The Core Paradox of University Oversight
The academic community's attempts to strictly filter student papers based on the "percentage of generated text" have reached a dead end. A practical experiment involving three popular verification services revealed that the exact same text, evaluated across different platforms and at different times, can be classified as either completely authentic or almost entirely machine-written. Discrepancies between algorithms frequently reach 30–40%.
At the same time, the detectors demonstrate surprising consensus on another front: they unfailingly identify authentic human writing. This suggests a fundamental reassessment of the entire anti-plagiarism methodology in the age of neural networks is urgently needed.
The Scale of the Problem: Explosive Growth in Three Years
According to statistics from the Antiplagiat service, the educational landscape is shifting rapidly:
- In 2023, traces of generative AI were detected in only 5.3% of student research papers.
- By 2026, this figure surged to an impressive 36%.
Universities responded to this challenge by introducing mandatory entry filters. But just how objective are the tools that decide the fate of academic degrees?
Who Was Tested and How
For comparative analysis, three tools actively used in academic environments were selected: domestic platforms Antiplagiat and Doomeit (the latter already implemented at MGIMO and Synergy University), along with the international tool GPTZero, widely adopted across foreign colleges.
The test pool consisted of three types of Russian-language materials:
- Reference human texts: grant applications and thesis papers from leading universities (TSU, TyumSU, UrFU) written strictly before 2022.
- Generated duplicates: copies of the same texts produced by the GigaChat model based on detailed prompts.
- Real-world 2026 cases: student papers from open sources and final thesis abstracts.
Detector Behavior: Strictness vs. Leniency
The two-month tests revealed substantial differences in how these algorithms evaluate Russian-language content:
- Antiplagiat: displayed excessive leniency toward students in the summer. A completely synthetic counterpart of a real thesis received only a 53% AI score. However, by October the database was updated, and sensitivity to generated text leaped by 30–35%. A paper that would have easily passed review in the summer was guaranteed to be sent back for revisions in the fall. The cost of a private check here starts from 483 rubles per 50,000 characters.
- Doomeit: proved to be a far stricter evaluator. It flagged an artificial analog of a scientific article with over 98% accuracy. The platform swiftly stabilized its models following a major August update and currently delivers the most reproducible results. A pleasant perk of the service is three free daily checks.
- GPTZero: showed absolute stability in its results throughout the entire testing period. Instead of calculating a percentage of fragmented matches, the tool estimates the overall probability of content origin. Notably, the service is extremely cautious with accusations: even on a fully machine-generated text, it might assign a 60–70% AI probability, categorizing the remainder as mixed to prevent false positives against genuine human authors.
Why the AI Percentage Is a Poor Metric for Thesis Defense Clearance
Continuous retraining of detection models causes a thesis paper's evaluation status to change in just a month without a single edit made by the student. The variance between different services renders any numerical thresholds set by universities (such as "no more than 20% AI") completely arbitrary.
The conclusion is obvious: evaluating student papers solely based on the share of neural network involvement is flawed. Semantically identical texts can receive a 30% score in one system and 70% in another.
A New Concept: Searching for the Human, Not Hunting the Robot
Rather than hunting for generated paragraphs, experts recommend evaluating the proportion of guaranteed human authorship. All detectors agree on one point: if the material is genuinely written by a human, machine markers do not exceed 10%, while originality and human phrasing metrics stay comfortably above 70%.
To make the assessment fair and transparent, adjusting the verification protocol is all it takes:
- Exclude formulaic sections: abstracts, introductions, and conclusions frequently rely on standard academic clichés that trigger detectors into falsely flagging them as machine-generated.
- Set aside bibliographies and literature reviews: using AI to structure open sources and perform preliminary research is entirely justifiable, serving as a digital assistant.
- Evaluate only the core contribution: the author's own results, calculations, methodology, and practical conclusions. For this essential core, a strict threshold should apply—no more than 10% third-party generation.
Such an approach will protect diligent students from the whims of updating algorithms and spare educators from endless disputes over the inaccuracies of detection tools.