, I was able to attend in person the 12th Asian Conference on Education and International Development (ACEID2026), held in Tokyo at the end of March this year, where I was the only representative from Russia.

Notably, colleagues from the international academic community saw in the proposed methodology something more than just a technical solution for a single course. First, they highlighted its universality (the approach can be applied to any online assessment where there is a risk of dishonest use of neural networks); second, its practicality (in addition to implementation examples, I provided guidelines for adoption). Third, they were surprised that the methodology had been developed and implemented by a single individual.

The Element of Surprise

Of course, the drawbacks of the method were also discussed across various platforms, and I do not deny their existence. To begin with, in order to apply it, a lecturer must still have a basic understanding of how a web page functions and be able to insert ready-made code snippets. It is not a matter of simply installing and running a programme. However, I am happy to help colleagues master the approach. As part of the Digital Teaching Consultants project, I recorded a training video, which is available in the Knowledge Base of Digital Consultants, accessible to all university staff.

At present, the methodology can only be applied to multiple-choice questions or questions where the answer is a word or a number. Traps for open-ended questions or essays are more difficult to design and automate on a large scale, although nothing is impossible here either. For example, one could insert extraneous information into the question text, which a student inclined to breach academic integrity is likely to reproduce in their essay.

When using this method, the element of surprise is crucial. If a dishonest student learns—say, from senior peers—that a lecturer employs it, they will most likely try to avoid being caught and devise ways to circumvent it (and such ways do exist, especially if the exam is taken from home without supervision). That said, I would hope that in such a situation the student would instead choose to prepare properly and sit the exam honestly.