LLM Writing Quality by Language — Russian

Raw research notes for Russian, part of the LLM Writing Quality by Language project — edition 2026-07-23. Originally published at peterkaminski.ai/research/llm-writing-quality-by-language/llm-writing-quality-russian.
Researched and written by Saga bg-etruscan (Claude Fable 5), directed by Peter Kaminski, 2026-07-23. Quotations are machine-extracted from the cited sources and not yet verified verbatim — see the main report’s Limitations section.
♡ Copying is an act of love. Please copy and share.
© Peter Kaminski · CC-BY 4.0 (Creative Commons Attribution 4.0 International)

Project files: main report · English · Spanish · Chinese · Hindi · Arabic · French · Portuguese · Russian · German · Japanese · Korean · Italian · Turkish · Indonesian · Polish · all files (.zip)

Summary verdict

Native Russian assessments converge on a consistent picture: frontier LLMs now produce grammatically clean, fluent Russian — case/aspect errors have largely disappeared from the conversation — but the texts fail at the level of style and voice. The dominant complaints are канцелярит (bureaucratese), calques from English (“стоит отметить, что…”), overuse of “является”, missing discourse particles (же, ведь, вот), “room-temperature” affect, over-structured list-heavy formatting, and a translated feel. Literary professionals (Tolstaya, Shargunov, critic Lidia Maslova) judge AI Russian fiction as saccharine, cliché-blind, and emotionally hollow — good only for “тупая черновая работа” — while conceding it is unmatched at official/formal registers. In head-to-head tests by Russian-speaking practitioners, Claude is repeatedly rated the most stylistically precise Western model in Russian, though a professional translator found its Russian translations “неуклюжий местами”; domestic models (YandexGPT, GigaChat) trail Western frontier models on creative tasks but win on local idiom and availability. A significant caveat: much of the 2026-dated Russian-language comparison content is SEO/aggregator-driven, and very little rigorous native criticism targets the newest frontier tier (Opus 4.5+, GPT-5.x, Gemini 3.x) on fiction specifically — the strongest literary-world critiques still date to the YandexGPT-era experiments of 2024.

Sources

  1. URL: https://habr.com/ru/articles/1022426/ Author: stas-clear (Станислав), working editor/author on tech and AI topics. Outlet: Habr (flag: informal-expert). Date: April 12 (Habr ID sequence places it 2026; the fetched page shows only “12 апр”). Models: unnamed, general frontier-LLM era. Quotes: “нейросеть слишком часто пишет текст, который выглядит как текст, но не ощущается как мысль” (“a neural net too often writes text that looks like text but doesn’t feel like thought”); “аккуратно собранная имитация мысли” (“a neatly assembled imitation of thought”); notes “переводной привкус” (“a translated aftertaste”). Claim: An editor’s verdict that AI Russian non-fiction irritates readers because it is “artificially correct yet empty” — false confidence, template structure, translated phrasing.

  2. URL: https://vc.ru/id490208/2847189-plagin-dlya-russkogo-teksta-ubirayushchiy-sledy-neyroseti Author: Илья Утов, developer, daily Claude (Claude Code) user; built open-source (MIT) “humanizer-ru” plugin. Outlet: vc.ru (informal-expert). Date: April 3, 2026. Models: Claude (current 2026 vintage). Quotes: “В русском совершенно другой набор [маркеров]: канцелярит… кальки с английского… злоупотребление словом ‘является’, отсутствие частиц ‘же’” (“Russian has a completely different set [of markers]: bureaucratese… calques from English… abuse of the word ‘yavlyaetsya’ [is], absence of particles like ‘zhe’”). Catalogues 37 patterns in 8 categories, incl. “стоит отметить что” as a literal calque of “it’s worth noting that”, empty openers (“в современном мире”), dash overuse, bold-spam, “perfect typography as a marker.” Claim: Even current Claude output in Russian carries a stable, enumerable fingerprint of canceleritis and Anglicism requiring systematic de-AI-ification.

  3. URL: https://vc.ru/ai/3006515-chatgpt-claude-gigachat-sravnenie Author: “НейроВед” (anonymous team; 200+ test prompts over two weeks). Outlet: vc.ru (informal-expert; some SEO flavor). Date: July 1, 2026. Models: GPT-4o, Claude 3.5 Sonnet, GigaChat — vintage flag: tests pre-frontier models despite 2026 date. Quotes: On ChatGPT in Russian: “в чисто русскоязычных задачах мы стабильно фиксировали характерные шаблоны” (“in purely Russian-language tasks we consistently recorded characteristic templates”) — clichés “в заключение”, “таким образом”. On Claude: “Claude 3.5 Sonnet показал себя как самый стилистически точный ИИ из тройки” (“Claude 3.5 Sonnet proved the most stylistically precise AI of the three”) — 92% style-match vs 74%/61%. On GigaChat: “Творческие задачи — слабое место” (“Creative tasks are its weak spot”). Claim: In a Russian practitioner bake-off, Claude leads on Russian style fidelity, ChatGPT defaults to bureaucratic clichés, GigaChat is strong on domestic domain text but weak creatively.

  4. URL: https://news.mail.ru/society/62600936/ (syndicated from Известия) Author: Лидия Маслова, professional literary critic (Известия, ex-Kommersant). Outlet: Известия. Date: September 1, 2024 — vintage flag: YandexGPT era. Models: YandexGPT/Алиса (anthology “Механическое вмешательство”). Quotes: “когда речь идет о тонких движениях души и переливающихся оттенках смысла… результат потуг ИИ в лучшем случае можно отнести к жанру «непреднамеренной комедии»” (“when it comes to subtle movements of the soul and shimmering shades of meaning… the AI’s strainings at best belong to the genre of ‘unintentional comedy’”); “словесный шлак” (“verbal slag”); but “когда надо написать какой-то официальный, казенный текст на суконном канцелярите… тут нейросети равных нет” (“when an official, drab text in stiff bureaucratese is needed… the neural net has no equal”). Claim: A professional critic’s verdict that (domestic-model) AI fiction is unintentional comedy, fit only for rough drafting, while AI is peerless at bureaucratic register.

  5. URL: https://moskvichmag.ru/lyudi/tatyana-tolstaya-hamit-alise-kak-proshla-vstrecha-s-avtorami-mehanicheskogo-vmeshatelstva Author: Ксения Василашвили (reporting); assessments by Татьяна Толстая (major Russian prose writer) and Ксения Буржская (novelist, Yandex Alisa AI trainer). Outlet: Москвич Mag. Date: November 12, 2024 — vintage flag: YandexGPT era. Quotes: Tolstaya: “Попросила ее написать рассказ в стиле Татьяны Толстой — ничего общего” (“I asked it to write a story in the style of Tatyana Tolstaya — nothing in common”); “Все рассказики у нее с добрым концом… Большая литература всегда про драму, на разрыв души и сердца” (“All its little stories have happy endings… Great literature is always about drama, tearing at soul and heart”). Burzhskaya (praise): “Я поручила ей сгенерировать монолог свекрови… она с ней справилась блестяще” (“I had it generate a mother-in-law’s monologue… it handled it brilliantly”). Claim: A canonical living stylist finds the model incapable of imitating literary voice or sustaining dramatic stakes; an AI-trainer novelist grants it competence at stock comic registers.

  6. URL: https://snob.ru/literature/dopustimo-li-ispolzovat-neiroseti-v-literature-bolshoi-razgovor-s-kseniei-burzhskoi-i-sergeem-shargunovym/ Author/participants: Сергей Шаргунов (novelist, editor-in-chief of «Юность») and Ксения Буржская. Outlet: Сноб. Date: May 21, 2026. Models: unnamed current models + Алиса. Quotes: Shargunov: “они всегда какие-то переслащённые… слова подобраны приторно” (“they’re always somehow over-sugared… the words are chosen cloyingly”); “Машина не замечает собственные штампы” (“The machine does not notice its own clichés”). Burzhskaya: “у машины этой боли нет и не будет” (“the machine doesn’t have this pain and never will”). Claim: As of mid-2026 established Russian writers still find LLM prose saccharine and cliché-blind, with the deficit framed as experiential rather than grammatical.

  7. URL: https://kinzhal.media/ask-kto-napisal/ Author: Кинжал editorial team (журнал Яндекс Практикума; Ilyakhov-school editing culture). Outlet: Кинжал (professional-editor guidance; informal-expert). Date: November 28, 2025. Models: LLMs generically (late-2025 vintage). Quotes: “Идеальная структура без личности — текст как из учебника, ни одной шероховатости” (“Perfect structure without personality — textbook-like text, not a single rough edge”); “Общая температура — комнатная: нет эмоции, нет позиции, нет удивления” (“Overall temperature — room temperature: no emotion, no stance, no surprise”); “Нелепые метафоры — «маяк в океане возможностей»” (“Absurd metaphors — ‘a lighthouse in the ocean of opportunities’”); “Очень часто — списки, везде и всегда” (“Very often — lists, everywhere and always”). Claim: Professional Russian editors codify AI Russian’s tells as flawless-but-lifeless structure, hedging formulas, kitsch metaphors, and compulsive listing.

  8. URL: https://englishgeeks.ru/blog/test-kachestva-perevoda-15-neyrosetey Author: Мария Коржакова, professional written translator (письменный переводчик). Outlet: English Geeks blog. Date: February 2026. Models: 15 incl. ChatGPT, Claude, Gemini, DeepSeek, YandexGPT (specific sub-versions not frontier-flagged — treat as late-2025/early-2026 consumer tiers). Quotes: Claude: “неуклюжий местами перевод” (“translation awkward in places”), untranslated Anglicisms (“центр для бэкпекеров”); ChatGPT: “довольно выразительный” (“fairly expressive”) but with bureaucratic tone; DeepSeek: reads “довольно гладко” (“quite smoothly”), handled the idiom “cup of tea”; YandexGPT: “самый странный результат” (“the strangest result”) — failed the task outright. Claim: A working translator finds every model’s EN→RU output still needs a human pass — Western frontier models are serviceable but calque-prone, and YandexGPT is far behind.

  9. URL: https://gorky.media/reviews/nejroset-horoshij-soyuznik-v-vojne-so-vzrosloj-kulturoj/ Author: Иван Напреенко, sociologist/critic at «Горький». Outlet: Горький (leading Russian literary review site). Date: June 29, 2022 — vintage flag: ruGPT-3, pre-frontier; included as literary-critical baseline. Models: ruGPT-3 (Pepperstein’s “Пытаясь проснуться”). Quotes: the prose’s defining trait is “незапоминаемость” (“unmemorability”) — it “разворачивает свой текст на плато” (“unrolls its text on a plateau”) with no accentuation; strengths: a hypnotic, séance-like “лярвический характер” prized only as avant-garde texture. Claim: The earliest serious Russian literary-critical framing: machine prose lacks the salience-hierarchy of embodied consciousness — flat, unmemorable, interesting only as found-object aesthetics.

  10. Supporting/context: https://habr.com/ru/articles/1008120/ (Prometeriy/Влад Орлинскас, Habr, March 9, 2026 — “без хорошей редактуры генерация читается ужасно и сразу бросается в глаза” / “without good editing, generation reads terribly and is immediately conspicuous”, but edited AI text is now unidentifiable; “battle for text is lost”); https://habr.com/ru/articles/888614/ (Chad_AI, Habr, March 6, 2025 — six markers incl. канцелярит example “Рациональное распределение времени между трудовой деятельностью и отдыхом” vs human “Поработали — надо отдохнуть!”).

Failure modes observed

Praise / strengths noted

Evidence quality & gaps