# Translationese formal register Part of [The AI Tells Index](https://feedsquad.com/ai-tells). A tell signals low effort. It does not identify an author. Skilled writers produce every shape listed here, some of them daily, and automated detectors misread those writers for it at rates measured above 60 percent on non-native English prose. Nothing in this index proves that a machine wrote anything. Read an entry as one piece of evidence to weigh against the false-positive notes printed beside it. ## Facts - Id: translationese-formal-register - Category: Model register (model-register) - Subcategory: non-english-register - Also known as: formal register in a casual context, non-English output that reads as translated - Status: Active. Currently signals low-effort writing. - Severity: medium - Evidence grade: community-observed (named by practitioners, no formal measurement exists) - Languages: id - Added: 2026-08-15 - Updated: 2026-08-17 - Page: https://feedsquad.com/ai-tells/translationese-formal-register ## Description Non-English output arrives in the highest formal register available, whatever the context asks for. Formal pronouns in every sentence, verb phrases turned into noun phrases, no discourse particles anywhere. It reads as a translation of a formal English source. Often it is one. ## Why it reads as machine-written The post-2022 lexical shift shows up outside English. A preprint covering 34 languages, committed to EMNLP 2026, reports uptake in 26 of them with a mean prevalence increase of 15.1 percent. What that work does not supply is a shape-level list per language. The only worked non-English taxonomy we hold is for Bahasa Indonesia, and it comes from a system prompt written to steer models away from these habits, which makes it inverted evidence of the habits. Two of its observations any speaker can test: the formal second-person pronoun Anda held in every register, and the absence of the particles that ordinary Indonesian carries. No equivalent list exists yet for Finnish, German, French or the other languages the multilingual measurement covers. Inventing one would be worse than naming the gap. ## Detection Type: judge (a rubric for a lightweight model judge) Rubric for a lightweight model judge: ```text Read the text in its own language. Decide the register it is written in, then the register its context calls for. Return a hit only when the text sits at the top formal register while the context is casual, such as a social post, a reply, or a message to a customer. Markers to scan for, with Indonesian as the worked case: the formal second-person pronoun used in every sentence with no shift, verb phrases converted into noun phrases where a plain verb exists, and no discourse particles anywhere in a conversational passage. Escape hatches, all returning clear: government, legal, academic and institutional writing, which are formal by requirement; a language or genre with no casual written form; an audience the writer addresses formally by convention. Return the register you judged, the context you judged it against, and one span that shows the mismatch. If you do not read the language, return nothing. ``` ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > Anda dapat melakukan pengecekan terhadap status pesanan Anda melalui halaman berikut. Kami senantiasa berupaya memberikan pelayanan terbaik bagi seluruh pelanggan kami. Kepuasan Anda merupakan prioritas utama bagi perusahaan kami, dan hal tersebut menjadi dasar dari setiap perbaikan yang kami lakukan. Kami mengucapkan terima kasih atas kesabaran dan pengertian Anda. After, repaired: > Kamu bisa cek status pesanan di halaman ini. Nomor pesanan ada di email konfirmasi, biasanya masuk 10 menit setelah bayar. Note: Indonesian, customer message. The formal version nominalises the verb and holds Anda; the repair drops to the register a customer writes in and adds where the order number is and when the email lands. Figures in this repair are invented for the specimen. ### Specimen 2 Before, exhibiting the tell: > Pelaksanaan pembaruan sistem akan dilakukan pada hari Senin. Pembaruan tersebut merupakan bagian dari komitmen kami untuk terus meningkatkan kualitas layanan bagi seluruh pengguna. Kami berharap seluruh pengguna dapat memaklumi kondisi tersebut. Kami mohon pengertian Anda atas ketidaknyamanan yang mungkin terjadi. After, repaired: > Sistem kami update hari Senin, jam 2 sampai jam 4 pagi. Selama itu kamu masih bisa lihat pesanan lama, tapi belum bisa checkout. Note: Same language, maintenance notice. The repair states the window and what still works during it. Figures in this repair are invented for the specimen. ## False positives Who legitimately writes this way. Civil servants, lawyers and institutional communications staff write the top register in every language they publish in, and that is the correct output for their genre. Second-language writers are taught the formal form first and many hold it, which is exactly the population that detection tooling already fails hardest on. Older writers in several languages read informal public address as rude, and some genres have no casual written form at all. The judgment has to be made by somebody who reads the language, against the real audience. A reviewer who does not read it should return nothing instead of guessing, and a translation into English destroys the evidence before the question is even asked. ## Model attribution No product can be named from it. The observation set we hold describes models writing Indonesian in general, and the source that supplies it is a steering prompt and not a measurement, so product-level attribution is unavailable and should not be inferred. ## Sources 1. Juzek, LLM lexical uptake across 34 languages, preprint (arXiv:2605.25358) https://arxiv.org/abs/2605.25358 (tier: primary-doc; accessed 2026-08-14) 2. anti-slop-writing system prompt and pattern list https://github.com/adenaufal/anti-slop-writing (tier: community; accessed 2026-08-14; inverted evidence, cited for what it catalogues rather than what it advocates) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-15, Active: Active and graded community-observed on 2026-08-15. The multilingual measurement establishes that the shift is not English-only, at 26 of 34 languages, and it is a preprint committed to EMNLP 2026. The shape-level detail comes from a single steering prompt for Bahasa Indonesia, carried as inverted evidence and never as sole support. No rate has been published for the register itself, so the entry claims a pattern and not a prevalence. ## License CC BY 4.0. https://creativecommons.org/licenses/by/4.0/ Attribution: The AI Tells Index, feedsquad.com/ai-tells Reuse the data, including commercially. Keep the attribution line. --- Dataset version 1.0.0. Schema version 1. Part of [FeedSquad](https://feedsquad.com). Built by [Herman Foundry](https://hermanfoundry.com) from Levi, Finnish Lapland.