The debate over how artificial intelligence is transforming the translation industry flared up again in mid-November 2026. While AI is accelerating translation workflows, current research shows that human translators remain essential for conveying cultural context, handling specialist texts and performing the final editorial check. This time, what is being discussed is not a prediction or the opinion of industry representatives; it is two scientific studies, conducted on tens of thousands of real translation segments, that speak with hard numbers.
In this article, we look at the large-scale production test presented at the 2026 conference of the European Association for Machine Translation (EAMT) and the comparative study conducted on legal documents in the United States, and explain step by step what the findings mean for processes involving sworn translation, notarization and apostille certification of official documents.
What Did EAMT 2026's 71,000-Segment Test Find?
The first of the studies was presented at the 26th annual conference of the European Association for Machine Translation, held in Tilburg, the Netherlands, on 16–18 June 2026. The paper, authored by Mara Nunziatini and Mercedes Speroni, is based on a production evaluation of 71,262 segments covering five domains, ten languages and five systems. The most striking aspect of the study is that it was conducted not under laboratory conditions but in a real production environment — that is, on work actually delivered to clients.
Of the 71,262 translation segments produced from English into 10 different languages and across five separate domains, 6,618 were additionally evaluated by 60 professional translators. The researchers developed an "AI post-editing" system that reworks machine translation output using linguistic guidelines and high-quality bilingual examples, and compared it against direct translation by a large language model, as well as output from Google Translate and DeepL. The result is clear: while AI-assisted post-editing methods outperformed the other systems in overall quality, direct translations produced by a large language model failed to reach the same quality level; having the text reworked by a human in line with domain knowledge, terminology and editorial rules proved decisive for translation quality.
Human–Machine Collaboration on Legal Documents: The Rutgers Study
The second study directly concerns anyone dealing with official document translation. In a study by researchers at Rutgers University in the United States, published in the journal "Digital Translation" on 9 July 2026, human translation by 27 professional translators, post-editing of machine translation and AI-assisted methods were compared for the English-to-Spanish translation of legal documents used by US government agencies.
The findings cut both ways. On one side, there is a speed gain: human post-editing of machine translation significantly reduced completion time compared with human translation; GPT-4 also received higher overall quality scores than DeepL and Google Translate. On the other side is the indispensability of the human: machine translation edited by a human was able to reach a quality level comparable to human translation; this demonstrated that AI's strongest use case is preparing the first draft while a human performs the final check, rather than removing the translator from the process. In other words, the machine is the party that starts the translation, not the one that finishes it.
What Exactly Is Post-Editing?
Post-editing — that is, the editing of machine translation output — is, roughly speaking, a three-stage workflow:
- Draft generation: The text is rendered into the target language by machine translation or a large language model.
- Expert editing: A professional translator with domain knowledge corrects terminology, institution names, date and number formats and legal wording by comparing the text against the source document.
- Final review and accountability: Textual integrity, consistency and intended use are checked; in official matters, the person who will sign the translation assumes responsibility for it.
The EAMT findings also explain why this workflow works: simply handing the text over to AI is not enough for a quality translation; terminology and domain knowledge must be built into the process. The correct use of specialist terminology and knowing the purpose for which the text will be used make human oversight necessary. A value in a lab report translated with the wrong unit, or a narrowed statement of authority in a power of attorney, are errors the machine finds "fluent" but an expert would never approve. This is why, in fields such as legal translation and medical translation, having a domain expert perform the post-editing is the critical threshold the research points to.
An Added Dimension for Turkish: Language and Data in the Age of AI
The debate also has a side facing Turkey. An opinion piece published on 16 November 2026 asked whether Turkish can remain a knowledge-producing language in the age of artificial intelligence. Today, millions of people use AI systems alongside search engines to access information; as books, articles and documents form the linguistic world of these systems, a critical inequality is emerging between languages. According to the piece, reliable domain-specific collections of Turkish-language knowledge need to be built.
This observation carries over into translation practice as follows: large language models deliver their strongest performance in data-rich languages such as English. In translation pairs between Turkish and lower-resourced languages (for example Azerbaijani, Persian, or languages of historical texts such as Ottoman Turkish), the margin of error can grow. Even on the EAMT 2026 programme, why large language models fail in low-resource languages was a research topic in its own right. This is why the language roster of an agency working with rare language pairs matters; you can see the scope of these pairs on Ziya Tercüme's translation languages page.
Why Is AI Alone Not Enough for Official Documents?
The studies measure the quality dimension; but official document translation also has a legal validity dimension, and in that dimension AI carries no weight whatsoever. For a diploma, criminal record certificate, court ruling or power of attorney to be accepted in official proceedings at home or abroad, the translation must be signed and stamped by a translator who has taken a sworn translator's oath before a notary; this is the procedure we call sworn translation. If the institution also requires notarization, the sworn translator's signature is certified before a notary; this is a second, separate step, detailed on our notarized translation page. If the document is going abroad, a third link is usually added to the chain: the apostille certificate. A translation agency cannot issue an apostille; in Turkey, this certificate is issued — depending on the nature of the document — by governor's or district governor's offices or by judicial authorities.
The digital side of this chain is also evolving. According to a statement by Abdulkadir Uraloğlu, Minister of Transport and Infrastructure, e-Apostille is a digital service that enables official documents bearing an apostille issued by authorized bodies to be created, approved, transmitted and verified electronically, and citizens can submit applications via the e-Government (e-Devlet) portal, with requests routed to judicial units under the Ministry of Justice and to governor's and district governor's offices under the Ministry of the Interior. As of the date of publication, e-Apostille covers specific document types; it is worth confirming the current scope for your document with the relevant official sources. When it comes to carrying out translation, apostille and certification in the correct order, our apostille translation service plans the process end to end.
Which Method for Which Text? A Comparative Look
If we translate the research findings into practice, the three basic approaches and their appropriate use cases can be summarized as follows:
| Method | Strength | Risk | Suitable For |
|---|---|---|---|
| Raw machine translation | Instant results, zero waiting | Terminology and context errors; no legal validity | Personal comprehension, informal correspondence, getting a first impression |
| Machine translation + expert editing (post-editing) | Speed gain; quality close to human translation according to the research | Errors slip through if the editor is not a domain expert | High-volume corporate content, technical documentation, web content |
| Sworn/human translation | Legal validity, signature and accountability, command of context | Requires more time than the machine | Official documents such as diplomas, court rulings, powers of attorney and medical reports |
Here is the point to watch: the speed gain in the Rutgers study was measured in a scenario where the editing was done by professional translators. A "the machine produced it, I gave it a quick read, we're done" approach is not the post-editing described in the research. In academic processes such as university applications, too, ensuring that transcript and diploma translations conform to institutional formats requires an expert eye; in these files, academic translation experience is decisive.
Common Mistakes and Pitfalls
When carrying the picture drawn by the research into everyday practice, these are the mistakes we encounter most often:
- Submitting machine output to an official institution: However accurate it may be, an unsigned, unstamped translation will not be accepted in official proceedings; institutions look for a sworn translator's signature.
- Confusing sworn translator certification with notarization: These are separate procedures; some institutions accept only the sworn stamp, while others also require notarization. The target institution should be asked before applying.
- Getting the apostille sequence wrong: Whether the destination country wants the apostille on the original document or on the notarized translation varies from country to country; if the sequence is set up incorrectly, the process starts over from the beginning.
- Inconsistent names and dates: Proper names being "translated" or date formats getting mixed up in machine translations is among the most common reasons for application rejections.
- Ignoring confidentiality risks: Uploading identity, health and court documents to publicly available translation tools can mean transferring personal data to third-party servers.
The common denominator of these pitfalls is that the process consists not of isolated steps but of a chain of trust: the right translation, the right certification, the right order. When any link in the chain breaks, what is lost is usually time rather than money; especially in files with an approaching deadline, this loss can have consequences that are difficult to remedy. For such situations, the urgent and fast translation option can step in, but the soundest course is to plan the process early.
What Affects Cost and Turnaround?
There is no single "standard fee" in official document translation; there are variables that determine the amount. Chief among them is the language pair: both expert availability and processing time differ between common and rare languages. The length and nature of the document is the second factor; a one-page civil registry extract and a multi-page court ruling do not require the same effort. Layers of certification are the third factor: whether only a sworn stamp, notarization or the full apostille chain is required changes the total time and the procedural steps. Finally, urgency directly affects planning.
The official fees and tariffs at the notarization and apostille stages are set by the relevant authorities and updated periodically; for the amounts in force as of the date of publication, consult the Union of Turkish Notaries (Türkiye Noterler Birliği) and the relevant governor's office or courthouse sources. If you would like a clear picture for your own document, you can request a free quote by stating the type of your document, its language and the destination country; your file will be reviewed and the required layers of certification reported together with the quote.
Conclusion: The Machine Accelerates, the Human Validates
The two major studies of 2026 converge on the same point: artificial intelligence is a powerful accelerator of the translation process, but it is the human — reworking the text with domain knowledge — who determines quality. Where official documents are concerned, a legal reality is added on top: the chain of signature, stamp, certification and apostille can only be established by people and authorized bodies. Technology is digitizing the links of this chain (as the e-Apostille example shows), but it is not eliminating them.
In the period ahead, hybrid workflows combining machine translation with expert editing are expected to become widespread. For document holders, the practical takeaway remains unchanged: find out what the target institution requires, set up the translation–certification–apostille sequence correctly, and for every document that needs official validity, leave the final word to an expert who assumes responsibility with their signature.
Sources
- Merhaba.info — Artificial intelligence sparks debate on the future of translation
- Ensonhaber — The future of the translation profession is being reshaped in the shadow of AI
- Bursa Hakimiyet (AA) — The future of translation in the shadow of AI debated on International Translation Day
- ACL Anthology — AI Post-Editing in Production: A 71,262-Segment Evaluation (Nunziatini & Speroni, EAMT 2026)
- EAMT 2026 — Proceedings of the 26th Annual Conference, Vol. 2 (Tilburg, 16–18 June 2026)
- Aydınlık — Can artificial intelligence think in Turkish?
- Ministry of Transport and Infrastructure — Official documents reach the world digitally with e-Apostille