Text: no public check
OpenAI's public provenance documentation does not provide a tool where you paste ChatGPT text and test for an official watermark.
ChatGPT watermark:
What is marked, and can you check it?
The important answer depends on the format. Text, images and audio do not use the same public verification system.
Supported provenance signals
The essentials in one minute.
Three facts prevent the most common misunderstandings.
Images: two signals
Supported images can be checked for signed C2PA Content Credentials and an embedded SynthID watermark.
Audio: SynthID
Supported audio can be checked for an embedded SynthID signal. A result is evidence of a signal—not a complete history.
OpenAI uses two different kinds of media provenance.
One travels as signed information in an image file. The other is embedded in the media itself.
SynthID: embedded in media
OpenAI says SynthID can mark supported images and audio. The signal is part of the media and may survive some transformations.
It is not a visible logo and cannot be found by reading image metadata alone.
C2PA: signed image history
Content Credentials can carry signed information about an image's origin. The verifier also checks whether the record is cryptographically valid.
Editing, conversion or sharing can remove this metadata.
What happens when you check a file?
The verifier looks for supported OpenAI signals. It does not guess from whether the file “looks like AI.”
1
You provide the original file
Use an image or a short audio file. The original is best because screenshots, exports and conversions can lose information.
2
Each applicable signal is checked
For images, the result can include both C2PA and SynthID. For audio, the documented check is SynthID.
3
C2PA checks the signed record
If an image has Content Credentials, the verifier checks the issuer and signature. Model or generation time may be present, but can also be unavailable.
4
SynthID checks the media itself
The detector looks for an embedded OpenAI-supported signal. It returns detected, not detected or an error—not the identity of a person.
Technical explanationC2PA signatures, SynthID and API limits
These systems answer different questions. C2PA is a signed manifest attached to an image file. SynthID is an embedded watermark that OpenAI says can be detected in supported images and audio. OpenAI's verifier reports each applicable signal separately.
C2PA validates a manifest
The verifier parses the image's Content Credentials, validates its cryptographic signature and reports the issuer. Optional details such as model and generation time may be missing.
SynthID checks an embedded signal
The detector examines the media for an OpenAI-supported watermark. Because it is embedded rather than stored only as metadata, it may survive some transformations—but edits can still weaken it.
The API returns bounded results
The Content Provenance API accepts one file per request, up to 50 MiB. Audio must be no longer than 60 seconds. Results are not a universal AI-origin verdict.
Supported public checks
Images: PNG, JPEG and WebP, with C2PA and SynthID results where applicable.
Audio: MP3, Opus, AAC, FLAC, WAV and PCM, with a SynthID result.
What is not publicly documented
- No pasted-text check appears in OpenAI's public provenance workflow.
- The documentation does not promise that every OpenAI output contains every signal.
- The verifier does not identify the prompt, account or person behind a file.
- The public guide does not provide a complete model-by-model coverage list.
Why use both? C2PA can carry useful signed context, but metadata is easy to lose. An embedded signal can be more durable, but it carries less human-readable context and can degrade after transformation. Together they provide stronger evidence than either one alone.
Read the result without overclaiming.
A provenance result is useful evidence, but it is not a verdict about authorship, truth or misconduct.
Detected
What you can sayA supported OpenAI signal was found.
Check which signal was detected and, for C2PA, whether the issuer and signature are valid.
What you cannot sayIt does not identify a person or prompt.
It also does not prove the file is true, unchanged in every way or entirely generated by AI.
Not detected
What you can sayNo supported signal was found in this file.
The result describes this file and the signals the verifier currently supports.
What you cannot sayIt does not prove human authorship.
The signal may be absent, removed, degraded, unsupported or created by another provider.
Remember: “Not detected” is not the same as “made by a human.”
What can you check right now?
Use OpenAI's own verifier for supported files. Do not confuse it with a writing-style detector.
OpenAI Verify
Checks supported OpenAI provenance signals in image and audio files.
- Images: C2PA and SynthID
- Audio: SynthID
- No pasted-text field in the public verifier
Generic AI text detector
Usually guesses from writing style or statistical patterns. It is not an official OpenAI provenance check.
- Can return false positives
- Cannot prove who wrote a text
- Should not support an accusation alone
Why can a signal disappear?
C2PA metadata and an embedded SynthID watermark fail in different ways.
What happens to the fileWhat the result may show
C2PA Content Credentials in images
The original image is checkedThe signed record may verifyThe issuer and available origin details can be read.
A screenshot or new export is madeThe metadata may be goneThe new file may not carry the original manifest.
SynthID in images and audio
The file is lightly transformedThe signal may remain detectableAn embedded watermark can survive some transformations.
The file is heavily edited or compressedThe signal can weakenCropping, compression, conversion and other changes can reduce detectability.
An unsupported or legacy output is checkedNo supported signal may be foundNot all OpenAI content contains every supported signal.
What is covered by the public verifier?
The public documentation defines formats and limits, not a promise that every OpenAI model or output is marked.
Images
PNG, JPEG and WebP files can be checked. Results may include C2PA and SynthID.
Audio
MP3, Opus, AAC, FLAC, WAV and PCM files can be checked for SynthID. Audio is limited to 60 seconds.
Text
The public provenance guide and verifier do not document a pasted-text watermark check. Do not turn that absence into a stronger claim.
Use the original fileIt gives the verifier the best evidence
One file per checkMaximum documented size: 50 MiB
Expect incomplete contextModel and generation time may be unavailable
What should you do with the result?
Use provenance as one piece of evidence and match the response to the situation.
Personal use
If you only want to understand an image or audio file, start with the original and read each signal separately. A missing signal is not proof of human origin.
School or work
Do not use a generic text detector as proof that ChatGPT wrote an assignment or email.
- Ask what the tool actually checks
- Review drafts, sources and version history
- Let a person assess the full context
Publishers and businesses
Machine-readable provenance and visible disclosure are separate questions. EU rules can require providers to mark output and can also require visible disclosure in specific public-interest uses.
Developers
The Content Provenance API can automate supported checks, but OpenAI says results should be combined with context, human review and clear uncertainty.
Writing in Danish? Fix spelling and grammar.
Paste your Danish text into SkrivSikkert and correct common mistakes.
Correct my Danish textFrequently asked questions.
Does ChatGPT watermark text?
OpenAI's public provenance documentation does not provide a pasted-text watermark check. That does not prove that no internal, experimental or future text signal exists; it means users do not currently have that public documented check.
Can I check a ChatGPT image or audio file?
Yes. OpenAI provides a public Verify page and a Content Provenance API for supported image and audio formats.
What is the difference between C2PA and SynthID?
C2PA Content Credentials are signed provenance metadata in an image file. SynthID is an embedded watermark used in supported images and audio. Metadata can be removed more easily; an embedded signal may survive some transformations.
Does “detected” prove OpenAI created the whole file?
No. It means the verifier found a supported OpenAI provenance signal. It does not identify a person, account or prompt, and it does not provide the file's complete history.
Does “not detected” prove a file is human-made?
No. A signal may have been removed or degraded, the output may be unsupported or older, or the file may come from another AI provider.
Why should I use the original file?
Screenshots, re-exports, format changes, cropping and compression can remove C2PA metadata or weaken an embedded watermark.
Can the verifier identify the ChatGPT account or prompt?
No. OpenAI's documentation says results should not be used to infer the prompt, account or person behind the content.
Does EU law require a visible label on every AI output?
No. Provider-side machine-readable marking and user-facing visible disclosure are separate obligations. Visible disclosure depends on the content and how it is used.
Sources and further reading.
OpenAI's own documentation comes first. EU and C2PA sources explain the legal and technical context.
OPENAI · PRIMARY SOURCEContent provenance and detectionOpenAI Developers ↗
OPENAI · PUBLIC TOOLVerify an image or audio fileOpenAI Verify ↗
EUROPEAN COMMISSIONQuestions and answers on Article 50European Commission ↗
EU CODE OF PRACTICETransparency of AI-generated contentEuropean Commission ↗
OPEN STANDARDContent Credentials explainedC2PA ↗