AI and Research Activity: Where Should the Boundary Lie Between Human Thought and the Algorithm?
TSU Rector Eduard Galazhinskiy has once again been confirmed in the position of Deputy Chair of the Higher Attestation Commission of the Russian Federation for the next five years. This edition of the blog presents an analytical text prepared by him, outlining the main reference points of the current situation connected with the mass arrival of generative agents, multifunctional AI algorithms, in the spheres of dissertations and academic publishing. The text is intended not only for experts, but for a broad readership as well, namely for all members of the Russian academic community.
Analytical Memorandum
On the Permissible Limits of the Use of Generative AI in Dissertation and Publication Processes in the Social Sciences and Humanities
1. Initial diagnosis
The system of academic certification throughout the world, including Russia, has entered a new phase of risk. Generative AI, represented by a wide range of specific generative agents, is now being used not only for language editing and technical assistance, but also for building outlines, compiling literature summaries, preparing reviews, writing evaluations, conducting preliminary editorial screening, and providing technical support for peer review. A 2025 Russian review published in the HSE University journal Language and Education notes that discussion within the academic community about the use of AI in the publication process has already shifted fr om general anxiety to normative work around transparency in the use of AI, human accountability for its use, and constant human oversight. At the same time, the practical operationalization of these principles remains insufficient. In the international publishing environment, this shift is confirmed by the 2026 editorial position of Nature Methods: the responsible use of generative tools in scholarly publishing is now viewed not as an optional matter, but as a systemic one.
For the Higher Attestation Commission, this means the following: the problem is no longer limited to text borrowing. A more complex threat has emerged, namely the substitution of an independent research contribution with plausible machine-generated output. This is especially dangerous in the dissertation process, because certification confirms not simply the coherence and correctness of a text, but the author’s personal contribution, novelty, inner unity of the work, and ability to bear responsibility for their own scholarly judgment. These requirements are закреплены? Better translate: enshrined in the current Regulations on the Awarding of Academic Degrees. Russian civil law directly links authorship of a scholarly work to the citizen whose creative labor produced it.
2. Why the risk in the social sciences and humanities is higher than it seems
In the social sciences and humanities, generative agents affect not only the form of the text, but the very fabric of justification and argumentation. Research here is often built through the interpretation of sources, reconstruction of context, conceptual distinction, historiographical positioning, and responsible reading of another person’s text. This is why even a stylistically flawless text generated by AI may turn out to be scientifically empty or, more dangerously, misleading. The same 2025 Russian review emphasizes that the real areas of application of generative tools already include idea generation, outline construction, summarization, and synthesis, that is, operations close to the core of the author’s own contribution. A 2025 study by Korean scholars published in Technology in Society, devoted specifically to the social sciences, reaches an even more important conclusion: ethical decisions cannot be developed in the abstract; they must be embedded in concrete disciplinary and institutional orders.
Fr om this, one may conclude that in the social sciences and humanities the chief threat is not the crude replacement of the author by the machine, but a much subtler erosion of scholarly subjectivity. A generative agent is capable of smoothing out contradictions, creating an appearance of historiographical completeness, offering plausible-sounding but standardized formulas, and filling logical gaps with rhetorical fluency. The result is the danger of an institutional error: mistaking textual maturity for the maturity of scholarly thinking.
3. The international situation: what has already become fact
Around the world, generative agents are rapidly becoming part of the publishing infrastructure. In March 2026, Springer Nature reported that in 2025 more than 1.5 million manuscripts passed through nearly 60 AI-based tools built into screening, editorial evaluation, and research integrity checks. More than half of this publisher’s journals already use its own peer review support platform based on such tools. By the end of 2026, more than 2 million manuscripts at this publisher will have passed through AI. In other words, this is no longer an experiment, but the industrial integration of artificial intelligence into the publishing cycle.
Even more important is the fact that machine assistance has entered peer review. A 2026 article in Nature Machine Intelligence, based on an analysis of more than 20,000 conference paper reviews, showed that automated feedback for reviewers prompted 27 percent of them to revise their reviews, incorporating 12,000 recommendations generated by AI. Blind assessment confirmed that the revised reviews receiving such feedback were significantly more informative. This marks an important threshold: the machine is no longer only helping to write academic text, but is beginning to structure the language and logic of expert evaluation itself.
At the same time, an even more troubling tendency is strengthening: the self-referentiality of generative agents. A 2026 Nature article on the automation of the full research cycle described a system that generated ideas, wrote code, conducted computational experiments, produced figures, wrote the article, and carried out its own peer review. One paper created entirely by AI passed the first stage of selection at a major conference and was withdrawn only at the special request of the researchers conducting the experiment. This means that generative agents have already come close to closing the entire loop: the production of the text, its preliminary evaluation, and its peer review processing are beginning to belong to one and the same technological family.
4. Self-referentiality as a new institutional danger
For the Higher Attestation Commission, self-referentiality is not an abstract philosophical category, but a direct practical problem. If a generative agent helps a degree candidate write a literature review, then a similar type of agent helps the editorial office preliminarily assess the article in a journal from the approved list, and then yet another artificial agent helps the reviewer draft their evaluation, scholarly communication gradually closes in on a machine circle. This creates the risk of circular verification: a text built according to a machine logic of plausibility receives positive signals in procedures that themselves rely on machine criteria of coherence, completeness, and formal persuasiveness. Real abuses are already being documented: in 2025, Nature reported cases in which hidden instructions for machine reviewing tools were inserted into manuscripts in order to induce a more favorable assessment.
For the social sciences and humanities, this danger is especially high. As noted above, unlike in most natural-science disciplines, a significant part of evaluation here is built not on laboratory experiments, but on work with texts. This is precisely why the self-referentiality of generative agents is capable not merely of accelerating the work here, but of deforming the very criterion of quality, shifting it from depth of thought to flawlessness of presentation.
5. Specific risks for the dissertation process
With regard to dissertations, dissertation abstracts, reviews, and formal evaluations, six central risks can be identified.
The first is the hidden generation of the semantic core of the work. A generative agent can not only improve style, but effectively propose the formulation of the problem, the hypothesis, the novelty, and the conclusions. For a dissertation in the social sciences or humanities, this is equivalent to intrusion into the very heart of the author’s contribution. Russian law, which recognizes as the author of a scholarly work a citizen rather than a technical system, provides grounds for considering such hidden substitution incompatible with the principle of independence.
The second is the fabrication of bibliography. The experience of many users shows that generative systems are capable of producing plausible but nonexistent bibliographic entries, as well as entries containing substantial errors. For the humanities and social sciences, this is not a secondary flaw, but the destruction of the very process of proof: a false reference here often undermines not an individual fragment, but the entire structure of the research.
The third is the machine standardization of dissertation language. If a significant part of the text is produced with the help of a generative agent, the likelihood increases that an averaged, impersonal style will appear. For the Higher Attestation Commission, this is directly relevant, because research in the social sciences and humanities must demonstrate not only correct wording, but also the author’s own work with concepts and sources.
The fourth is the compromise of reviews and evaluations. As soon as generative agents begin to be used by opponents, reviewers, and editors without transparent rules, the risk emerges of weakening genuine expert responsibility. A special study has shown that even reviewers themselves see threats here related to privacy, bias, and opacity, despite recognizing the possible usefulness of AI.
The fifth is false accusations based on detector results. Two major studies from 2023, published in Patterns and Computer Science, as well as a 2026 study in Nature Machine Intelligence, confirmed both a troubling increase in suspicion toward generative AI in the academic environment and the fact that current systems for detecting generated text are not absolutely accurate or reliable. In particular, popular machine-text detectors systematically produce false positives on English-language texts written by authors for whom English is not a native language. In other words, simpler and less lexically diverse English turns out to be statistically “suspicious” to the algorithm, as a result of which the system detects not generated text, but simplified English. One may assume that similar errors can also occur with Russian-language texts written by authors for whom Russian is not their native language. For the Higher Attestation Commission, this leads to a direct conclusion: an indicator identified by machine means cannot serve as independent proof of academic misconduct. Results from detectors used to identify generated text can only serve as grounds for a deeper substantive review.
The sixth is the overloading of the scientific certification filter itself. If generative agents accelerate the production of outwardly acceptable texts, then the flow of publications and manuscripts increases, and it becomes difficult to distinguish them quickly from genuinely independent work. One of Nature’s 2026 articles directly warns that the automation of the full cycle is capable of placing additional strain on an already overburdened peer review system.
6. The Russian situation: what already exists and what is still missing
In Russia, the problem has already entered the normative field. Since January 1, 2025, Russian National Standard R 71657–2024 has been in force, directly concerning the creation of scholarly publications using artificial intelligence. In 2025, the Association of Scientific Editors and Publishers adopted the second edition of its Declaration of Ethical Principles for Scientific Publications. It explicitly states that authors bear responsibility for the use of artificial intelligence, that AI cannot be an author or co-author of a scholarly article, and that editors must make decisions independently of machine recommendations. In the same year, the Association held a special roundtable devoted to the first experience of regulating and monitoring generative artificial intelligence in scholarly publications. This means that the professional environment has already formulated its basic guidelines.
But in Russian certification practice, there is still no sufficiently detailed typology of what is permissible and what is not. The Regulations on the Awarding of Academic Degrees set out the criteria of independence, novelty, and internal unity of the dissertation, but do not specifically describe the situation of AI use. At the same time, the current requirements for journals on the Higher Attestation Commission’s approved list still proceed from an earlier model of scholarly publication: they concern peer review, the presence of abstracts, keywords, and bibliographic lists, but do not contain any direct requirement for journals to have a public policy on the use of generative agents by authors, reviewers, and editors. Consequently, this is precisely wh ere an institutional gap exists today.
7. The normative position the Higher Attestation Commission would be well advised to adopt
It would seem that for the Higher Attestation Commission, especially in the social sciences and humanities, it is fundamentally important to adopt not a prohibitive position, but a differentiating one.
The first norm: a generative agent cannot be recognized as an author, co-author, official reviewer, opponent, or subject of expert judgment in state scientific certification procedures. This conclusion is sufficiently grounded both in Article 1257 of the Civil Code and in the general logic of the Regulations on the Awarding of Academic Degrees.
The second norm: the use of AI that does not affect scholarly judgment as such should be considered permissible. This may include spell-checking and punctuation, formatting, technical verification of bibliographic descriptions, and translation, provided that the result is mandatorily checked by a human being.
The third norm: the use of generative agents for the initial selection of literature, the categorization of material, draft planning, and language polishing should be considered permissible, but only if this fact is mandatorily disclosed. However, even here full responsibility for the correctness of facts, references, quotations, and interpretations must remain with the author. This logic corresponds both to the 2025 Russian Declaration of Ethical Principles for Scientific Publications and to the international movement toward transparency and human oversight over AI.
The fourth norm: hidden use of generative agents is impermissible for writing literature reviews, formulating novelty, conclusions, the theoretical and methodological section, reviews, opponent evaluations, dissertation council conclusions, and expert opinions. It is precisely at these stages of the dissertation process that not technical, but scholarly and institutional responsibility is manifested.
8. Practical steps the Higher Attestation Commission can take without changing the underlying law
The Higher Attestation Commission can already initiate five practical decisions.
The first is to introduce a mandatory declaration of AI use for dissertations, dissertation abstracts, and publications included in a certification file. In this declaration, the author should indicate whether generative agents were used, at what stage, and for what tasks. This would not be a prohibition, but a new standard of transparency. Such a measure is directly consistent with both international and Russian approaches to accountability.
The second is to develop unified methodological recommendations for dissertation councils and the Commission’s expert councils on the use of AI in the dissertation process and on the impermissibility of substituting the author’s scholarly contribution with the results of generative agents. The main risk today is that different councils will act differently: some excessively leniently, others suspiciously and arbitrarily.
The third is to update the requirements for journals on the approved list. Since the Ministry of Science and Higher Education order of May 31, 2023, as amended in 2025, already sets out the mandatory requirements for such journals, it would be logical to supplement it with the requirement to have a public policy on the use of generative agents by authors, reviewers, and the editorial office. This does not require dismantling the system; it simply requires bringing it into line with the new publishing reality.
The fourth is to state directly that automatic generated-text detectors cannot serve as an independent basis for concluding that a violation has taken place. They may be used only as an auxiliary indicator, to be followed by substantive review: analysis of references, quotations, drafts, comparison of the author’s oral and written discourse, requests for explanations, and requests for working materials. The need for such caution is indicated both by research on detector bias and by recent publications on the growth of suspicion as a specific pathology of the academic environment.
The fifth is to initiate a special interdisciplinary study aimed at identifying the changes that generative agents are now introducing into the dissertation process. At present, we have mainly international cases, but very few Russian studies, and there is also no sufficient corpus of Russian-language certification texts for precise diagnosis. Without this, any regulation will remain partly speculative.
9. Forecast for the Higher Attestation Commission in the coming years
In the short term, over the next two or three years, generative agents will definitively become part of everyday authorial and editorial practice. The growth in their use will proceed faster than the growth of normative clarity. This is already evident both fr om the international publishing infrastructure and from the speed with which generative feedback has entered peer review.
In the medium term, on a five- to six-year horizon, the main object of dispute will no longer be the fact of use itself, but the boundary of the author’s contribution. The focus will shift from the question “may one use it?” to the question “what exactly can still be considered independent scholarly work?” In the social sciences and humanities, this line will run first of all through interpretation, historiography, conceptual work, and expert judgment.
In the long term, on a ten-year horizon, the most likely outcome is a mixed model: AI infrastructure will accompany almost the entire path of the manuscript, but the legality and legitimacy of scientific certification will remain only wh ere the human being continues to bear final judgment and responsibility. If this is not fixed in advance, the dissertation process risks facing two distortions moving toward one another: on the one hand, the inflation of plausible but weak text; on the other, arbitrary suspicion and distrust toward any manifestation of an unusual authorial style. Both are destructive for the Higher Attestation Commission.
10. Final conclusion
For the Higher Attestation Commission today, it is fundamentally important to call the problem by its proper name. What is at stake is not simply a new technology for creating scholarly text. What is at stake is a redistribution of functions within scholarly communication: generative agents, as quasi-subjects of scientific cognition, are beginning to participate in the creation of the text, in its editorial processing, and in its peer review evaluation. This is wh ere their most dangerous property manifests itself: self-referentiality. Science begins to look at itself in the mirror rather than beyond the horizon. For the social sciences and humanities, this is especially risky because the object of certification here is not only the text as a result, but also the intellectual effort that stands behind it. International practice is already moving toward a regime of transparency, accountability, and human oversight; the Russian professional environment has also articulated this logic, but the certification system has not yet brought it to the level of unified procedures.
Therefore, the strategic task of the Higher Attestation Commission is not to prohibit the technology as such, but to protect the zones of inalienable scholarly responsibility. A person may use artificial intelligence. But scholarly judgment, novelty, interpretation, peer review, and the certification decision must remain unconditionally human. If this boundary is drawn clearly, the Higher Attestation Commission will be able not only to respond to the challenge, but to strengthen anew the very meaning of scientific certification.
Rector of TSU Eduard Galazhinskiy,
Member of the Council for Science and Education under the President of the Russian Federation,
Vice President of the Russian Academy of Education,
Vice President of the Russian Union of Rectors,
Deputy Chair of the Higher Attestation Commission of the Russian Federation