Use the Content Provenance API to check whether an image, audio file, or PDF
contains supported OpenAI provenance signals. Send a file to
POST /v1/content_provenance_checks to receive the completed verification
results in the same response. Use these signals in content review,
fact-checking, labeling, and trust and safety workflows.
To check a file in your browser, use the web tool at openai.com/verify.
For request parameters and response schemas, see the Content provenance API reference.
To check text, apply for access. Text verification is currently available only to approved organizations including AI research and academic institutions. OpenAI reviews each application on a case-by-case basis.
A not_detected result means the tool didn’t find supported signals in the
uploaded file. Content may still have been generated by OpenAI if its metadata
was stripped or shows evidence of tampering, its watermark was degraded, it
came from a legacy generation model, or it was created before provenance
signals were available. The tool doesn’t currently detect content generated by
another company’s AI model, so a not_detected result doesn’t rule that out
either.
What content provenance checks
Content provenance checks supported files for the following signals:
| Signal | Applies to | What it checks |
|---|---|---|
| C2PA Content Credentials | Images and PDFs | Signed metadata with issuer and AI-use details |
| SynthID | Images and audio | A watermark embedded directly in supported media |
C2PA metadata provides more context about a file’s origin. Editing, converting, or sharing a file can remove its metadata. A SynthID watermark is part of the image or audio itself and may survive some transformations.
Supported PDF exports from ChatGPT and the Responses API include embedded C2PA metadata. These exports can include PDFs you upload and download again without changing their contents. Upload the exported PDF to the Content Provenance API to check its metadata.
The API checks for supported OpenAI signals. It isn’t a general-purpose AI detector and doesn’t identify content generated by every AI system. Visible watermarks and labels are separate from the provenance signals checked by the API.
Verify a file
Send an image, audio file, or PDF as the file field with the OpenAI SDK. The
SDK builds the multipart request and reads your API key from the
OPENAI_API_KEY environment variable:
import { createReadStream } from "node:fs";
import OpenAI, { toStreamingFile } from "openai";
const client = new OpenAI();
const result = await client.contentProvenanceChecks.create({
file: toStreamingFile(createReadStream("myimage.png"), "myimage.png", {
type: "image/png",
}),
});
console.log(result);Use these OpenAI SDK versions or later: Python 2.52.0, Go 3.49.0, and Ruby 0.75.0.
To verify Opus audio, use the same endpoint and set the uploaded file’s media
type to audio/ogg:
curl https://api.openai.com/v1/content_provenance_checks \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-F "file=@./example.opus;type=audio/ogg"
To verify a PDF with the same command, use the path to your PDF and set the
file part’s media type to application/pdf.
The response contains the completed result. For example, an image returns:
{
"object": "content_provenance_check",
"created_at": 1778000000,
"results": [
{
"type": "c2pa",
"outcome": "detected",
"validation_state": "trusted",
"issuer": "OpenAI OpCo, LLC",
"model": "gpt-image",
"generated_at": "2026-07-27T18:34:12Z"
},
{
"type": "synthid",
"outcome": "not_detected",
"model": null,
"generated_at": null
}
]
}
The object field identifies the response, and created_at is the check’s
creation time as a Unix timestamp in seconds. The entries in results depend on
the uploaded file. Images include C2PA and SynthID results, audio includes a
SynthID result, and PDFs include a C2PA result. The API omits checks that don’t
apply instead of returning not_detected.
The API completes verification before it returns. You don’t need to create a background job, poll another endpoint, or upload the file to the Files API.
If a request fails, check the HTTP status and error.code when available. A
malformed, unsupported, or blocked file returns 400; an organization without
access receives 404; and requests over the rate limit return 429. Retry only
transient failures, such as rate limits or server errors. For general guidance,
see API error codes.
Understand verification results
Read each applicable entry in results independently. The response doesn’t
include a top-level outcome.
C2PA results
A C2PA result describes the state of the Content Credentials in an image or PDF. This example shows an image result.
{
"type": "c2pa",
"outcome": "detected",
"validation_state": "trusted",
"issuer": "OpenAI OpCo, LLC",
"model": "gpt-image",
"generated_at": "2026-07-27T18:34:12Z"
}
Use the fields as follows:
outcomeindicates whether a supported OpenAI C2PA signal wasdetectedornot_detected.validation_stateindicates whether the manifest istrusted,valid,invalid, ornot_present.issueridentifies the manifest issuer when that information is available.modelidentifies the model recorded in the provenance signal, when available.generated_atidentifies the content’s generation time when that information is available.
For images, the outcome is detected only when a trusted or valid manifest
identifies OpenAI as its issuer and includes an AI-generation action. A third-party
manifest, a manifest without an AI-generation action, an invalid manifest, or
a not_present manifest produces not_detected. The issuer and
validation_state can still describe a manifest even when the outcome is
not_detected.
For PDFs, a detected outcome requires a trusted OpenAI manifest with a
supported provenance signal. A trusted manifest
meets the requirements for a valid manifest
and has a signer certificate chain that leads to an accepted trust anchor.
A detected PDF signal indicates that an OpenAI tool processed the file. The
signal alone cannot establish whether, or to what extent, the tool generated
or edited its contents.
Don’t treat an invalid manifest as reliable provenance evidence. A
not_present result means the file has no available C2PA manifest.
SynthID results
A SynthID result describes whether the verifier detected a supported watermark in an image or audio file:
{
"type": "synthid",
"outcome": "detected",
"model": null,
"generated_at": null
}
An outcome of detected means the file contains a recognized watermark. An
outcome of not_detected means the verifier didn’t detect that watermark. It
doesn’t rule out AI-generated or AI-modified content. model and
generated_at provide the generating model and generation time when available;
either field can be null.
Supported uploads
By default, the API supports the following file formats:
- Images: PNG, JPEG, WebP, and GIF.
- Audio: MP3, Opus, AAC, FLAC, WAV, and PCM.
- Documents: PDF.
Limit each uploaded file to 50 MiB. Audio must be 60 seconds or shorter after decoding.
Set the uploaded file part’s media type. Use image/gif for a GIF, image/png
for a PNG image, audio/ogg for Opus audio, or application/pdf for a PDF.
Don’t add a separate type field or manually set the multipart/form-data
request header. The curl -F option sets the request content type and multipart
boundary. Send one file per request.
Rate limits and Zero Data Retention (ZDR)
The public Content Provenance API is eligible for Zero Data Retention (ZDR) when you verify images or audio at default rate limits.
Strict rate limits help protect the API against misuse. Organizations can apply for higher limits, and OpenAI reviews each application on a case-by-case basis.
Access with higher rate limits is ineligible for ZDR.
If the API returns 429 rate_limit_exceeded, reduce your request rate and
honor the Retry-After header when present. See
rate limits for general retry guidance.
Use verification results responsibly
Use verification results as evidence in a broader review process:
- Treat
detectedas evidence of a specific supported signal, not a complete history of a file. - Treat
not_detectedas an absence of detected evidence, not proof that the content is human-created or wasn’t generated with OpenAI. - Check the C2PA issuer before attributing an image or PDF to a particular provider.
- Verify the original file when possible. Compression, cropping, screenshots, metadata removal, and format conversions can erase or weaken a signal.
- Account for the originating product, model, file format, and creation date. Not all OpenAI-generated content contains a supported signal.
- Pair automated decisions with human review in high-stakes workflows.
- Don’t use repeated queries to reverse-engineer, remove, or evade a watermark.
- Don’t infer a prompt, account, or individual creator from a verification result.
Using the Content Provenance API is subject to the OpenAI Services Agreement.
For information about platform-wide monitoring and retention settings, see data controls.