Comparison

Free AI watermark checkers compared

“Free AI watermark checker” is four products wearing one name. Some count hidden characters. Some read a C2PA manifest. One vendor page reads SynthID or C2PA on images and audio. Most of the pages that rank are style detectors that never look at a watermark at all.

This comparison is about what each class can honestly report. It is not a horse race of accuracy percentages, because the tools are not measuring the same event. A hit on a manifest and a hit on a prose classifier can both be true. A miss on either one is not a human-written certificate.

The four classes

Sort a tool by the evidence it uses before you sort it by the adjective in the title.

• Character scans look for Unicode that does not show on screen: zero-width spaces, a BOM, bidi marks. The result is a count. It is reproducible.
• Container readers look inside PNG, JPEG, PDF, or DOCX for C2PA, XMP, EXIF, or generator properties. The result is “this segment is present” or not.
• Keyed watermark detectors rebuild a secret pattern from the pixels, the audio, or the token choices. Without the key, a third party is guessing.
• Classifiers score how the prose or the picture statistics resemble known generators. No key is required, and no key is read.

OpenAI’s textGrain sits in the third class. The public description is in how the ChatGPT text watermark works. Invisible paste characters sit in the first class, and they are not textGrain. That distinction is the whole of the hidden-character guide.

What each free surface checks

Coverage below is what the operator documents, not a lab retest of every logo. “Free” includes freemium tools whose unpaid tier is the one most people actually run.

What this site checks

The ChatGPT watermark detector on the homepage is a listed-carrier tool. On text, it counts invisible Unicode and can remove those codepoints without changing visible words. On PNG and JPEG, it looks for C2PA segments and AI-looking EXIF or XMP. Markdown, HTML, PDF, and Word go through the matching tabs when the mark lives in metadata or markup rather than in the sentences.

It does not score “how AI this sounds.” It does not hold OpenAI’s textGrain key, Google’s production SynthID key, or Anthropic’s key. A clean result means the carriers it lists were absent. People still download an image and run it back through the same checker, which is a fair test of the metadata layer and a useless test of a pixel watermark. The pixel layer is out of scope on purpose: stripping a manifest is not the same operation as repainting a picture.

Use it when the question is hygiene or disclosure of a file you own. Do not use it when the question is whether Turnitin will highlight a paragraph. Those answers do not travel.

Vendor verify pages

OpenAI’s verify page is the right free tool for a specific question: does this supported image or audio file contain an OpenAI provenance signal the page knows, such as SynthID or a trusted C2PA manifest? Upload one original file. Cropping, recompression, and screenshots are how that signal gets harder to read. A miss has the usual list of causes: the file predates the signal, the export path was not covered, or the mark was degraded. None of those causes is “a person painted it.”

There is no equivalent public paste box for textGrain. OpenAI says organizations with a qualifying research use can apply. That is a review, not a free tier. API customers who opt in to watermarking still do not receive the detector with the switch.

Anthropic describes Claude’s text watermark as a version of SynthID-Text and says file outputs such as PNG and JPEG get C2PA. Their help center is the source for which products are covered. A third-party page that claims a universal Claude score is running a classifier unless it is Anthropic’s own detector. The side-by-side of the three labs is in Claude, Gemini, and ChatGPT watermarks.

Content Credentials

contentcredentials.org/verify reads a C2PA manifest if the file still has one. It is the right second opinion when this site reports a manifest, or when you want the signed claim displayed by the standard’s own tooling. It will not see a manifest that an export already dropped, and it will not see textGrain, because text has no C2PA box.

Camera makers and newsrooms use the same standard for photographs that were not model-generated. A credential is an origin record, not a synonym for “AI.” Reading the claim matters more than seeing the badge. The longer account of hard-bound versus soft-bound signals is the C2PA watermark guide.

Style detectors sold as watermark checkers

GPTZero, ZeroGPT, Copyleaks, Originality, Writer, and Sapling are useful only if you wanted a classifier. Several of them have a free or metered box. Their pages often rank for “ChatGPT watermark detector” because that is what people type, not because the model inspects a watermark. OpenAI describes this category directly: the tool looks at patterns such as word choice after generation, instead of detecting an embedded signal.

Disagreement inside the category is expected. Thresholds differ, languages differ, and a paraphraser moves the score without touching a key the tool never had. A publisher workflow that treats any one of them as the watermark check will both accuse human drafts and clear marked drafts that were rewritten. How to place the score next to a real intake process is in how to check a blog writer. How the two measurements diverge in principle is in watermark versus AI detector.

Developer tooling is a special case. Google’s open-source SynthID-Text detector can score text when you hold the matching configuration, which is how a lab watermarks its own model. It does not become a ChatGPT detector by installing Transformers. Running it on a freelancer’s essay and reading the number as OpenAI’s opinion is a category error.

How to choose for a real file

Match the surface to the object in front of you.

1. A paste of prose. Run a character scan. If you also want a style estimate, run a classifier and label it as a classifier. Do not expect either one to report textGrain.
2. A PNG or JPEG from a generator. Run a container check here, then Content Credentials Verify if you want the signed claim. For an OpenAI export, openai.com/verify is the surface that also looks for SynthID.
3. A PDF or DOCX. Check document metadata. A credential in the file can outlive a rewrite of the paragraphs, and a rewrite can outlive a credential if someone saved a clean copy.
4. A claim that “the watermark was removed.” Ask which layer. Character cleaning, manifest stripping, and paraphrasing are different, and only the first two are verifiable by re-running the same scan. See can you remove a ChatGPT text watermark.

If the result you needed was “a person wrote this,” none of the rows provide it. A missing signal has too many ordinary causes, collected in what a missing watermark means.

How to make a check repeatable

Two people can disagree about a “free check” because they did not check the same bytes. Fix the object first. For prose, store the text file you pasted, not a screenshot and not a recap. For an image, store the original export and the exact file you uploaded. A re-saved JPEG from a preview pane is a different specimen.

Write down the surface, not the category. “We used a watermark checker” is how a character scan and a classifier get merged in the next meeting. Name the site, the date, and the result in the tool’s words: codepoint counts, manifest present, or a vendor score with the vendor’s label. If you clean the file, store the cleaned copy beside the original and say which carriers the cleaner claimed to remove. Re-run the same surface. A second tool that still flags the prose has not failed. It was measuring something else, which you would know if the note named the class.

Keyed checks have an extra line: who held the key. A SynthID-Text notebook you ran locally is a check of your configuration. An OpenAI verify result is a check of OpenAI’s image and audio signals. Neither note should be copied forward as “ChatGPT text: negative.” The missing-watermark page is the glossary for that mistake.

Free tiers and what they quietly drop

Metered classifiers truncate. A free box may score the first few hundred words of a 3,000-word post and let you talk as if it scored the post. Paste the sections that matter, or pay for the whole file, and say which you did. A score on the introduction is not a score on the methods.

File tools drop in a different way. Some browsers resize an upload before it leaves the machine. Some checkers accept a PDF and rasterize the first page, which destroys both the text tokens and the embedded metadata, then run a classifier on the raster. You think you tested the document. You tested a picture of the first page. Prefer a tool that names the carrier. This site’s document path looks at metadata and markup. It will not invent a style score to fill a quota.

Privacy belongs in the comparison because “free” is often the product. OpenAI says files sent to its verify page are processed for the signal and not used for training, with the legal-retention caveat in that help article. Read the same clause on any classifier before you paste a client draft or a student’s essay. A school that already has Turnitin has a contract. A random free box does not become appropriate because the search result said watermark.

Price is a bad proxy for class. A paid classifier is still a classifier. A free manifest reader is still a manifest reader. Choose the class that matches the question, then choose a tool inside that class whose quota and retention terms you can live with. The editorial version of that choice is how to check a blog writer.

FAQ

← All AI watermark guides