Does removing supported image metadata change image pixels?

No. For supported removals, the tool copies PNG image data and the JPEG compressed scan byte-for-byte. It removes only identified metadata containers.

Free watermark checker

Check and remove hidden watermarks

Q: Are hidden Unicode characters proof that text was written by AI?

No. Invisible Unicode characters can be used for formatting, emoji, right-to-left text, or intentional watermarking. Their presence does not identify the author or software that added them.

Check text for invisible Unicode or inspect supported files for signed Content Credentials. Review each finding before creating a cleaned copy where safe removal is available.

Check your content

Your content stays in this browser

Choose input

Run a watermark check

Text to check

Checks exact invisible Unicode, decoded hidden payloads, and mixed-script lookalikes. It does not infer AI authorship.

Invisible Unicode Available · exact local check
Mixed-script lookalikes Available · review-only check
SynthID Text detector Local reference score · not a Claude verdict

Lookalikes are flagged only when common Greek or Cyrillic characters appear inside a Latin-script token; they are never changed automatically. Our SynthID Text reference engine matches Google DeepMind's published token-level implementation and runs locally with a pinned GPT-2 tokenizer. The resulting test-profile statistics are not a Claude verdict. Anthropic's official detector is in private preview; Claude attribution here starts only when a compatible tokenizer, keys, algorithm profile, thresholds, and test vectors become publicly reproducible. Rewriting can change tone and precision without proving that a statistical watermark is gone. Read Anthropic's announcement.

Check complete

Text checked

SynthID Text reference detector

Reference engine verified · Claude key unavailable

How our SynthID Text detector works

Supported checks

What this watermark checker can find

A clean result means no supported signal was found. It does not prove that content has no watermark or that it was created by a person.

Hidden Unicode

Finds invisible Unicode, decodes tag and variation-selector payloads, and reviews mixed-script lookalikes.

C2PA and AI metadata

Checks supported files for embedded C2PA credentials, including Anthropic signing chains. Cleanup remains PNG/JPEG only.

Important limits

Pasted text receives a local reference-profile score, but it cannot become a Claude verdict while Anthropic's compatible detector profile remains private. Pixel marks and C2PA soft bindings remain unsupported.

Visual guide

See where a supported signal lives

Text findings are individual Unicode code points. Image findings live in metadata containers beside the pixels. The tool shows the layer it can inspect before you edit anything.

A text example revealing an invisible U+200B character between Hello and world, followed by an edited copy. — Text
Reveal exact characters

Inspect the code point, category, and position before choosing what to remove.

Diagram separating visible image pixels from C2PA, AI, and other metadata containers. — Images
Keep pixels separate from metadata

Supported metadata can be removed without repainting or recompressing the image pixels.

Statistical text watermarks

The signal builds one decision at a time

A generative watermark does not add hidden characters. It changes how an AI chooses among plausible next tokens, leaving a keyed pattern that can accumulate across a text.

One context, several plausible continuations

The harbor grew quiet as evening arrived.

quiet
still
dark

Ordinary generation

Fresh randomness makes the choice

Watermarked generation

A secret key and recent tokens shape the choice

How it works

Check first, remove only what you reviewed

The tool separates detection from removal. Your original text or file stays unchanged while you review supported findings and decide whether to create a cleaned copy.

1

Choose text or a file

Paste up to 500,000 characters or choose a supported file up to 25 MB.
2

Review the findings

See matching character types, code points, positions, metadata containers, or C2PA verification results.
3

Download a cleaned copy

Remove supported findings, re-check the result, then copy or download it.

Read the detection and removal method

Plain-language definition

What is a hidden watermark?

A hidden watermark is a signal that is not meant to be obvious in normal viewing. In text, that may mean invisible Unicode characters or statistical word-selection patterns. In images, it may mean metadata such as C2PA Content Credentials or a signal embedded in pixels. These methods require different detectors; no single check covers all of them.

Questions

Watermark checker and remover FAQ

Are hidden Unicode characters proof that text was written by AI?

No. Invisible characters can be legitimate in emoji, right-to-left writing systems, typography, or document formatting. A match shows which supported character exists; it does not identify who or what added it.

Does this checker detect SynthID or statistical text watermarks?

Watermark Tidy tokenizes pasted text locally and runs a verified DeepMind 0.2.1 reference scorer with public test keys. It reports raw statistics for an explicit GPT-2 test profile, not a provider verdict. Anthropic's official detector is in private preview and its compatible profile is not public, so Claude attribution remains unavailable.

Can this remove Claude's embedded text watermark?

No. It can remove supported invisible Unicode from text and supported C2PA structures from PNG or JPEG files. Anthropic describes its embedded text watermark as a separate signal. Its official detector is in private preview, but the compatible detector profile is not publicly reproducible.

What file watermark metadata can it check?

It validates embedded Content Credentials in supported image, video, and audio files. PNG and JPEG also receive structural metadata inspection and are the only formats with byte-preserving metadata removal. Anthropic signing-chain identity is reported separately.

Does removing supported image metadata change the pixels?

No. For supported removals, PNG image data and the JPEG compressed scan are copied byte-for-byte. The tool removes only identified metadata containers and re-checks the output before download.

Are my text and images uploaded?

No. Checks and removals run in your browser. The application has no upload endpoint, account system, analytics, or database. Loading the website itself still makes a normal request to its Cloudflare host.

Does a clean result prove an image or text has no watermark?

No. A clean result means only that no supported signal was found. Visible overlays, pixel-level marks, C2PA soft bindings, statistical text watermarks, and unknown vendor formats remain outside this check.

Technical terms follow the C2PA Content Credentials specification and the Unicode Standard. Last reviewed September 3, 2026.

Check and remove hidden watermarks