Text watermarks can live in ordinary words
OpenAI and Anthropic describe statistical watermarks in supported generated text. These signals come from patterns in the model's word choices. They do not depend on zero-width spaces or other hidden Unicode characters.
That distinction matters when choosing a cleaner. Free AI Wiper can find and remove supported hidden characters in copied text. It does not detect or remove OpenAI textGrain or Claude's statistical text watermark.
The provider details below were checked against their official explanations on 9 October 2026. Coverage depends on the model, product and rollout date.
Does ChatGPT add a text watermark?
On 5 October 2026, OpenAI announced textGrain, its statistical text watermark. It described a rollout over the following weeks for eligible ChatGPT and Codex text outputs in the European Union, across all plans. API customers worldwide can opt in for supported models; API watermarking is off by default. This announcement does not establish whether a particular response or account has received the rollout. Read OpenAI's announcement.
OpenAI says textGrain changes how the model chooses between possible words or word pieces. It adds no hidden characters, invisible spaces or watermark-only tokens. Text-detector access is restricted to approved organizations; enabling watermark generation does not grant detector access. See OpenAI's provenance FAQ.
How does Claude watermark text?
Anthropic describes Claude's watermark as a version of SynthID-Text, the approach published by Google DeepMind. It uses a secret key to create a statistical pattern across word choices, without inserting hidden characters. Anthropic says it applies watermarking globally at launch and is adding coverage for older models over time. Its detection API is in private preview for eligible organizations. Read Anthropic's explanation.
Public SynthID research code provides tools for experiments with known keys and configurations. It does not provide Claude's production key or establish compatibility with Claude's deployed watermark. See Google's reference implementation.
Does copying into plain text remove the watermark?
Copying and pasting preserves the wording. A watermark carried by those word choices can therefore remain in a plain-text file. Removing file metadata, stripping formatting or deleting zero-width characters does not, by itself, test or remove this statistical signal. OpenAI explains how its signal survives reuse.
Can rewriting remove a statistical text watermark?
Substantial rewriting can weaken detection. OpenAI reports reduced detection after word substitutions, and Anthropic says light editing may leave its watermark detectable. These statements do not establish a universal removal method or a guarantee for a particular passage. Rewriting also changes your content, and rewriting with a watermarking model can produce another watermarked output. OpenAI's evaluation, Anthropic's editing limitations.
A low score from a general AI-writing classifier is not verification that a provider's watermark is gone. The matching watermark detector and its limitations matter.
What can a hidden-character cleaner help with?
A hidden Unicode character is a real code point in the text. A zero-width space (U+200B), for example, can be located at a particular position and removed without rewriting the surrounding words. Free AI Wiper's text cleaner reports supported characters by code point, line and column. Processing happens locally in your browser.
You may find hidden characters in human-written text copied from another application. Generated text may contain none. Neither result establishes who wrote the passage.
To investigate copy-paste artifacts, follow the hidden-character guide. For questions about authorship claims, read what AI detectors can prove. A report with no supported hidden-character findings does not rule out a statistical watermark.