Swapping the Letter Yo and Guillemets Cut a Thesis's AI Score in Antiplagiat from 64% to 0%
Russian AI-text detectors nitpick typography: em dashes, the letter yo and guillemet quotes (the so-called "herringbone" quotes). In a philology student's thesis, the author replaced them with hyphens, the plain letter ye and straight-style curly quotes. Antiplagiat then showed 0% AI instead of 63.86%, and Doumate 28.17% instead of 70.12%. The text stayed the same; only the symbols changed. So an AI percentage does not prove authorship, and it is better to normalize typography before a check.
AI-processed from Habr AI; edited by Hamidun News
A student writes a thesis on their own, defends it with top marks, and then the university's detector finds a quarter of the text to be AI. The instructor doubts it, the student protests, and there is no way to check who is right. Experiments that an author described on Habr point to one cause of such stories.
The Russian AI-text detectors Antiplagiat and Doumate react not only to words but also to formatting: em dashes, the Russian letter yo and guillemet quotes (the Russian "herringbone" quotes). When the author replaced these symbols with hyphens, the plain letter ye and curly quotes (“”), Antiplagiat's score for a graduation thesis fell from 63.86% to 0% AI, and Doumate's score fell from 70.
12% to 28.17%. Not a single word in the text changed.
How It Started: Four Dashes and 100% AI
After the author's first article, a reader named Kamil Gadeev contacted her. He took a generated text of about 150 words. Doumate rated it 100% AI. Kamil replaced just four em dashes with hyphens, and the score became 0% AI.
The author first assumed the short length was the issue. She asked Doumate's support team what size is needed for a reliable score. The answer: preferably about 10 pages. In the author's view, the developers are playing it safe, and at least half of that is realistically enough. Support also confirmed that the company has seen cases where an em dash affects the result. To her knowledge, there are no published studies on the subject.
The Experiments: The Letter Yo, Guillemets and Double Translation
The second case involved the author's own text. Doumate rated the original, written by a human, at 0% AI with 90% originality. The text was then translated into English and back into Russian using an AI chat (an RU-EN-RU chain). The AI share rose to 51.1%, while originality barely changed.
The author asked Claude to compare the two texts character by character. Besides stock word substitutions, it found typographic changes: the chat put the letter yo in place of the letter ye and guillemet quotes in place of the author's curly “” quotes. The author manually switched 33 instances of the letter yo back to ye and 12 pairs of guillemets back to curly quotes. The AI share fell from 51.1% to 16.83%, even though the text remained exactly the same.
Next, they checked three works by a philology student from 2024, 2025 and 2026, provided by her academic supervisor. The student often writes the letter yo and quotes heavily, so her texts contain many guillemets. In the 2026 graduation thesis, 33 instances of the letter yo, 54 guillemet quotes and 4 em dashes were replaced. The results:
- Doumate: 70.12% → 28.17% AI;
- Antiplagiat (version 1.9.2 dated 10.11.26): 63.86% → 0% AI;
- GPTZero is almost insensitive to Russian typography: the difference was a few percentage points.
Where the Limits Are
Typography is not a magic button. In the 2025 term paper, swapping the symbols changed almost nothing: the AI share stayed around 50%. The author concludes that here the detector reacted to content, which has many signs of AI generation. In her model, the final score is made up of two kinds of signals, content-based and formatting-based: they either add up, or one absorbs the other. Typography decides the outcome only if there is a lot of it in the text and there are few objections to the content.
The sample is tiny: one author, one student, three works. This is not a peer-reviewed study, and the detector developers did not confirm the results beyond a general acknowledgment that the dash has an effect. The author considers the student's rising AI share from year to year a possible sign of closer work with chatbots. That is a supposition, not an established fact.
What to Do in Practice
For those building products on text classifiers, the lesson is broader than student theses. A detector trained on model-generated texts easily latches onto surface features such as dashes and quotation marks, because chatbots use them more often than people do. Such a feature can be changed in a second. A person with a philologist's habit of using the letter yo and guillemets falls under suspicion, while someone who generates text can strip those symbols out and get a low score.
If you check other people's texts, do not take the AI percentage as proof, and look at the individual highlighted fragments. If your own text is being checked, the author advises replacing em dashes with hyphens, the letter yo with ye, and guillemets with curly quotes before submitting. The goal, in her words, is not to fool the detector but to help it avoid false positives. If you develop detectors, normalize such characters before classification, otherwise the verdict will depend on keyboard habits.
- Nov 21, 2026Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
- Nov 21, 2026Ego lite Gives Each AI Agent Its Own Workspace Space in the Browser
Key Questions Answered
How It Started: Four Dashes and 100% AI?
After the author's first article, a reader named Kamil Gadeev contacted her. He took a generated text of about 150 words. Doumate rated it 100% AI. Kamil replaced just four em dashes with hyphens, and the score became 0% AI. The author first assumed the short length was the issue. She asked Doumate's support team what size is needed for a reliable score. The answer: preferably about 10 pages. In the author's view, the developers are playing it safe, and at least half of that is realistically enough.
What are the key details on The Experiments: The Letter Yo, Guillemets and Double Translation?
The second case involved the author's own text. Doumate rated the original, written by a human, at 0% AI with 90% originality. The text was then translated into English and back into Russian using an AI chat (an RU-EN-RU chain). The AI share rose to 51.1%, while originality barely changed. The author asked Claude to compare the two texts character by character.
Where the Limits Are?
Typography is not a magic button. In the 2025 term paper, swapping the symbols changed almost nothing: the AI share stayed around 50%. The author concludes that here the detector reacted to content, which has many signs of AI generation. In her model, the final score is made up of two kinds of signals, content-based and formatting-based: they either add up, or one absorbs the other. Typography decides the outcome only if there is a lot of it in the text and there are few objections to the content.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.