Karmactive Team
If you use ChatGPT or Codex within the European Union, eligible text outputs are expected to carry an invisible watermark as OpenAI rolls out its new textGrain system over the coming weeks. OpenAI announced the technology on October 5, 2026, describing a statistical signal that subtly alters word choices during text generation so that detection tools can later assess whether the text came from an OpenAI model. What it cannot do is prove who wrote it, whether any of it is accurate, or whether it was ever edited.
OpenAI published the announcement on October 5, 2026, covering ChatGPT and Codex users in the EU in a phased rollout over the coming weeks, and opening it as an opt-in option for API customers globally on selected models. The company says textGrain works by introducing a statistical bias into which words are selected during generation — a process invisible to any ordinary reader. Initial access to the detection tool is limited to approved researchers and expert organisations. OpenAI has said it plans to make the textGrain technology available as open source in the future.
Eligible outputs generated using ChatGPT or Codex within the European Union are expected to carry an invisible statistical watermark as the rollout proceeds, to satisfy EU AI Act transparency requirements. While this watermark does not alter readability or trigger any browser warning, approved researchers and organisations with access to OpenAI’s provenance system may be able to assess whether a passage originated from an OpenAI model. Businesses publishing AI-assisted content should review their disclosure practices under EU AI Act Article 50, which requires providers of certain AI systems to ensure outputs are marked in machine-readable form.
What textGrain actually does — and where it fails
The EU AI Act’s Article 50 requires providers of certain AI systems to ensure outputs are marked in a machine-readable format detectable by automated means. textGrain meets this by biasing token selection at the moment of generation. Think of it like a very faint pattern woven into fabric: the cloth looks the same to the naked eye, but a scanner tuned to the pattern can find it.
OpenAI states that the watermark is harder to detect on shorter or more constrained text. Editing, paraphrasing, or translating a watermarked passage also weakens the signal. Heavy rewriting can make the pattern undetectable. This means a student who generates an essay and then rearranges sentences may not produce a text that any detector flags with confidence.
This is the detail most coverage has skipped: textGrain establishes that text was probably generated by an OpenAI model at some point — it does not establish who, when, or in what proportion. It does not identify the user. It does not measure how much of a final document was AI-written versus human-edited. It does not determine ownership or authorship. And it carries no assessment of whether the content is factually correct.
For EU API customers, textGrain is opt-in on selected models, giving businesses time to update disclosure workflows before compliance deadlines.
Does this mean every ChatGPT output in Europe is permanently traceable?
Not in any absolute sense. Detection requires specialised tools running statistical analysis, not a browser plugin or a human reader. And as noted, heavily edited or short text produces unreliable results. The watermark is a probabilistic signal, not a cryptographic proof.
The open-source direction matters here. If the textGrain technology becomes publicly available as planned, publishers and platforms could eventually verify AI provenance without going through OpenAI’s own servers. That matters for publishers under the EU AI Act and for employers or academic institutions concerned about undisclosed AI use. It does not create a perfect detection system — it creates a better starting point than currently exists.
For context on how similar transparency rules are developing, see our piece on the EU AI Act and what businesses need to know and our overview of AI content detection tools and their limits.
FAQ
**Can people tell if text was generated by ChatGPT in the EU?** Detection requires specialised analysis tools, not ordinary reading. OpenAI’s textGrain embeds an invisible statistical pattern into word choices. Automated detectors — initially limited to approved researchers and expert organisations — can assess whether a passage likely came from ChatGPT or Codex, but the system produces less reliable results on short passages or heavily edited text.
**What is textGrain and how does it work?** textGrain is OpenAI’s statistical watermarking system. During text generation, it subtly biases which words are selected — creating a pattern that is invisible to readers but assessable by specialised software. It does not change the meaning, fluency, or speed of the generated text.
**Does the EU AI Act require all AI text to be watermarked?** Article 50 of the EU AI Act requires providers of certain AI systems to ensure outputs are marked in a machine-readable format detectable by automated means where technically feasible, subject to scope and exceptions. OpenAI’s textGrain is its announced approach for compliance.
**Can you remove the textGrain watermark by paraphrasing?** Heavy editing, paraphrasing, or translation weakens the watermark’s statistical signal and can make it undetectable. OpenAI acknowledges the system’s limitations on short or significantly rewritten text. The watermark is probabilistic, not permanent.
OpenAI’s plan to make textGrain technology open source and the European Commission’s compliance assessment guidelines for Article 50 are both expected in the coming months. Check back for updates as the regulatory timeline develops.
