LLM Writing Quality by Language — Turkish

Raw research notes for Turkish, 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-turkish.
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-Turkish assessment of LLM prose splits sharply by genre and vintage. The most-cited failure is captured in a phrase that recurs across journalism and publishing: AI “Türkçeye çevirebiliyor ama Türkçeleştiremiyor” — it can render text into Turkish but cannot make it Turkish. Working translators and editors converge on “üslupsuzluk” (stylelessness) and “yavan” (bland/flat) as the core diagnosis: semantically correct output in a colorless register, poor with idioms, argot, and wordplay. Yet the single most rigorous native study (Prof. Gökhan Tunç, Anadolu University, published 2025–2026) found GPT-4.5 pastiches of Sait Faik, Ferit Edgü, and divan poetry fooled majorities of Turkish-literature students and experts — imitation of canonical style is a demonstrated strength even as original literary voice is judged absent (“Dümdüz yazıyor” — “it writes flat,” per novelist Dilge Güney, March 2026). Informal native communities (Ekşi Sözlük) have crystallized concrete tells: the em dash alien to Turkish typography, uniform sentence rhythm, and noticeably worse quality when generating directly in Turkish versus translating from English. Crucially, almost no serious Turkish critic has published a named assessment of the late-2025/2026 frontier tier (Opus 4.5/4.6, Fable 5, GPT-5.x, Gemini 3.x); the 2026-dated material that does name those models is SEO-adjacent tech blogging, which rates their Turkish as “workable to natural” while flagging English-syntax drift in long literary text. Coverage is genuinely thin at the intersection of {frontier vintage} × {credentialed literary judgment} — that gap is itself a finding.

Sources

  1. URL: https://www.farukbildirici.com/2024/02/12/yapay-zeka-cevirilerinde-komik-hatalar/ (also ran at T24 and Gazete Duvar) Author: Faruk Bildirici — Turkey’s best-known media ombudsman (former Hürriyet okur temsilcisi, ex-RTÜK member); T24 “Medya Ombudsmanı” columnist. Outlet: own site / T24. Date: 2024-02-12. Models: unnamed AI translation tools used by Turkish newsrooms — vintage flag: pre-frontier (early 2024). Quotes: “yapay zekâ… Türkçeye çevirebiliyor ama Türkçeleştiremiyor” — “AI can translate into Turkish but cannot naturalize it into Turkish.” Examples: “training wheels” rendered literally as “eğitim tekerlekleri” where a human would write “destek önlemleri” (support measures); “sorumlu” (responsible) flipped to “suçlu” (guilty) in BBC Türkçe copy. Claim: AI-translated Turkish news is recognizably machine-made — literal idiom transfer and register errors that a human translator would never commit.

  2. URL: https://kayiprihtim.com/roportaj/yapay-zeka-destekli-edebiyat-cevirisi/ (context: https://www.milliyet.com.tr/kultur-sanat/yapay-zekayla-ceviri-tartismasi-6978149) Author: Faruk Akhan, publishing director of Dedalus Kitap — the publisher at the center of Turkey’s biggest AI-translation scandal (nine books AI-translated under pseudonyms). Interview by Hakan Tunç. Outlet: Kayıp Rıhtım (established genre-literature site). Date: 2023-07-15 — vintage flag: DeepL-era, pre-frontier. Quotes: “Makinenin çevirirken en büyük eksikliği üslupsuzluk.” — “The machine’s greatest deficiency in translation is stylelessness.” “Çoğu zaman anlam olarak doğru olsa da yavan denebilecek Türkçe karşılıklar verdi.” — “Though usually semantically correct, it produced Turkish equivalents one could call bland.” Claim: Even the pro-AI publisher concedes machine Turkish is semantically right but stylistically dead — weakest on author voice, wordplay, argot, and idiom; the surrounding controversy (translators/editors uniformly hostile; Haber Global headline “Makine çevirir ama Türkçeleştiremez”) is the profession’s verdict.

  3. URL: https://www.iyikitap.net/2026/03/01/yuvarlak-masa-yapay-zeka-yasamadan-anlatabilir-mi/ Author(s): Roundtable — Dilge Güney (lawyer, novelist: Mavi Yıldız, 1 GB Adalet), Toprak Işık (author-engineer), moderated by author Barış İnce. Outlet: İyi Kitap (children’s/YA literature review magazine), issue 157. Date: 2026-03-01. Models: ChatGPT named generically — current-era but versions unspecified. Quotes: Güney: “Edebî metinde cümle yazabilecek seviyede değil şu an… Dümdüz yazıyor şu anki hâliyle.” — “It’s not at the level of writing a literary sentence yet… In its current state it writes flat.” But also: “Bütün ustaları… okuyacak. Ve çok iyi taklit edecek.” — “It will read all the masters… and imitate them very well.” Işık: “Duygusu yok… dolayısıyla onunki daha mekanik kalmak durumunda.” — “It has no emotion… so its writing is bound to stay mechanical.” Claim: Practicing Turkish authors in 2026 judge LLM Turkish fiction prose flat and mechanical at the sentence level, while expecting imitation ability to keep improving.

  4. URL: https://www.haberturk.com/edebiyatin-dijital-ciraklari-turk-edebiyati-nin-ustalarini-gecti-mi-3881798 and https://dergipark.org.tr/en/pub/akaded/article/1708572 Author: Prof. Dr. Gökhan Tunç, Turkish Language and Literature Dept., Anadolu University. Outlets: Habertürk (2026-05-02) reporting his study; peer-reviewed article in Akademik Dil ve Edebiyat Dergisi (2025-07-20). Model: ChatGPT/GPT-4.5 — vintage flag: early-2025 model, published 2025–2026. Findings/quotes: In a blind test with 160 Turkish-literature students, 40.6% misidentified Ferit Edgü’s real story as AI, and only 29.4% caught the AI pastiche of Edgü; participants rated AI prose “daha okunur/daha anlaşılır” (more readable/understandable) and even “daha insani” (more human). Poetry study: expert judges misattributed AI divan poetry as human 88.3% of the time; contemporary free verse was the weakest genre — “uzman katılımcıların çağdaş Türk şiirini kimin yazdığını belirleme konusunda kararsızlık içinde oldukları” (“expert participants were undecided about who wrote the contemporary Turkish poems”). Claim: The strongest empirical native evidence that frontier-adjacent models already produce Turkish literary pastiche indistinguishable from the canon to trained readers — strongest in rule-bound classical forms, weakest in modern individual voice.

  5. URL: https://tezmakale.com/ai-detector/ Author: Aybars Kutay Dağlı, TezMakale (Turkish academic-editing/AI-detection service). Date: updated 2026-06-13. Models: ChatGPT, Gemini, Claude, Copilot named generically — current-era. Quotes: “Cümle uzunluklarındaki çeşitliliği analiz eder. Düşük varyans, mekanik yazım sinyali olabilir.” — “It analyzes variance in sentence length; low variance can signal mechanical writing.” “Edilgen fiil yapılarının ve belirsiz özne kullanımının yoğunluğunu ölçer” — “It measures the density of passive verb structures and indefinite subjects.” Also flags over-regular agglutinative suffix patterns and formulaic -dır/-mektedir copulas. Claim: A practitioner-codified inventory of what AI Turkish looks like in 2026: uniform sentence rhythm, passive overuse, over-regular morphology — a commercial artifact of the failure modes natives perceive.

  6. URL: https://eksisozluk.com/chatgpt-uzun-cizgisi—8048066 ; https://eksisozluk.com/chatgptye-makale-yazdirip-yakalanmak—7712165 ; https://eksisozluk.com/dedalus-kitapin-yapay-zekayla-kitap-cevirmesi—7683292 Author: Ekşi Sözlük users (anonymous native speakers; Turkey’s largest collaborative dictionary/forum). Date: ongoing, threads active 2025–2026. Models: primarily ChatGPT; versions mixed. (Direct fetch blocked, 403 — characterized via search excerpts; quotes paraphrased, not verbatim.) Key points: the em dash (uzun çizgi) “doesn’t exist on Turkish keyboards, isn’t part of Turkish punctuation tradition” and has become the signature of AI-contaminated Turkish text; users report that asking ChatGPT to write directly in Turkish yields prose so poor “you don’t even need a detector,” while English-first-then-translate evades detection; readers describe a “yapay zekâ kokusu” (AI smell) in careless machine prose. Claim: The informal native consensus — direct Turkish generation is visibly weaker than the models’ English, with typographic and rhythmic tells natives spot instantly.

  7. URL: https://vidoport.com/blog/claude-vs-chatgpt-vs-gemini-hangisi-daha-iyi-2026-karsilastirma-rehberi (similar: https://celiluyanikoglu.com.tr/chatgpt-vs-claude-vs-gemini-kapsamli-karsilastirma-2026) Author: unbylined Turkish tech-blog guides — low-credential but the only found Turkish-language assessments naming the actual frontier tier. Date: February 2026. Models: Claude Opus 4.6 / Sonnet 4.5, GPT-5.2 / GPT-4o, Gemini 3 Pro/Flash — on-scope vintage. Quotes: “ChatGPT genel Türkçe konuşmada en dengeli performansı verir” — “ChatGPT gives the most balanced performance in general Turkish conversation.” “Claude resmi ve kurumsal yazışmalarda akıcıdır” — “Claude is fluent in formal and corporate correspondence.” “Üçü de iş yapılabilir seviyededir” — “All three are at a workable level [in Turkish].” The celiluyanikoglu guide adds (per search excerpt, unverified by direct fetch) that Gemini 3.1 Pro scores 73.6% on Turkish MMLU and that in long academic/literary Turkish “structures that slip into English syntax can be observed.” Claim: 2026 frontier models’ Turkish is rated functional-to-natural for utilitarian registers, with residual English-syntax drift in long-form literary text — but this is tech-blog, not literary, authority.

  8. URL: https://erkansaka.net/2025/10/14/turna-ai-kumru-2b-launch-reactions-turkey-llm/ (context: https://blog.monsternotebook.com.tr/kumru-ai-incelemesi/, https://eksisozluk.com/kumru-ai—8034656) Author: Prof. Erkan Saka, İstanbul Bilgi University (media anthropologist), compiling Turkish community reactions. Date: 2025-10-14. Models: Kumru-2B/7B (VNGRS, Turkish-only trained) vs. multilingual models. Key points: Turkish users found Kumru outperformed much larger multilingual models (LLaMA-70B, Qwen-72B, Gemma-27B) on purely Turkish tasks — summarization, grammar correction — validating the “Türkçe tokenizer” thesis that multilingual models handle Turkish morphology inefficiently; but Kumru fell far behind frontier models on reasoning (confident arithmetic errors, missing the interrogative when a question mark was omitted where ChatGPT inferred it). Claim: The domestic-LLM comparison point — Turkish-native training buys fluency on mechanical Turkish tasks that even large multilingual models lack, while frontier models retain a decisive inference advantage.

Also noted: developmental psychologist Dr. Tolga Yıldız (İstanbul University) in K24 argued “dil modellerinin metin üretmesi değil, giderek bir yazı normu üretmeye başlaması” is the real problem — “the real issue is not that language models produce text, but that they are increasingly producing a norm of writing” (located only via search excerpt; K24 tag page: https://www.k24kitap.org/etiketler/yapay-zeka). Ankara Edebiyat Dergisi ran a “Yapay Zeka ve Edebiyat” dossier in its July–August 2026 issue (via İttifak Gazetesi), confirming the literary press is engaging, though quote-level content wasn’t retrievable.

Failure modes observed

Praise / strengths noted

Evidence quality & gaps