Generation of problems: how Russia is struggling with the use of AI in theses
Experts studied almost 13,000 theses for 2013-2026 and found out how massively students use neural networks to write final qualifying papers (WRC), which sections are most often written using artificial intelligence (AI) and which neural network is in demand more than others for their creation. The issue of criteria for evaluating final work remains acute — a regulatory document of standard recommendations for universities is planned to be released by the fall. Details can be found in the Izvestia article.
How to distinguish a real WRC from an AI fake
The share of AI-generated text in students' graduation papers increased from 9.9% in 2022 to 42.3% in 2025, a fourfold increase in less than three years, according to a new study by the special platform. Retext.ai (available from Izvestia).
The analysis of the AI detector showed that the use of neural networks in the text of student papers is unevenly distributed. The highest "AI density" was recorded in the conclusions (about 56%) and introductions (about 49%) — these sections are written according to standard templates, and the AI for writing a thesis copes with them the easiest. The main part of diplomas, where students describe empiricism, data and their own calculations, turns out to be less susceptible to generation — here the share of artificial intelligence is about 41%.
The leaders in the use of neural networks are graduates of economics, business and IT fields (up to 60% of the "AI share"). Humanities and natural sciences specialties show a more restrained growth — up to 15%. The survey showed that the proportion of jobs with almost no signs of artificial intelligence has fallen from about 70% in 2022 to 23% in 2025.
The experts managed to establish that the distribution of generator models in 2024-2025 shifted towards OpenAI: the query "how to write a diploma using ChatGPT" describes the most common scenario. Other LLMs (large language model) — Gemini, Deepseek and Llama - also occupy a significant share. At the level of individual papers, there are also combined scenarios where the introduction for the diploma is generated by one neural network model, and the conclusion and speech for the diploma defense are generated by another.
"We noticed that the demand for search queries like "neural network for writing a diploma" and "AI thesis" has doubled in two seasons of defenses," says Olga Shkryaba, one of the authors of the study. — And in 2026, there is no longer a question of whether a student uses AI, but how exactly.
In Retext.ai It is predicted that, while maintaining the dynamics of the last three years, the share of research papers written using AI will reach 50-60% in 2026. However, analysts point out that this does not mean "mass plagiarism" — some of the work will only be edited by AI tools. A certain percentage is usually transferred, but the other, unfortunately, is almost completely generated.
March of the Imitators
Currently, a whole class of platforms has emerged on the market that generate term papers and theses in a matter of minutes with automatic registration according to GOST, imitation of bibliographic lists and built-in detector bypass tools.
Plekhanov Russian University of Economics explains that currently the demand for freelancers who write papers for students is shifting towards services that require deep subject expertise, non-standard analytics, adaptation to the unique requirements of a particular teacher or assistance in preparing for oral defense, that is, where neural networks are still ineffective.
—Our university and other universities will have to complicate assignments, introduce unconventional individual trajectories and oral verification formats in order to make ordering finished work economically impractical," says Marina Kholod, a leading researcher at the Laboratory of Artificial Intelligence, Neurotechnology and Business Analytics at Plekhanov Russian University of Economics.
At the same time, Andrey Druzhinin, an expert at Rosobrnadzor and adviser to the Rector of the Moscow State University of Economics, notes that at the moment there are already precedents for litigation around neural network simulations of work. A fourth-year student was expelled for a significant proportion of the AI-generated text in the WRC, the court supported the position of the university.
— At the same time, this year there were massive complaints about false alarms of detection systems: students of Plekhanov Russian University of Economics, KemSU and a number of other universities faced a situation when the system labeled self-written works as AI—generated, - said the interlocutor of the editorial board.
An illustrative case is that a KemSU student was forced to rewrite her thesis 13 times, prepared without using neural networks. The situation gained such resonance that representatives of the Federal Assembly had to intervene in it.
In response to the growing tension, the University Consortium of Big Data Researchers, together with Antiplagiat, established an expert council on academic ethics to develop common standards and standard local acts for Russian universities.
At the same time, the Ministry of Education and Science of the Russian Federation refrains from strict federal regulation: the agency supported the initiatives of universities to create ethical codes, making it clear that there would be no single mandatory rules. The current national standards (GOST R 59277-2020, GOST R 59895-2021, etc.) form a terminological and methodological framework, but are only advisory in nature. Izvestia sent a request to the Ministry of Education and Science of the Russian Federation, but at the time of publication of the article, no response had been received.
According to Anti—Plagiarism, in 2023, signs of the use of AI were recorded in 5.3% of student papers, in 2024 — already in 17.8%, in 2025 - in 24% (among the WRC, the figure was 22.7%). According to data for the first quarter of this year and the same period last year, "the share of student papers containing signs of generative tools increased from 17% to 27% (about 60% year-on-year growth), with almost the same volume of verified papers at 1.15 million."
The demand for traditional "gray" services for writing works manually
The massive segment of "gray" services will be under pressure from AI and is likely to "evolve" to the level of working with more complex services, where substantive expertise is required, says Evgeny Lukyanchikov, executive director of Antiplagiat. The writing sector, estimated by experts at 10-12 billion rubles, is experiencing a structural crisis due to the sharp increase in the use of neural networks and AI agents.
The organization's experts note a qualitative transformation in demand: the average volume of student papers has increased from 45 to 48 pages, indicating a demand for more complex, structured materials. Antiplagiate emphasizes that the market is not disappearing, but "evolving": simple performers are being replaced by algorithms, and demand is shifting towards services that require deep subject expertise and human intelligence, where neural networks are still ineffective. Thus, "gray" authors are forced to improve their skills, moving from writing texts to complex compilation and analysis.
However, it is important to distinguish between tools and violations: by itself, working with AI is not an academic crime; the problem becomes passing off the generated text as your own independent work. But how to catch those who abuse AI? And where is the measure of the use of neural networks? The UNESCO report notes that "at this stage, most higher education institutions have adopted only framework ethical principles for the use of AI, but monitoring and verification mechanisms remain the subject of discussion, since reliable technical solutions for identifying content created using AI do not yet exist."
Russian universities are trying to "fight" the problem in their own way. At the Higher School of Economics, for example, in addition to the Declaration of Ethical Principles for the Creation and Use of Artificial Intelligence Systems (adopted in 2024), the documents governing the rules of the educational process include procedures that require students to declare the use of AI, and teachers are given the authority to cancel the assessment if undeclared use of generative models is found.
"And these standards work," says Anna Korovko, Senior Director for Basic Educational Programs at the Higher School of Economics. — For example, the share of theses with declared use of AI in 2025 was about 20%. At the same time, according to the plagiarism verification system, there were about 21% of texts with traces of AI. This indicates the acceptance of the rules of the game by the students.
In turn, the St. Petersburg State University Graduate School of Management has adopted an AI Application Policy that contains a quantitative norm unique to Russian practice. If the teacher has not set their own rules, the student should limit the amount of generated content to 20% of the total work. The document is of a framework nature and assumes that teachers will develop their own memos. Without notifying the teacher, the use of AI may be considered a violation of academic integrity — a form of plagiarism.
The compact rules for the use of artificial intelligence tools in the educational process, based on the principle of "default allowed", were also approved by the Faculty of Economics of Moscow State University. The student is required to specify the model, the parts of the work, and the purpose of the use. It is allowed not to declare AI if it was used "exclusively for information retrieval" — not for content creation, retelling, analysis, or comparison.
The teacher has broad powers: he has the right to demand detailed prompta and restrict the use of AI up to a complete ban. But the key norm is different: "Due to the imperfection of modern means of detecting AI, the final decision on the correctness of the use of AI cannot be made solely on the basis of automatic verification and is always the result of an expert assessment by the teacher." It is recommended to restructure the point rating system so that the student cannot successfully complete the discipline, "completely delegating all work to generative AI," including the introduction of oral "blocking" control elements.
At the moment, universities have mostly created only ethical declarations on the extent to which students use AI when creating works. And in general, so far everyone agrees that it is not so easy to "catch" both students and teachers. The Moscow State Pedagogical University has adopted a memorandum on mindfulness. MIPT adheres to industry self-regulation instead of internal regulations. The RANEPA has approved a "Policy of academic integrity and the use of artificial intelligence." And this is perhaps the most legally formalized document. Its structure is closer to the compliance policy of a corporation than to the university's code of ethics.
As of 2026, RGGU is working on the creation of an internal code of ethics. Dmitry Guzhelya, Vice—rector for Development of this university, even publicly outlined the scale of the problem: "In universities, AI is exploited almost universally by both teachers and students. Another thing is that it is institutionalized in only a few organizations, while the rest do it independently and not always correctly."
Do I need to change the criteria for evaluating student papers due to the spread of AI
Several key measures need to be taken to solve existing problems, says Alexander Kondakov, President of the Institute of Mobile Educational Systems LLC, former Deputy Minister of General and Vocational Education of the Russian Federation, Corresponding Member of the Russian Academy of Education. First, in his opinion, it makes sense to develop uniform regulations for the use of AI in the system of continuing education — and not at the level of individual universities or schools. Such a document should define acceptable boundaries for the use of neural networks in the educational process, establish norms of behavior in the media space, set requirements for the ethical use of technology and eliminate legal uncertainty.
Secondly, it would be reasonable to create media awareness and literacy in the field of AI among all participants in the educational process. The new PISA—2029 concept will help in this, which will measure a person's ability to behave correctly and safely in the media space and effectively create and consume content.
Perhaps, according to Ivan Karlov, head of the Laboratory of Digital Transformation of Education at the HSE Institute of Education, one of the answers lies in the transition to such formats as "startup as a diploma" or "project as a diploma". In such cases, the use of generative artificial intelligence is not so critical, because it is not the text that is important, but the actual result.
Anti-Plagiarism believes that the educational system needs uniform, understandable rules.
— To do this, we, together with leading universities and partners, are launching an expert council on the academic ethics of AI applications, within which we will prepare standard recommendations for universities. We expect that such a document will be ready by the autumn of 2026," Evgeny Lukyanchikov said.
Переведено сервисом «Яндекс Переводчик»
