An AI watermark can support a limited finding: a particular provider or supported tool was likely involved in creating or processing the content. It cannot, by itself, prove who used the tool, how much AI contributed, whether the item was edited, or whether anyone misrepresented it. If a check finds nothing, the result is inconclusive.
For example, if a supported OpenAI provenance signal is detected in an image, a careful report says that a supported OpenAI signal was found in that file. It does not say that the person who submitted it created the image, that the image is unchanged, or that AI produced every part of it. Provenance records, embedded signals, and handling history are separate evidence layers.
A detected signal supports a bounded claim about likely provider involvement. No detected signal is not proof of human origin.
What an AI watermark can and cannot prove
A recognized watermark is evidence that a supported system probably contributed to the content. The strength and specificity of that evidence depend on the detector, the signal it recognizes, the file, and the provider’s documentation.
A watermark does not establish the user’s identity, intent, ownership, or the amount of AI contribution. Someone may substantially edit AI-generated material, or use AI to edit material they created themselves. A signal alone cannot distinguish those cases. A detector’s not detected
result means only that it did not find a signal it supports in the checked item.
Use this three-layer model as a practical editorial framework, not an official standard:
- Provenance metadata: Records such as C2PA Content Credentials may describe an item’s origin or editing history, depending on what the issuing tool recorded.
- Embedded signal: A provider-specific watermark, such as SynthID, is carried within the media or reflected statistically in generated text.
- External context: The original file, source, upload history, transformations, and custody records establish what was checked and how it was handled.
Keep each layer’s result separate. A validated metadata record is not the same observation as a detected embedded watermark, and neither replaces surrounding evidence.
Three different things people call an AI watermark
Content Credentials are provenance records based on the C2PA standard. A manifest can include claims about origin or editing history when the issuing tool records them. C2PA is not inherently an AI label: a camera or ordinary editing application can create provenance information too. Read the recorded claims and validation status rather than assuming that any manifest means AI-generated.
See the Content Credentials explanation for the distinction between provenance and a particular generation method.
Visible marks are labels, logos, or text placed on content for people to see. They can communicate disclosure, but they are separate from machine-readable credentials and invisible watermarks. A visible mark may be cropped or edited out, so its presence or absence is not a complete provenance check.
Invisible watermarks are signals embedded in media or represented in text-generation patterns. They are designed to be unobtrusive and detectable by compatible tools, but detection is not guaranteed after every transformation. Metadata may be stripped during conversion or screenshotting, while an embedded signal may remain detectable after some changes.
How text and media watermarks work
Google describes SynthID Text as a method that adjusts token-selection probabilities during text generation. In simplified terms, the system nudges choices among plausible next tokens to create a statistical pattern that a compatible detector can test across enough text. It does not depend on one unusual character or a repeated word substitution. Google says detection is more reliable with longer, varied passages and can weaken after thorough rewriting or translation, or when the text has few plausible alternatives, as in some factual answers or code. Read Google DeepMind’s explanation of SynthID Text for the method and stated limits.
For images, video, and audio, SynthID implementations embed signals in generated media. Google says its image and video marks are designed to withstand some common transformations, including cropping, filters, frame-rate changes, and lossy compression. Designed to withstand
is not a guarantee that a detector will identify a signal after every edit. Google also describes detection coverage for Google and certain partner signals, including OpenAI, NVIDIA, and Kakao. That does not make the detector universal. See Google’s SynthID overview for current scope.
A screenshot can lose provenance metadata and always has a different file hash from the original. An embedded signal may still be detectable, but the screenshot must be checked and recorded as its own file. Preserve and check the original whenever possible.
What major providers document today
This provider snapshot was checked on November 12, 2026. Support varies by product, model, region, media type, and version. A provider’s general statement should not be read as a promise about every output.
Google documents SynthID for images, video, audio, and text. Its current overview describes detector support for Google and certain partner signals, including OpenAI. This is documented partner coverage, not universal detection of all models or watermark methods. Google explains the text method and its limitations, but the sources reviewed do not establish a general public Google text-detection endpoint.
OpenAI
OpenAI says supported image outputs can include C2PA credentials and SynthID, while supported audio outputs can include SynthID. Its current public verification tool and documented Content Provenance API support supported image and audio files. The reviewed documentation does not establish general video verification. OpenAI also documents textGrain watermarking for ChatGPT users in the EU and limited text-detector access for qualifying organizations.
Anthropic
Anthropic says Claude’s text watermark is a version of the SynthID Text approach. Its detector is not documented as a general public self-serve tool. Anthropic describes private-preview access for eligible organizations, including certain regulators, media organizations, researchers, educational organizations, civil-society groups, and qualifying enterprises.
Google, OpenAI, and Anthropic describe different products and access routes. Before choosing a detector, confirm its current supported provider, modality, file type, access route, and result semantics in official documentation: OpenAI’s provenance overview, its public verification tool, developer documentation, or Anthropic’s watermark announcement.
Where EU regulation fits
The European Commission states that Article 50 transparency obligations apply from August 2, 2026, for covered roles and circumstances. The precise obligation depends on the provider or deployer, the content, and the use case. Covered approaches can include marking, labeling, or making certain AI-generated or manipulated content detectable.
The law does not require every provider to use one technology such as SynthID. The Code of Practice on Transparency of AI-Generated Content is a voluntary compliance pathway, while the underlying legal obligations apply within their scope. Signing the code does not prove that every product output carries a particular signal. Consult the Commission’s transparency guidance and Code of Practice page for current scope. This is context, not legal advice.
A practical workflow for checking a file
For a publisher or operations team, the input is the original file plus an asset identifier and intake details. The content operations owner should keep the asset and verification events in an asset registry or case-management system. Use a provider’s documented tool only for supported formats and with appropriate authorization.
The following is a proposed internal record, not an OpenAI response schema. The values are illustrative application statuses. Preserve vendor fields and responses separately rather than assuming these are their exact names.
{
"asset_id": "asset_1042",
"file_sha256": "illustrative_hash_of_exact_received_file",
"checked_at": "illustrative_timestamp",
"verifier_provider": "OpenAI",
"verifier_configuration": "illustrative_configuration_identifier",
"c2pa_validation_state": "value_from_documented_response",
"c2pa_issuer": "value_if_returned",
"synthid_outcome": "value_from_documented_response",
"review_status": "pending"
}One verification event means one check of one exact file hash by one verifier configuration at one time. An event may contain separate C2PA and embedded-signal observations. If a crop or re-encoding is checked later, create another event linked to the new hash and, where known, the original file’s hash.
For concurrent processing, use a database-enforced uniqueness constraint or transactional upsert on the intended event key. A read-then-insert check alone can race. For content-level deduplication, a proposed key might include the content hash, verifier, verifier configuration, and normalized verification version. For chain-of-custody auditing, preserve separate intake events even when their content hashes match. These are proposed application rules, not vendor schemas.
- Confirm the selected verifier supports the exact media type and access route.
- Preserve the original binary, exact hash, raw response reference, and check timestamp.
- Keep C2PA validation and embedded-signal outcomes as separate fields.
- Route invalid, conflicting, unsupported, and consequential findings to a named human owner.
Use deterministic rules for file type and size checks, hash calculation, manifest validation, required-field parsing, and duplicate handling. If a file is unsupported, a manifest is invalid, or verification fails, record that specific state rather than converting it to human-made.
Before sending sensitive audio, personal data, biometric material, or confidential files to an external verifier, check authorization, privacy, retention, and data-handling requirements.
Teams designing dependable operating workflows may also need to define ownership, retention, and review routes around these checks. This link concerns broader workflow design, not a watermark detector.
Text watermark checks need a different path
The provider documentation reviewed here does not establish a universal public text-watermark detector. Anthropic documents private-preview detector access for eligible organizations. OpenAI documents textGrain for ChatGPT users in the EU and limited detector access for qualifying organizations. Google explains SynthID Text and its limitations, but the reviewed sources do not establish a general public Google text-detection endpoint.
If you are authorized to use a provider-specific detector, preserve the full passage and record the claimed provider, language, translation history, editing or rewriting history, detector used, access basis, and check time. A short excerpt is a poor basis for a decisive assessment. Lower confidence is appropriate for factual answers, code, mathematics, fixed quotations, structured data, translated text, and heavily rewritten passages.
A positive text result supports likely provider involvement, not exclusive authorship, user identity, ownership, or a measurement of AI contribution. Attach the result to an editorial or review record, not an automatic rejection rule.
Frequently asked questions
Does no watermark prove content is human-made?
No. It means only that the selected check found no supported signal in that file or passage. The source may be unsupported, metadata may be absent, or a transformation may have affected detection.
Can a screenshot preserve a watermark?
An embedded signal may survive some transformations, but metadata may be lost. A screenshot is a separate file and should have its own hash and verification event.
Is C2PA the same as an AI watermark?
No. C2PA is a provenance standard that can record different kinds of origin and editing information. It does not automatically mean an item was generated by AI.
Can one detector check all AI content?
No. Coverage depends on the supported provider, signal, modality, and detector. Google documents some partner coverage, but no universal detector was established in the sources reviewed.
Can OpenAI’s current public verifier check video?
The current OpenAI documentation reviewed supports supported image and audio files. Do not claim general video verification without a newer official OpenAI reference.
Can anyone use Anthropic’s text-watermark detector?
Anthropic documents private-preview access for eligible organizations, not a general public self-serve detector.
Sources and scope
Provider capabilities and access routes change. Consult the current Google SynthID overview, OpenAI Content Provenance documentation, OpenAI verification tool, Anthropic announcement, and European Commission guidance before implementing a production process. This article explains evidence interpretation and proposes an operating pattern; it is not legal advice or a vendor integration guide.