AI Watermark Center

AI Watermark Center

Understand What’s Hidden in AI Content

Inspect hidden Unicode, image metadata and Content Credentials with evidence-based tools that run in your browser.

0 characters

Open full Text Watermark Scanner
  • Runs in your browser
  • No account required
  • No content storage
  • Evidence-based results

Inspection tools

Inspect what is actually present

Live tools report characters and file signals you can independently observe. Provider-side watermarks stay labelled as provider-side.

  • Text

    Hidden Unicode

    • Zero-width characters
    • Bidi controls
    • Formatting marks
    • Normalization differences
    Open Text Scanner

  • Provider signals

    Documented, not detected here

    • SynthID — provider verification
    • Statistical watermarking — provider-specific
    • Official verification tools
    • Not an independent detector
    View provider research

Capabilities

What AI Watermark Center can verify

Local inspection means the browser reports what is in the bytes or characters you supply. Provider verification stays with the vendor.

  • Hidden Unicode

    Local inspection

  • EXIF / Metadata

    Local inspection

  • Embedded C2PA

    Local inspection

  • SynthID

    Provider verification

  • Statistical watermark

    Provider-specific

  • AI authorship

    Not determined

Not a scan result · product coverage only

How it works

Three steps. No invented scores.

Inspectors report what is present. They do not output confidence percentages or human-vs-AI classifications.

  1. 01

    Inspect

    Paste text or select a supported image. Processing stays in your browser.

  2. 02

    Understand

    See exact characters, metadata fields, or embedded Content Credentials when present.

  3. 03

    Verify

    Compare findings with dated provider documentation and separate evidence labels.

Provider research

Compare AI provenance systems

Sourced profiles for how major providers currently document watermarks and Content Credentials — and what this site can inspect independently.

Comparison table

  • Provider

    Claude

    Research published
    Text
    Statistical
    Generated file
    C2PA
    Public verify
    Announced
    AWC C2PA
    Available
    Last verified

    Anthropic documents a statistical text watermark for supported new models and C2PA Content Credentials on supported generated files. AI Watermark Center can inspect Unicode and embedded credentials; it cannot currently detect Claude’s keyed text watermark.

    View research

  • Provider

    OpenAI

    Research published
    Image
    C2PA + SynthID
    Audio
    SynthID
    Public verify
    Available
    AWC C2PA
    Available
    AWC SynthID
    Not available
    Last verified

    OpenAI documents layered image provenance (C2PA and SynthID) on supported ChatGPT, Codex, and API outputs, SynthID on supported generated audio, a public verifier, and a Content Provenance API. A current OpenAI text watermark is not established. Research profile and image cluster available; AWC inspects embedded C2PA only.

    View research

  • Provider

    Gemini

    Research published
    Text
    SynthID
    Image
    C2PA + SynthID + Visible mark
    Audio
    SynthID
    Video
    SynthID
    Public verify
    Available
    AWC C2PA
    Available
    AWC SynthID
    Not available
    Last verified

    Google DeepMind documents SynthID across image, video, audio, and Gemini text. Gemini Apps add invisible SynthID and Content Credentials on generated or edited media. Gemini’s verifier identifies Google AI SynthID, not every company’s watermark.

    View research

  • Provider

    Adobe Firefly

    Research published
    Image
    C2PA
    Public verify
    Available
    AWC C2PA
    Available
    AWC SynthID
    Not available
    Last verified

    Adobe documents automatic Content Credentials on fully Firefly-generated assets, with Adobe as issuer and optional cloud storage for recovery. SynthID use is not established from the documentation reviewed here.

    View research

Database

Documented claims, not a complete catalogue

Four current records from the evidence database. Claim status, evidence authority, and inspection support remain separate fields.

Explore all evidence records
Claude, Statistical, Documented. Expand for full claim, sources, and limits.

Anthropic · Claude · Text · Statistical watermark

ClaimClaude uses a statistical / SynthID-Text-derived text watermark on supported new models

Current status
Documented
Evidence authority
Official documentation
AWC support
Not available
Provider public detection
Announced, not generally released
AWC integration
Not integrated
Last verified
Support detail
Statistical watermark detection is not available. Anthropic’s keyed detector is announced, not integrated here.
Product / model / surface scope
Supported new models, including models launched in the EU on or after 2 August 2026.
Limitations
Detection depends on a key. Short, factual, proofread, and code-heavy text may carry less usable signal. Older outputs are not assumed to be marked.
Notes
Hidden Unicode is not the official watermark mechanism. Independent verification requires Anthropic’s detection capability.
Primary sources
OpenAI, C2PA, Documented. Expand for full claim, sources, and limits.

OpenAI · ChatGPT image generation, OpenAI API, Codex · Image · C2PA / Content Credentials

ClaimSupported OpenAI-generated images include C2PA Content Credentials

Current status
Documented
Evidence authority
Official documentation
AWC support
Available
Provider public detection
Publicly available
AWC integration
Not integrated
Last verified
Support detail
Embedded Content Credentials can be inspected for JPEG, PNG, and WebP. Generic C2PA is not proof the file is OpenAI-generated; provider identification requires credential data. OpenAI’s verifier is not integrated here.
Product / model / surface scope
Current supported OpenAI image generation (ChatGPT, Codex, OpenAI API) as documented. Coverage varies by model, time, export path, and file type.

Surface: supported current image generation

Limitations
Metadata can be stripped. Screenshots typically omit C2PA. AWC does not claim OpenAI authorship from a non-OpenAI manifest. Missing C2PA does not mean the file is not OpenAI-generated.
Notes
Current Help Center documentation lists C2PA on supported ChatGPT, Codex, and OpenAI API images. OpenAI began adding Content Credentials to DALL·E 3 images in 2024; that chronology is historical and does not mean every subsequent file still carries a readable manifest.
Gemini, SynthID, Documented. Expand for full claim, sources, and limits.

Google · Gemini image / Nano Banana family · Image · SynthID

ClaimGoogle AI-generated image outputs in supported products contain SynthID

Current status
Documented
Evidence authority
Official documentation
AWC support
Not available
Provider public detection
Publicly available
AWC integration
Not integrated
Last verified
Support detail
SynthID detection is not available. Gemini’s verifier recognizes Google AI SynthID, not every company’s SynthID.
Product / model / surface scope
Supported Google generative image products, including Gemini Apps and documented Nano Banana / Gemini API image models.

Surface: Gemini Apps, Gemini API, and other documented Google tools

Limitations
Gemini currently identifies Google AI SynthID, not a universal SynthID detector. Historical Nano Banana files are not assumed to use identical policy. AI Watermark Center does not detect SynthID.
Notes
DeepMind documents SynthID on AI-generated images. Gemini Apps say all AI-generated media created or edited with Apps includes invisible SynthID. Gemini API image-generation docs say all generated images include SynthID. Nano Banana is a versioned Gemini image family, not one timeless model.
Adobe Firefly, C2PA, Documented. Expand for full claim, sources, and limits.

Adobe · Adobe Firefly · Image · C2PA / Content Credentials

ClaimAdobe Firefly automatically applies Content Credentials to fully generated assets such as Text to Image

Current status
Documented
Evidence authority
Official documentation
AWC support
Available
Provider public detection
Publicly available
AWC integration
Not integrated
Last verified
Support detail
Embedded Content Credentials can be inspected for JPEG, PNG, and WebP. Adobe cloud recovery is not performed. Adobe Inspect is a separate official tool.
Product / model / surface scope
Adobe Firefly features such as Text to Image where 100% of pixels are generated with Firefly.
Limitations
100%-generated Firefly assets are the documented automatic case. Edited or mixed-pixel workflows are not assumed to behave identically. Cloud-only credentials are not retrieved here.
Notes
Credentials can include Adobe as issuer, issue date, application or device, AI tool used, and actions. They may be attached to the file and may also be stored in Adobe’s public Content Credentials cloud.
Primary sources

Research

Independent research, with dates

Lifecycle labels come from the published research registry. Does Claude Hide Watermarks in Unicode is collecting. Nano Banana Image Provenance Test: SynthID, Metadata & Content Credentials and OpenAI Image Provenance Test: C2PA, Metadata & Provider Verification remain dataset pending. No fabricated sample counts.

All research

Research & regulation

Latest research & regulation

Short paths into regulation and provenance topics that sit beside the live inspectors.

All guides

Methodology

We only claim what we can verify.

Pages distinguish official documentation, independent testing, limited observations, and unverified claims. Claim status is a separate field from evidence authority.

Read the methodology

These are independent axes. Evidence authority describes source strength. Claim status describes polarity. Official evidence can document a mechanism or rule one out; it is not a positive result by itself.

Evidence authority

Source strength

Neutral with respect to whether a claim is true.

  • Official
  • Verified
  • Observed
  • Unverified

  • Official documentation

    An official source supports the evidence. That is not the same as confirming the claim.

  • Independently verified

    AI Watermark Center independently reproduced the observation.

  • Observed

    A limited or single-path observation. Not a completed study.

  • Unverified

    The claim is recorded, but this site has not independently verified it.

Claim status

Factual polarity

Independent of evidence labels.

  • Documented
  • Ruled out
  • Conditional / partial
  • Unknown

  • Documented

    Provider documentation supports the mechanism as described.

  • Ruled out by provider documentation

    Provider documentation rejects the proposed mechanism.

  • Conditional / partial

    Coverage is conditional on product, file type, or export path.

  • Unknown

    Unknown is not “does not exist.”

FAQ

Straight answers about what this product does

Answers reuse the same evidence boundaries used across tools, research, and provider pages.

What is an AI watermark?

The phrase is used loosely for visible labels, hidden Unicode in copied text, cryptographic manifests such as C2PA Content Credentials, and statistical patterns a model owner may detect. Those mechanisms are not interchangeable.

Can AI Watermark Center detect SynthID?

No. AI Watermark Center does not detect SynthID. Where providers publish verification tools, those results are labelled provider verification — not an independent AWC detection.

What can the text scanner detect?

It lists supported hidden and unusual Unicode characters that are actually present in the pasted text, including zero-width marks, bidi controls, and related formatting characters. It does not detect statistical text watermarks.

What can the image inspector detect?

It inspects ordinary metadata (such as EXIF, GPS, XMP/IPTC where supported) and embedded Content Credentials. It does not detect SynthID, recover remote credentials, or decide whether an image is AI-generated.

Are images or text uploaded?

No. Text and image inspection run in your browser. Content is not uploaded to AI Watermark Center for analysis.

Does no detected signal mean human-created?

No. Absence of a detectable signal is not proof that content was not generated by AI. Successful removal of an invisible watermark is not claimed unless it can be independently verified.

What does C2PA prove?

Embedded Content Credentials can show signed provenance claims attached to a file when a manifest is present. Missing credentials do not prove a file is human-made, and credentials can be stripped or never written.

More detail in the methodology and guides.

Guides

Learn the mechanisms

All guides

AI Watermarks

What Are AI Watermarks?

A primer on visible labels, invisible statistical watermarks, cryptographic provenance, and why a missing signal does not prove content is human-written.

Published

Read guide

C2PA

What Is C2PA and Content Credentials?

An explainer of the C2PA standard and Content Credentials: signed provenance attached to files, how it differs from invisible watermarks, and what a missing manifest does and does not mean.

Updated

Read guide

Independent inspection

Start with what you can inspect

Paste text or open an image in the browser. Provider-side marks stay labelled as provider-side. We only claim what we can verify.

Scan AI TextInspect an Image