One of the most talked-about technology developments of August 2026 is that AI-generated text will now carry a "trace". According to a technical briefing published by Anthropic, new Claude models released to the EU market on or after 2 August 2026 will support machine-readable marking infrastructure; hidden watermarks will be embedded in generated text, while digitally signed provenance metadata will be integrated into supported file formats. At first glance, this may look like news that concerns only the tech world. Yet it has significant implications for anyone seeking to have official documents translated — diplomas, powers of attorney, court rulings or medical reports: text that has passed through an AI system is now becoming technically detectable.
In this article, drawing on up-to-date sources, we look at how invisible watermarking technology works, which legal framework it rests on, and how it may affect official document translation and certified (sworn) translation processes.
What happened? EU transparency rules and the invisible watermark era
Behind this development lies the European Union's artificial intelligence regulation. The new transparency obligations under the EU AI Act came into force on 2 August, with a four-month compliance period granted for existing AI products already on the market; Anthropic announced that newly released models will mark content from day one, while integration work continues for existing models. In other words, this is not a single company's choice; the change is being implemented, alongside other major AI providers, to comply with the EU AI Act. Similar marking systems are expected to become widespread across the industry in the coming period.
The scope is also remarkably broad. It is stated that machine-readable marks will be applied globally, including via the API, chat interface and other products, and because the system operates at the model level, it activates automatically regardless of which interface is used to access it. Put differently, no matter which application a user generates text through, the output may carry this trace.
How does the invisible watermark work, and what does it prove?
The "watermark" here is not a visible mark like a stamp on paper. The method is said not to affect output quality or content, with watermarked and unwatermarked texts being indistinguishable to readers; the watermark carries no information identifying any person, organization or conversation, instead leaving an invisible statistical signature in the text. The technical foundation is not new either: the technique used is based on an adapted form of the SynthID-Text approach developed by Google DeepMind and described in research published in the journal Nature.
The watermark's durability, however, is graduated. The company explained that the watermark will go unnoticed by readers, will probably not be entirely removed by light editing, but will disappear if the text is completely rewritten with every word changed. The most critical detail for official translation is this: the watermark is also applied in assistive operations such as editing and translation; detection results therefore need to be interpreted with caution, and the mark alone does not prove that the content was written entirely by AI. In other words, the watermark does not say "this text was written by AI"; it conveys the information that "this text passed through an AI system at some point".
Why does this development matter for official document translation?
In official translation, the chain of responsibility is clear: the translation is carried out by a sworn translator whose oath is registered with a notary, who vouches for its accuracy with their signature and stamp; if the receiving institution requires it, the translation is additionally notarized. Sworn translator certification and notary-approved translation are two distinct procedures: the first is the translator's own signature liability, the second is the certification of that signature by a notary. In this chain, the answer to the question "who produced the text?" must always be a human being, because legal liability can only be attached to a real person.
Invisible watermarking technology adds a new dimension to this picture. If a document's translation is done from start to finish by a chatbot and presented as a "sworn translation", it may later become technically detectable that the text passed through an AI system. At a time when institutions and academic bodies are increasingly interested in AI detection tools, "official" documents prepared through unsupervised machine translation carry both reputational and validity risks. On the other hand, it should be remembered that the absence of a watermark proves nothing either; the real guarantee is not a technological trace, but the human signature behind the translation and the institutional quality process.
Who is affected? Concrete scenarios
Consider a sample scenario: a student applying for a master's degree abroad has their transcript and statement of purpose translated by a chatbot and adds them to the application file. The transcript translation is returned because it does not meet the institution's "sworn/certified translation" requirement; the statement of purpose is flagged by AI detection tools, raising questions about the candidate's authenticity. Had the same student entrusted the transcript to a sworn translator with academic translation experience, the document would have been accepted the first time.
Similar risks apply to companies as well. A firm entering an international tender that submits its draft contract, certificate of activity or signature circular via unsupervised machine translation may face serious consequences, both due to terminology errors and the liability gap. The situation is even more sensitive for documents to be submitted to court: in the field of legal translation, a single mistranslated term can affect the course of a case. Medical reports, patents and technical specifications are likewise document types that require expert human oversight.
AI translation versus sworn translation: A comparison
It would not be realistic to say AI tools have no place at all in the translation process; professional agencies can also use the technology in a supervised manner at the drafting, terminology and consistency-checking stages. The distinction lies in whether the technology is used as an "unsupervised final product" or as an "assistive tool under human supervision". The table below compares the two approaches specifically for official documents:
| Criterion | Unsupervised AI translation | Sworn translation (human-signed) |
|---|---|---|
| Legal liability | None; no accountable party in case of error | Assumed by the sworn translator with signature and stamp |
| Acceptance by official bodies | Generally not accepted | Accepted per institutional requirements; notarization added if needed |
| Detectability | Can be flagged by invisible watermarks and detection tools | Human-produced; verified by signature and oath record |
| Confidentiality | Document content is uploaded to external servers | Processed through an institutional workflow under confidentiality obligations |
| Terminology and format | High risk of errors in seals, signatures, tables and similar elements | Document layout and official terminology preserved exactly |
As the table shows, the issue is not merely "translation quality"; acceptability, liability and confidentiality must be assessed as a whole. At Ziya Translation, we treat it as standard practice that the final text of every official document translation is prepared and checked by a human translator; you can explore the language pairs we work with on our translation languages page.
The privacy dimension: Before uploading your document to a chatbot
Another issue overshadowed by the watermark debate is data privacy. The use of "shadow AI" is increasingly being discussed in Turkey as well: In examples such as a public official uploading a tender file to a chatbot on their personal account, or a physician turning patient notes into a report, users do not intend to harm their institutions — they are merely trying to speed up their work — yet the result can be the uncontrolled transfer of sensitive data to third-party systems.
Official documents are, by definition, dense in personal data: identity numbers, health information, financial records, criminal record content... Uploading such documents to publicly available AI tools may also pose risks with respect to obligations under Turkey's Personal Data Protection Law (KVKK). One of the unseen benefits of working with a professional translation agency is that your document is processed under a confidentiality commitment, within a workflow with restricted access. This principle should not be abandoned even for rush jobs; indeed, with urgent translation requests, speed, confidentiality and accuracy can be managed together.
The right process step by step: Translation, notary and apostille sequence
If you plan to use your document abroad, or a foreign document in Turkey, the path to follow is broadly established. First, the requirements of the receiving institution are clarified: is a sworn translator's signature alone sufficient, is notarization required, or is an apostille certificate needed? The answer to this question determines the entire process. Let us stress in particular that the apostille certificate is not issued by the translation agency: apostilles are issued by provincial and district governorships for administrative documents, and by the competent judicial authorities for judicial documents; the agency's role is to organize the translation and certification steps in the correct order.
- The level of certification required by the receiving institution (sworn signature / notarization / apostille) is clarified.
- The document is translated by a translator with a registered oath in the relevant language pair; names, dates and numbers are checked one by one.
- If required, the translation is notarized.
- If an apostille is required, the certificate is obtained from the competent authority depending on the document type (provincial/district governorship or judicial authority); for some document types, this step can also be carried out electronically.
- The document is delivered in the format the receiving institution requires (wet-signed, e-signed or physical).
Digitalization is transforming the apostille leg of this chain as well. According to the official announcement, with the e-Apostille system, citizens can create applications via e-Government (e-Devlet) according to document type; requests are forwarded electronically to judicial units under the Ministry of Justice and to provincial and district governorships under the Ministry of the Interior, and the resulting documents are recognized as valid in more than 130 countries that are party to the Hague Convention. As of the publication date of this article, the e-Apostille does not cover every document type, so it is advisable to confirm the current scope for your own document with the relevant authority. To have all of these steps planned from a single point, you can get support through our apostille attested translation service.
Factors affecting cost and next steps
The total cost of an official document translation is not made up of a single item and varies from document to document. The main factors determining the amount are as follows:
- Language pair: Translator availability differs between common and rare languages.
- Length and nature of the document: Page/character volume and the need for legal, medical or technical expertise affect the process.
- Level of certification: Whether only a sworn signature or additional notarization is required changes the total; notary fees are collected by the notary office according to the official tariff.
- Urgency: Same-day or weekend delivery requests affect planning.
- Third-party official fees: For example, if you are entering the equivalence process for a foreign diploma, by decision of the Executive Board of the Council of Higher Education (YÖK), a fee of TRY 1,250 has been set for 2026 for diploma/graduation equivalence certificate applications and re-examination requests (as of the publication date of this article; please confirm the current amount with YÖK). Such official charges are items independent of the translation service.
If you would like a clear picture for your own document, simply send it via our free quote form; we will clarify the certification requirements together and map out a route tailored to you. Ultimately, the invisible watermark era makes visible, through technology, a principle that already applies in official translation: behind an official document there must be a human being who assumes responsibility with their signature. AI tools can assist; but the translation carrying your diploma, court ruling or medical report is secured by the sworn translator's signature and, where required, notarization.
Sources
- BirGün — Claude now adds invisible watermarks to texts
- Mynet Teknoloji — How will Claude's 'invisible' watermark work? Anthropic explains the details
- Cumhuriyet — Anthropic makes a new statement on the watermark system introduced for Claude
- 12punto — Anthropic to add invisible watermarks to text and images generated with Claude
- Teknoblog — Why are invisible watermarks being added to Claude texts?
- Dünya Gazetesi — Measures should be taken against shadow AI use
- Ministry of Transport and Infrastructure — With e-Apostille, official documents reach the world digitally
- YÖK Recognition and Equivalence Services — Equivalence Application