OpenAI’s textGrain announcement and rollout scope
OpenAI’s October 5, 2026 announcement describes textGrain, a statistical signal in word choices. It announces opt-in watermarking for select API models, with the API setting off by default, and an upcoming rollout for eligible ChatGPT and Codex text in the EU. Initial detector access is restricted to approved researchers and expert organizations.
That phased scope matters. It is not evidence that every answer everywhere is marked, and an earlier article saying text watermarking is only research is no longer the complete picture. Check the current product documentation when provenance matters. A Unicode scan does not establish whether a particular passage falls within the rollout.
Are hidden ChatGPT characters text watermarks?
A scan can tell you that a character exists in the text you pasted. It cannot establish when it entered the workflow. The character might have come from the source answer, a rendering layer, an intermediary editor, or a later manual change.
If you want to investigate, compare a controlled sample at successive stages. Preserve the original and record how you copied it. A reproducible observation about one workflow is useful; generalizing it to every ChatGPT answer or every AI detector would require much stronger evidence.
ChatGPT HTML attributes and clipboard formatting
A browser may copy a styled HTML representation alongside plain text. Attributes in that HTML can describe elements or application behavior. A plain-text paste may not carry those attributes at all. Always distinguish literal source markup from the characters in the passage being reviewed.
A claim that a data attribute proves the origin of all words in a document is too broad. The document may include quoted material, imported text, human edits, or multiple sources. Interface context can be relevant evidence in some circumstances, but it is not a complete authorship history.
What a ChatGPT text cleaner can remove
If an unexpected character causes a lookup or formatting issue, inspect it and make a narrow correction. If Markdown is the problem, use formatting conversion. Check the output in the application where it will be used and preserve meaningful citations and structure.
Do not describe that operation as a verified removal of a provider watermark. The tool has only measured and changed its supported character patterns. It has not analyzed a private statistical scheme or authenticated the document’s source.
How to evaluate a ChatGPT watermark claim
Which product and version does the claim cover? What primary source supports it? What exactly does the proposed checker inspect? Has the method been validated for that specific signal? What does a negative result leave unresolved?
Those questions are more useful than looking for a suspicious punctuation mark. If the real task is assessing a document, verify its facts and sources and review its drafting evidence through a fair process. If it is simply moving text between apps, solve that concrete problem without adding an unsupported authorship conclusion.
Sources & further reading
Primary references for the technical points in this guide.