The issue is not just about academic integrity. It is also about how assignments are designed. Consider an analytical paper on CBAM, the EU's Carbon Border Adjustment Mechanism. A student prepares a policy brief on the mechanism's impact on export-oriented industries in the Eurasian Economic Union (EAEU), draws on UNCTAD data, presents tables, and formulates conclusions. The final document may look polished and even convincing. But if the grade is based solely on that document, the instructor cannot determine whether any verification took place: whether the student cross-checked model-generated estimates against official statistical databases, accounted for logistics costs, exchange rate fluctuations, and sector-specific exemptions, or simply incorporated a plausible-sounding conclusion produced by an algorithm.
Simply disclosing the use of AI does not solve this problem. A generic statement such as "the model was used to help with wording" explains nothing. Conversely, when an assignment includes a record of the decisions made along the way, even the legitimate use of AI tools no longer undermines the learning outcome.
What kind of evidence matters? Not just any evidence. What is useful is evidence that reveals the process of verification and decision-making. A list of sources the student used to verify the data. A brief record indicating which claims generated by the model were accepted, which were rejected, and why. An intermediate draft or revision history showing how the argument evolved. And a specific description of AI's role: not simply "AI was used," but rather "the model helped formulate a hypothesis, after which the conclusion was verified against specific databases and adjusted to account for particular limitations."
This approach changes the logic of assessment. Instead of focusing on the polished appearance of the final text, the instructor evaluates the student's ability to work with information: to distinguish verified facts from plausible claims, identify weaknesses in calculations, and explain why an initial conclusion was abandoned. Students, in turn, come to understand that pretending to work independently requires more effort than demonstrating the actual course of their work.
There is also a downside. Requirements to submit intermediate materials increase the workload, not only for students but also for instructors. If every draft, prompt, and comment has to be reviewed manually, assessment will quickly become unmanageable. The goal, therefore, should not be maximum documentation but a minimum meaningful set of materials: a one-page verification report, a decision table, and a brief reflection. Additional requirements can be determined within individual courses: sometimes a source log will be necessary, while in other cases a half-page commentary will suffice.
The easiest way to start is with a single course, rather than waiting for university-wide guidelines covering all programs. The instructor announces in advance that, along with the final policy brief, students must submit a verification table and a short note explaining which conclusions generated by the model they accepted, which they rejected, and why. After one semester, the results can be evaluated. Has the number of submissions that are difficult to assess on their merits decreased? Are students relying more heavily on primary data? Or have they simply learned to attach a meaningless checklist? Such a trial is more useful than abstract discussions about the future of education.
The rules should take the form of a short checklist. What must be submitted. Where it will be stored. Who will have access. And, most importantly, what students will receive credit for when they transparently document their work with an AI model.
What deserves assessment is neither the statement "AI was used" nor the polished structure of the final report. The focus should be on decisions: why a particular database was selected, why a model-generated conclusion was rejected, and how that rejection changed the calculations. In this way, the model becomes an ordinary tool, alongside spreadsheets, calculators, and academic articles. It has just one distinctive feature: without a record of verification, its contribution is indistinguishable from work done by someone else.
Author: PhD in Economics, Associate Professor, Department of World Economy and International Finance, Financial University under the Government of the Russian Federation, Elizaveta Valentinovna Ogloblina.