Claude vs ChatGPT vs Gemini: 4 AI Watermark Layers and the Free Tools That Work in 2026

Short answer: Claude now adds a model-level watermark to supported new text and C2PA provenance to supported files. ChatGPT currently marks supported images and audio, not public text output. Gemini uses disclosed SynthID systems across text and media. No single free tool removes all four mark layers, and “no mark found” never proves human authorship.

Checked against vendor documentation on 12 August 2026.

The rush to “remove the Claude watermark” started before Anthropic published the detector or the technical method. That gap matters. A cleaner can prove that it deleted an invisible Unicode character or a C2PA block. It cannot prove that it defeated an undisclosed model-level text signal.

The open-source tool in the viral post, watermarks-remover, is more honest than most of the posts sharing it. It labels Unicode and common metadata cleaning as verifiable, statistical text rewriting as best effort, and pixel-level SynthID removal as out of scope. I cloned the current repository and ran its suite: all 33 tests passed. That makes it useful, but not a universal eraser.

Here is the vendor picture first.

What changed with Claude's AI watermark in 2026?

Anthropic's new guidance says Claude models launched on or after 2 August 2026 support machine-readable marking from launch. The system applies worldwide, not only in the EU, and covers supported output from Claude, Claude Code, Claude Cowork, Claude Tag, the API and supported cloud routes.

There are two parts. Generated text carries an imperceptible watermark embedded at model level. Supported files, including SVG, PNG and JPG, receive signed provenance data based on the C2PA standard. Anthropic says the text mark survives copy and paste and may persist through some editing.

What Anthropic does not say is just as important. The company has not disclosed whether the text method uses token sampling, Unicode, lexical choices or a hybrid. It has not released the detector, a public test endpoint or the keys. The official page says technical detection details are forthcoming.

So a tool that removes zero-width characters can truthfully say “no suspicious Unicode remains”. It cannot truthfully say “the Claude watermark is gone”. Those are different tests.

The timing comes from the EU AI Act Article 50 transparency duties, which began to apply on 2 August 2026. The related Code of Practice is voluntary, but the legal transparency duties are not. This is why several model providers are expanding machine-readable provenance now.

What are the 4 AI watermark layers you need to test in 2026?

Treat AI provenance as four separate layers. The right removal method depends on where the signal lives.

1. Is it only a visible label or corner badge?

A visible badge is ordinary pixels or text placed over the asset. Cropping, inpainting or editing can remove it. This is the easiest layer and the least informative one.

2. Is it C2PA, EXIF, XMP or document metadata?

C2PA Content Credentials are cryptographically signed provenance records attached to a file. EXIF, XMP and document properties are older forms of metadata. They can record the creating tool, edit history, timestamps or software names.

Metadata is useful because it can carry rich context, but it is fragile. Re-saving, format conversion and many social platforms can remove it. Both Anthropic and OpenAI warn that stripped metadata does not prove the content was never AI-generated.

3. Is the signal embedded in image, audio or video content?

An embedded media watermark changes the content itself. Google's SynthID distributes a signal through pixels, audio waveforms or video frames. OpenAI now says supported images carry both C2PA and SynthID, while supported audio carries SynthID.

This is the layer that metadata-only tools miss. ExifTool can remove XMP. A format conversion can drop a C2PA manifest. Neither action proves that an embedded pixel or audio signal has gone.

Google says its media marks are designed to survive common changes such as cropping, filtering, compression, noise and speed changes. Heavy transformations can weaken detection, but there is no free universal cleaner that reliably removes current SynthID media marks while preserving quality.

4. Is the mark statistical and woven into token choices?

Generative text watermarking changes how the model chooses the next token. The signal is spread across the wording rather than inserted as one removable character.

Google has published this method. The SynthID-Text paper in Nature says the system changes only the sampling process, and its detector does not need to run the underlying language model. Google's production experiment covered nearly 20 million Gemini responses and reported no change in its quality checks or human ratings.

Heavy paraphrasing, translation or regeneration can disrupt a statistical signal because it replaces many token choices. It also replaces the original voice and facts. Light edits, heading changes and paragraph shuffling are much less likely to help.

Is invisible Unicode the Claude watermark?

There is no public evidence for that claim as of 12 August 2026.

Invisible Unicode is a real edit-based marking technique. Zero-width spaces, bidirectional controls, tag characters and unusual space code points can survive copy and paste. They are easy to inspect and remove deterministically.

But Anthropic only says its new mark is embedded in text at model level. It does not identify Unicode as the carrier. A Unicode cleaner remains good document hygiene, especially when copied text behaves oddly, but a clean Unicode scan is not proof that Claude's undisclosed signal is gone.

How do Claude, ChatGPT and Gemini watermarks compare?

What does Claude mark?

Claude has the broadest new text commitment in the three vendor documents reviewed here. Supported new models mark all generated text across product surfaces and cloud partners where supported. Claude also signs supported files with C2PA.

The trade-off is opacity. Anthropic has announced coverage and limits but not the technical text scheme or public detector. Any removal claim today has to stop at the layer it can inspect.

Does ChatGPT watermark text in 2026?

Not in OpenAI's current public provenance guidance. The company lists C2PA plus SynthID for supported images and SynthID for supported audio. It says expanding signals to text is a goal.

OpenAI did publish research on a text watermark in 2024. It said the method was accurate and survived local edits, yet was easily bypassed by global changes such as translation or rewriting with another model. It also raised false-positive and fairness concerns, especially for non-native English speakers. That research should not be confused with a deployed ChatGPT text watermark.

How does Gemini's SynthID work?

Gemini has the most technically disclosed cross-media system. SynthID-Text biases token selection. SynthID for images, audio and video embeds signals in the media itself. The detector and robustness tests differ by medium.

Google's SynthID overview and Nature paper give researchers far more detail than Anthropic currently gives for Claude. That still does not make removal simple. An open reference implementation explains the class of method, but production keys and configurations remain controlled.

Which free AI watermark removal tools actually work?

Is watermarks-remover the best all-in-one free option?

For transparent, local inspection across text and common document formats, it is the best open-source starting point I found. The MIT-licensed repository includes:

• deterministic inspection and cleaning for suspicious Unicode;
• metadata cleaning for PNG, JPEG, SVG, PDF, DOCX, ODT, HTML and Markdown;
• optional C2PA inspection through c2patool;
• a rewrite hook for statistical text marks;
• an optional reverse-SynthID image scorer for research.

Its limitation statement is the reason to trust it. Pixel-level removal is out of scope. Statistical rewriting is best effort. The optional reverse-SynthID integration scores images but does not remove the mark. No vendor detector means no certified clean result.

After installation, its safest sequence is inspection first, then a lossless clean:

python3 skills/remove-ai-marks/scripts/inspect_file.py draft.md
python3 skills/remove-ai-marks/scripts/clean_file.py draft.md -o draft.cleaned.md
python3 skills/remove-ai-marks/scripts/inspect_file.py draft.cleaned.md

Run it only on material you own or have permission to process, and keep the original.

When should you use c2patool and Content Credentials Verify?

c2patool is the strongest free local inspector for C2PA manifests. It reads a summary or low-level report and can add manifests. Use it to establish whether a signed claim exists before and after a file operation.

c2patool image.jpg

For a no-install browser check, use Content Credentials Verify. Both are verification tools. Neither claims that a clean C2PA result means there is no embedded media watermark.

When should you use ExifTool, ImageMagick or FFmpeg?

ExifTool is the best free general metadata tool. It can inspect and remove EXIF, XMP and many other tag groups across a wide set of file types. Work on a copy, then compare the result:

exiftool -all= -o cleaned.jpg original.jpg

ImageMagick's strip option and FFmpeg's metadata controls can do the same job during a media conversion. These tools clean file containers. They do not reliably remove SynthID embedded in pixels, frames or sound.

Can a free local model remove a statistical text watermark?

A local model through Ollama can heavily rewrite text without sending it to another hosted vendor. Using a model from a different provider also reduces the risk of applying the same vendor's mark again.

This is not lossless. A strong rewrite changes wording, rhythm, citations and sometimes facts. Back-translation has the same problem. If preserving voice matters, keep the original text and accept that the mark may remain. If provenance policy matters, disclose the AI role instead of treating absence of detection as proof of human authorship.

What is the best 5-step removal workflow in 2026?

1. Confirm authority and the disclosure rule. Only clean content you own or may process. Check school, employer, publisher and legal requirements before removing provenance.
2. Identify the layer. Look separately for a visible badge, file metadata, an embedded media mark and a text-level signal. Do not call them one watermark.
3. Preserve the original and baseline it. Save a copy. Record Unicode counts, metadata, C2PA status and any provider verification result before editing.
4. Use the lowest-impact technique. Scrub Unicode for Unicode, strip metadata for metadata, and edit a visible badge as pixels. Only use a full rewrite when you accept the loss of voice and accuracy.
5. Recheck every channel and report residual risk. A clean metadata report says nothing about SynthID. A clean Unicode scan says nothing about an undisclosed statistical mark. Write down what was verified and what remains unknown.

This layered workflow is also the right way to audit generated media before it reaches a public site. Schmitdy's free tools, security approach and AI-search audit use the same rule: test the exact claim rather than relying on a vague detector score.

Can you ever prove an AI watermark is gone?

Only for a defined signal and a defined detector.

You can prove that a file contains no C2PA manifest that c2patool can read. You can prove that a cleaner removed the Unicode categories it targets. You can compare EXIF and XMP before and after. You cannot turn those results into a universal claim that the asset is unmarked.

A provider detector can also return no signal. Anthropic, OpenAI and Google all warn that absence is inconclusive. The content may come from an older model, an unsupported route, a short passage or a heavily edited file. The detector may simply have too little signal left.

The honest output is therefore a small evidence record: signal checked, tool and version, result, transformation applied, and residual uncertainty.

Should businesses remove AI provenance by default?

No. Provenance is useful when a company needs an audit trail, an origin record or a clear edit history. Stripping it by default can destroy evidence your legal, security or editorial team would prefer to keep.

Do not use watermark removal to claim that generated work is human-made. A watermark only says a supported tool may have processed content. Its removal changes the evidence, not the history.

Which conclusion survives the hype?

Claude's new watermark is real. The viral Unicode explanation is unproven. ChatGPT does not currently publish a deployed text watermark. Gemini's SynthID-Text is the best documented text scheme of the three. Metadata stripping is straightforward, embedded media marks are harder, and statistical text marks require destructive rewriting.

The best free stack is inspection-led: watermarks-remover for Unicode and common files, c2patool or Content Credentials Verify for C2PA, ExifTool for metadata, and the vendor's own verifier where one exists. Use a rewrite only when you accept the quality cost and the lack of proof.

That is less exciting than “one command removes every AI watermark”. It is also the answer the evidence supports.