Codex

Light

Artificial Intelligence

OpenAI has begun labeling text generated by its AI systems ChatGPT and Codex for users in the EU region, the firm stated on Monday, due to the requirement of transparency under the EU Artificial Intelligence Act implemented on August 2.
Those rules require AI companies to label machine-generated content in a form other systems can read, with Article 50 breaches carrying penalties up to €15 million or 3% of global annual turnover. The rollout reaches eligible accounts on every plan over the next few weeks, but only inside the bloc.
It is possible for developers accessing the API from any location to activate the feature for certain models starting from today, although the feature will remain inactive until a customer activates it, according to TechCrunch.
<blockquote class="twitter-tweet">
 We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology.
 Our tools already help verify whether an image or audio file was created with our…
 — OpenAI (@OpenAI) <a href="https://x.com/OpenAI/status/2107164650249101695?ref_src=twsrc%5Etfw" rel="noopener noreferrer nofollow">October 5, 2026</a>
</blockquote>
<h3 id="textgrain-hides-the-watermark-in-the-model-s-word-choices">textGrain hides the watermark in the model’s word choices</h3>
OpenAI’s textGrain system works by slightly steering the words ChatGPT chooses when several alternatives would fit equally well. Over a long enough text, those little decisions build up a statistical pattern that is invisible to the naked eye but can be picked up by special software.
Because the watermark is built into the wording itself rather than stored in file metadata, it can remain detectable even after the text is copied and pasted somewhere else.
OpenAI developed the system with researchers from the University of Pennsylvania and Yale. Their t<a href="https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf" rel="noopener noreferrer nofollow">echnical paper</a> explains that textGrain uses a secret key to rank possible next words and subtly favor certain choices. When enough of those choices appear in a passage, software with access to the same key can identify the pattern and determine that the text was likely generated or processed by an OpenAI model.
OpenAI says the watermark does not contain any information that identifies the user who generated the text. The company also says its testing found no meaningful drop in GPT-6 Astra’s performance when textGrain was enabled.
<h3 id="light-editing-can-sharply-weaken-watermark-detection">Light editing can sharply weaken watermark detection</h3>
OpenAI also published the limits of the system. In its tests, replacing just 10% of the words with synonyms reduced detection from about 92% to 66%. The watermark is also harder to detect in short passages, mathematical answers and translated text.
Due to these limitations, OpenAI will be restricting the use of its detection tools to authorized researchers and expert organizations for the time being. It also mentions that the inability to detect watermarks cannot confirm that the text is human-written. The text is perhaps just too short, extensively edited or even produced by a different company’s model.
<blockquote>
 A watermark can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.
 – OpenAI
</blockquote>
<h3 id="anthropic-chose-global-watermarking-while-openai-starts-with-europe">Anthropic chose global watermarking while OpenAI starts with Europe</h3>
In August, Anthropic started to watermark Claude’s outputs globally. Initially, it seemed like a more wide-ranging policy compared to OpenAI, which was doing it only in Europe, but the reason for that was technical; Anthropic did not have a system to restrict regional watermarking yet.
Both companies had already signed the EU’s Code of Practice on Transparency of AI-Generated Content in July, alongside roughly 190 other organizations, including Google, Meta, Microsoft and Mistral. Consequently, their efforts at watermarking are in response to the same EU transparency requirements that became effective from August 2.
Anthropic’s rollout still drew criticism from some users, who argued that they were providing the ideas, instructions and judgment while Claude was mainly helping with the writing.
According to the Wall Street Journal, OpenAI had indeed created a water-marking system for texts years before but opted not to make the system public because of fears that the users may shift to other AI systems that do not watermark their outputs.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. <a href="https://www.cryptopolitan.com/newsletters/?utm_source=cp&amp;utm_medium=web&amp;utm_campaign=inlineAds" rel="noopener noreferrer nofollow">It's free</a>.

OpenAI adds invisible text watermarks to ChatGPT and Codex in Europe

Table of Contents

Cryptopolitan_News

Other articles published on Oct 6, 2026

Join our free newsletter for daily crypto updates!