LLM Writing Quality by Language — Indonesian

Raw research notes for Indonesian, 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-indonesian.
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 (4-8 sentences)

Native Indonesian commentary on LLM writing quality is abundant for ChatGPT-generically but almost nonexistent for named late-2025/2026 frontier models (Claude Opus 4.5/4.6, GPT-5.x, Gemini 3.x) — Indonesian critics, academics, and writers overwhelmingly say “AI/ChatGPT” without version, so most assessments must be vintage-flagged as GPT-4/4o-era or unspecified. The consistent native verdict: frontier LLMs now produce Indonesian that is grammatically clean, structurally tidy, and publication-competent at the sentence level — several assessors say output “exceeds students’ actual proficiency” — but stylistically it defaults to stiff bahasa baku (formal register), reads as translated-from-English (imported em-dash habits, parallel-clause rhythms, “terlalu rapi” over-neatness), and lacks lived experience, original diction, and emotional depth. A June 2026 academic study comparing ChatGPT poems with human poets found the machine wins on structure and coherence and loses on diction originality, symbolic complexity, and existential grounding. Literary-establishment voices range from categorical rejection (semiotician Narudin Pituin: AI literature is “secara semiotik cacat dan rendah”) to pragmatic collaboration (poet Hasan Aspahani’s year-long “with AI” short-story experiment; Martin Suryajaya’s corpus-trained poetry machines). Indonesian NLP practitioners frame the gap structurally: Indonesian training data is thin and skewed formal, so models handle baku register far better than colloquial/slang and regional variation. No credible native assessment specifically of Claude Opus 4.5+, Fable 5, GPT-5.x, or Gemini 3.x prose in Indonesian was found; that is a genuine coverage gap, not a search failure.

Sources

1. URL: https://geotimes.id/kolom/ketika-ai-menulis-dilema-dan-kekhawatiran-seorang-pengajar-sastra/ Author: Donny Syofyan — lecturer, Faculty of Cultural Sciences, Universitas Andalas (teaches creative writing and literature). Outlet: GEOTIMES (opinion). Date: 12 Dec 2025. Models: ChatGPT, DeepSeek, Llama 3 (recent but pre-frontier-specific; vintage: late-2025 mainstream). Quotes: He identifies AI tells in Indonesian student prose: overused openings like “Air menahan napasnya” (“The water held its breath”), spaced-less em-dashes, repetitive parallelism (“hanya air. Hanya gerakan” — “only water. Only movement”). “Saya sangat takut bahwa menulis menjadi begitu mudah dengan teknologi ini sehingga tidak ada yang mau melakukannya sendiri” — “I am very afraid that writing becomes so easy with this technology that no one will want to do it themselves.” Claim: AI Indonesian prose is technically publication-quality (grammar, vocabulary above student level) but formulaic, emotionally shallow, and recognizable by imported stylistic tics.

2. URL: https://journal.unram.ac.id/index.php/kopula/en/article/view/10550 Authors: Muhammad Bahrul Ulum, Nur Fuadi, Yusuf Nur Arifin — Universitas KH. Mukhtar Syafaat. Outlet: Kopula: Jurnal Bahasa, Sastra, dan Pendidikan (Universitas Mataram). Date: 20 June 2026. Model: ChatGPT, version unspecified (vintage flag: likely GPT-4o/5-era given date). Quotes: “puisi hasil ChatGPT memiliki keunggulan pada keteraturan struktur, koherensi bahasa, dan kemudahan pemahaman makna” — “ChatGPT poems excel in structural regularity, linguistic coherence, and ease of comprehension.” “puisi karya manusia lebih menonjol pada aspek orisinalitas diksi, kedalaman emosi, kompleksitas simbolik, dan pengalaman eksistensial” — “human poems stand out in originality of diction, emotional depth, symbolic complexity, and existential experience.” “kreativitas yang dihasilkan masih bersifat komputasional” — “the creativity produced is still computational.” Claim: Peer-reviewed Indonesian comparison: AI wins on form/coherence, loses on diction originality and depth in Indonesian poetry.

3. URL: https://www.indonesiana.id/read/186086/sastra-demokratisasi-dan-artificial-intelligence Author: Gunoto Saparie — poet/novelist, chair of Satupena Central Java, head of Central Java AI Creator Forum. Outlet: Indonesiana (Tempo platform). Date: 5 Sept 2025. Models: ChatGPT, Sudowrite, Bing, GPT-3, Cohere (vintage flag: mixed, some older tools). Quotes: On Hasan Aspahani’s year of AI-assisted cerpen: “bukan AI yang menulis, melainkan dirinya yang bekerja keras menyusun prompt, kerangka, dan karakter” — “it wasn’t the AI writing, but he himself working hard building prompts, outline, and characters.” “AI bisa merangkai kalimat, tetapi ia tak tahu bagaimana rasanya menunggu seseorang di tepi senja yang tak kembali” — “AI can string sentences together, but it doesn’t know what it feels like to wait for someone at the edge of a twilight that never returns.” Claim: A senior literary-community figure judges AI Indonesian competent at sentence assembly but authorially empty; the human still does the creative labor.

4. URL: https://jakarta.suaramerdeka.com/nasional/13414494076/… (also mirrored at hatipena.com and kbanews.com — all three intermittently 403; content confirmed via search snippets) Author/subject: Narudin Pituin — sastrawan, literary translator, semiotics scholar (author of Sintesemiotik: Teori dan Praktik, 2023; CSH 2018 poetry award). Outlet: Suara Merdeka Jakarta / Hatipena. Date: Feb 2025. Models: AI generically (vintage flag: unspecified). Quotes: Writing literature with AI is “secara semiotik cacat dan rendah” — “semiotically defective and low”; writing poetry with AI is “haram” — “forbidden” — “sebab tak ada penyair besar yang menulis puisi dengan AI” — “because no great poet has written poetry with AI.” Claim: Categorical establishment rejection: AI-produced Indonesian literature is semiotically deficient regardless of surface fluency.

5. URL: https://tengara.id/blog/penyair-sebagai-mesin-pembacaan-dekat/ Author: Saut Situmorang — poet (seven collections), S2 in Indonesian literature, prominent (combative) literary critic. Outlet: tengara.id (originally Bali Politika). Date: Feb 2024 — vintage flag: pre-frontier, ChatGPT-3.5/4 era. Models: ChatGPT; reviews Martin Suryajaya’s corpus-trained poetry models (Penyair sebagai Mesin, 2023). Quotes: Adopts Nick Cave’s verdict — “Lagu ini sampah” (“this song is garbage”) — for machine poems; rejects them as lacking existential struggle. Notably offers no line-level linguistic analysis. Claim: The strongest-worded critic dismissal of AI Indonesian poetry, but aesthetic/ideological rather than linguistic, and dated.

6. URL: https://medium.com/@feliksmakarios/mengapa-ai-lebih-pandai-berbahasa-inggris-menjelajahi-ketimpangan-bahasa-dalam-teknologi-523d540f6b81 Author: Feliks VP Samosir — informatics lecturer, Universitas Pelita Harapan, NLP researcher. Outlet: Medium (personal). Date: 22 Nov 2025. Models: SEA-LION, Sailor, IndoBERT, Llama 2, mT5, XLM-R, Aya (regional/research models; frontier models discussed generically). Quotes: “Masalahnya bukan hanya soal jumlah data (quantity gap), tetapi juga kualitas data (quality gap)” — “The problem isn’t just data quantity but data quality”; machine-translated training data causes “hilangnya nuansa lokal” — “loss of local nuance.” Notes “SEA-LION ternyata bisa mengungguli Llama 2 dalam tugas berbahasa Indonesia” — “SEA-LION can actually outperform Llama 2 on Indonesian tasks.” Claim: An Indonesian NLP academic’s structural explanation of why LLM Indonesian trails English: thin, formal-skewed, partly machine-translated training data; regional models (SEA-LION/Sahabat-AI) close some of the gap.

7. URL: https://analysis.netray.id/lokalisasi-ai-untuk-terjemahan-yang-akurat/ Author: Winda Trilatifah. Outlet: Netray (Indonesian media-monitoring/NLP firm). Date: 17 Dec 2025. Models: LLMs generically. Quotes: “Mengubah bahasa ‘secara literal’ tidak cukup untuk membangun pengalaman yang inklusif dan profesional” — “Converting language ‘literally’ isn’t enough to build an inclusive, professional experience.” Identifies Indonesia’s register spectrum — “sangat formal di pemerintahan, semi-formal di korporasi, dan sederhana di media sosial” — and regional variation (“Nuansa daerah memengaruhi pilihan kata, cara bertutur, hingga preferensi gaya komunikasi”) as what generic AI misses. Claim: Indonesian industry view: generic-LLM Indonesian answers “what does this sentence mean” but not “how would this sound natural to an Indonesian” — register and regional nuance are the failure axis.

8. Supporting/community: https://medium.com/@ansarassamad/apakah-tanda-pisah-tanda-tanda-ai-45293766bb02 (Ansara Samad, Aug 2025 — “Bahasa Indonesia, secara tata bahasa baku, tidak terlalu banyak menggunakan tanda pisah”: em-dashes read as Anglophone import in Indonesian, though not proof of AI authorship); https://www.threads.com/@shireishou/post/DI-FhmcTrRc/ (novelist community pushback on “em-dash + too-neat = AI” heuristics); https://fkip.umm.ac.id/benarkah-tulisan-chatgpt-mengalahkan-kualitas-tulisan-manusia-ini-kata-ahli-bahasa-fkip-umm/ (UMM language-faculty expert: ChatGPT Indonesian shows markedly lower lexical richness than human writing — site unreachable at fetch time, content via search summary); https://ririsatria.id/2025/07/23/teknologi-dan-transformasi-digital/tantangan-dunia-sastra-di-era-kecerdasan-buatan/ (Riri Satria, UI computer-science lecturer + poetry-community founder, Jul 2025: machine poetry is “konstruksi bahasa” only — “Mesin tidak memiliki pengetahuan tentang dunia nyata dan imajinasi… yang dimiliki manusia ketika membuat puisi” — though his model references are dated, vintage flag).

Failure modes observed

• Default to stiff bahasa baku / over-formal register — output is “terlalu baku dan kaku” (too formal and stiff); models struggle to hold natural colloquial (bahasa gaul/santai) register without heavy prompting (IDN Times, DQLab, erafone, Netray).
• Translated-from-English feel — Anglophone punctuation and rhythm imported into Indonesian: em-dash overuse (tanda pisah is rare in standard Indonesian typography), semicolon-heavy logic connectors, parallel-clause cadences (Ansara Samad; Donny Syofyan; bik.uma.ac.id).
• Lexical poverty / stock diction — measurably lower vocabulary richness than human Indonesian writing (FKIP UMM expert); recycled poetic openings and clichéd closures (Syofyan); “pilihan kata netral” uniformity.
• Over-neatness as a tell — “terlalu rapi, terlalu runut” (too tidy, too orderly), no spontaneity, no typos, uniform sentence structure — widely cited as the community’s #1 AI marker in Indonesian.
• Register/regional flatness — no handling of Indonesia’s formal→casual spectrum or regional speech coloring; loss of “nuansa lokal,” partly because Indonesian training data skews formal/translated (Netray; Samosir).
• Emotional/experiential emptiness in fiction — competent sentence assembly without lived experience: original diction, symbolic complexity, and existential grounding consistently rated inferior to human poets (Kopula study; Gunoto Saparie; Riri Satria; Narudin).

Praise / strengths noted

• Grammar, syntax, and academic-register coherence at or above typical student level; useful for self-editing and academic writing polish (Kopula; KUNKUN journal review; Syofyan grudgingly).
• Structural regularity and comprehensibility — ChatGPT poems beat humans on form and clarity in the June 2026 Unram-journal study.
• Effective as a creative sparring partner rather than author: Hasan Aspahani’s “dengan AI” cerpen collection (some publishers interested) and Martin Suryajaya’s corpus-trained poetry experiments (Penyair sebagai Mesin) show productive human-in-the-loop use.
• Regional models (SEA-LION v3, Sahabat-AI — 50B Indonesian tokens, plus Javanese/Sundanese) outperform general LLMs on localized Indonesian tasks (SUPA/Chemin benchmarks, sea-lion.ai), implying the ceiling for natural Indonesian is rising.

Evidence quality & gaps

• Coverage is real but model-blind. Native assessments are plentiful (national dailies’ platforms, literary organizations, university lecturers, a peer-reviewed 2026 study), but almost none name model versions — “ChatGPT/AI” generically. Zero Indonesian-language native assessments found of Claude Opus 4.5/4.6, Fable 5, GPT-5.x, or Gemini 3.x prose specifically; claims that “Claude writes the most natural Indonesian” come only from English-language comparison blogs and the Artificial Analysis multilingual benchmark, not native critics.
• Fiction commentary skews philosophical over linguistic. Literary-establishment critiques (Narudin, Saut Situmorang) argue authenticity/semiotics, mostly without line-level analysis of the Indonesian itself; the most concrete linguistic tells come from a creative-writing lecturer (Syofyan) and writing-community/SEO-adjacent posts.
• Vintage caveats: Situmorang (2024) and Riri Satria’s model references (GPT-2 era) predate frontier models; the freshest credible items are Syofyan (Dec 2025), Netray (Dec 2025), and the Kopula study (June 2026).
• Paywalls/blocks: Kompas.id’s “Bermitra dengan Kecerdasan Buatan dalam Karya Sastra” (likely the best mainstream literary-press piece) is paywalled; Hatipena/KBA/Suara Merdeka mirrors of the Narudin interview returned 403 — quotes for that item rest on search-index snippets from three independent mirrors.
• No Tirto/Tempo/The Conversation Indonesia critique of frontier-model Indonesian output quality was found; The Conversation Indonesia’s relevant piece is about AI-detector false positives, not generation quality.