The share of materials in Russian media with clear signs of neural network generation has grown from zero before the advent of ChatGPT in late 2022 to 18% by the end of Q2 2026, according to a study by Kirill Pshinnik, co-founder of the online university Zerocoder, titled "Newsrooms After ChatGPT," presented at the Russian publishers' forum "Publishing/Praktikum."
Using AI tools, the author analyzed 254,614 articles from 174 Russian publications from 2020 to the end of June 2026. Short news items were not included in the sample. The publications were divided into three tiers: 13 federal socio-political and business outlets, 99 second-tier publications (thematic, lifestyle, and major regional), and 62 city and niche outlets. Each article was assessed by two open-source AI text detectors — OpenPangram-2B and Qwen3-4B-Detect. An article was considered "AI" if the average score of the two models was at least 0.5. The GigaCheck detector (Sber) was used for verification. It is not included in the index because since 2025 it has been less accurate at recognizing texts from new models. In Q2 2026, the index detectors classified 14% and 23% of articles as AI, while GigaCheck classified 8.3%. The final index was 18%.
Until the end of 2022, the share of AI articles remained within 0–0.08% per quarter: neural networks for mass text generation did not yet exist, and these values indicate the level of false positives. After the release of ChatGPT on November 30, 2022, the indicator grew every year: 0.88% in Q4 2023, 5.7% in Q4 2024, 14% in Q4 2025, and 18% in Q2 2026, according to the report.
According to the study, the share of AI texts depends not so much on the status of the publication as on the topic: neural networks most often write about home, food, and health, and almost never touch on politics and incidents. By publication tier, the second tier leads — thematic, lifestyle, and large regional media: one in four to five articles there shows signs of AI, while in federal outlets it is one in ten. The market is heterogeneous: almost half of the publications have less than 5% AI articles, while one in five has more than 30%.
Direct speech is disappearing from publications along with AI, states Kirill Pshinnik. In articles with signs of neural network generation, direct quotes are 2.1 times fewer, and expert remarks are 2.6 times fewer than in human-written texts, according to his observations. Such materials are shorter: on average 773 words versus 1,090. A similar picture is seen within newsrooms: where AI articles exceed 20%, experts per 1,000 words decreased by 22% compared to 2020–2022, while in publications with almost no AI, it increased by 6%. "This is a correlation, not a proven cause," the study authors emphasize.
Neural network language is also penetrating human-written texts, notes Kirill Pshinnik. For example, 11 words appear 4.1 times more often in AI articles than in human ones: "key," "form," "balance," "critically," and others. In Q2 2026, they are present in all articles 1.9 times more often than on average in 2020–2022. Even in texts that detectors consider human, such words increased by 13%. The share of the gray zone — texts with AI signs below the threshold, likely edited by a neural network — grew from 1.7% to 27%.
"Neural networks have already become a common tool in newsrooms, and the question is not whether to use them, but what disappears from the text along with manual work. Our data show: where there are more AI articles, direct speech from experts and officials is heard less often," says Kirill Pshinnik. "The model will quickly gather background information, but it will not get a comment or call a source back."
