Blog
AI watermark blog
People search “ChatGPT watermark” and land on three different mechanisms. One is a statistical pattern in word choice. One is invisible Unicode in a paste. One is a signed manifest inside an image file. These guides separate them, then point at the check that matches the thing you actually have.
The working checker is the ChatGPT watermark detector on the homepage. It reports listed carriers: invisible Unicode in text, and C2PA, EXIF, or AI metadata in files. It does not guess writing style, and it does not hold OpenAI’s text-watermark key. A clean result means those listed carriers were absent. It does not mean a person wrote the draft.
Start with the question you have
- How does the ChatGPT text watermark work? OpenAI’s textGrain method, the key, and what a detector can and cannot say.
- Is ChatGPT text watermarked in 2026? EU ChatGPT and Codex, the API opt-in, and everything that is still off.
- Can Turnitin see that watermark? The AI writing report is a classifier. It does not hold OpenAI’s key.
- Did a freelancer send watermarked copy? What an editor can check on the draft, the images, and the disclosure.
- Why does pasted ChatGPT text contain characters I cannot see? Zero-width spaces and related format characters, and how to count them.
- Does this PNG or JPEG carry a C2PA watermark? Content Credentials, and how that differs from a pixel watermark.
- I have text or a file in front of me and I want a check. Paste or upload on the tool. The report lists matches. It does not score “how AI this sounds.”
How the guides fit together
The pillar on how textGrain works is the mechanism. The pillar on 2026 coverage is where that mechanism is switched on. Read them in that order if you are updating an old internal FAQ that still says ChatGPT does not watermark text, or that still treats a zero-width space as the watermark.
Schools and publishers then split. Instructors comparing an Originality percentage with a rumor about a hidden mark should read Turnitin and ChatGPT watermarks and watermark versus an AI detector. Editors hiring writers should read how to check a blog writer and then which free checker does which job. The removal page is there so “clean” does not get applied to the wrong layer: can you remove a text watermark. A miss on any of those tests is explained in what a missing watermark means.
Other labs are not a footnote. Claude’s text mark is a SynthID-Text variant. ChatGPT’s text mark is textGrain. Gemini’s public story is SynthID-Text again, with a different key. The comparison is Claude, Gemini, and ChatGPT. The legal duty those rollouts are answering is Article 50 and the transparency code. None of that changes what the homepage tool can see. It changes which question you bring to the tool.
All articles
How the OpenAI ChatGPT text watermark actually works
textGrain, in plain language: token choice, the key, and who can detect it.
CoverageDoes ChatGPT watermark text in 2026?
EU ChatGPT and Codex, the API opt-in, and the paths that stay unmarked.
SchoolsCan Turnitin detect ChatGPT watermarks in 2026?
Why an AI writing percentage and a textGrain hit are different results.
Two testsChatGPT watermark vs an AI writing detector
Embedded signals versus classifiers, and why both can be true.
EditorsHow to check if a blog writer used watermarked AI text
What to collect, what a character scan proves, and what a contract can require.
ComparisonFree AI watermark checkers compared
Character scans, C2PA readers, vendor verify pages, and style detectors.
RemovalCan you remove a ChatGPT text watermark?
Characters, manifests, and rewrites are three different operations.
NegativesWhat a missing ChatGPT watermark actually means
A quiet detector is not a finding that a person wrote the text.
Other labsClaude, Gemini, and ChatGPT watermarks compared
SynthID-Text, textGrain, and C2PA, and why one key does not read the others.
Article 50What the EU AI Act requires for marking AI text
Machine-readable marks, the code of practice, and what a watermark does not settle.
Text you can inspectChatGPT hidden characters and invisible Unicode
Zero-width spaces, BOM, and bidi marks that ride along when you paste.
Images and filesWhat a C2PA watermark is, and how to check a file
Content Credentials in PNG and JPEG, and how they differ from pixels.
Background these pages rely on: OpenAI’s note on EU text provenance, OpenAI’s provenance help, Anthropic on Claude’s text watermark, and the C2PA standard.