OpenAI has launched a verification page and developer API that check images and audio for provenance signals associated with its own generation tools. The service looks for C2PA metadata and SynthID watermarks, then reports whether one of those signals was detected.
The distinction between a positive and negative result is central to the product. A detected signal can support a conclusion that the file came through a compatible OpenAI workflow. A result with no signal does not prove that the file is human-made, nor does it rule out earlier OpenAI generation.
Two signals carry different kinds of evidence
C2PA is metadata attached to a file to record information about its origin and editing history. It can provide structured provenance, but metadata may be removed when a file is copied, transformed, captured again or processed by software that does not preserve it.
SynthID is an embedded watermark designed to survive more of that handling. OpenAI's checker can look for SynthID in supported images and audio, while the C2PA check applies to images. Watermarks can still be degraded, and legacy content may have been created before a supported signal was added.
Using both methods gives the service more than one route to a positive result. It does not turn the checker into a forensic verdict. OpenAI explicitly says it is not a general-purpose detector and does not identify material made by other providers.
The browser handles individual checks
The public-facing verification page accepts PNG, JPEG and WebP images. For audio, it accepts MP3, WAV, AAC, FLAC, Opus and PCM files. The result can indicate that C2PA was detected, SynthID was detected, or that the service found no supported signal.
OpenAI says uploaded material is processed for the verification request, is not stored unless legally required and is not used to train its models. Those statements address the browser workflow at a high level, but organisations dealing with confidential recordings or regulated material should still assess the service against their own handling requirements.
The API is synchronous and deliberately narrow
Developers can submit a file to POST /v1/content_provenance_checks. The endpoint returns a synchronous result, making it suitable for a review screen, moderation queue or labelling process in which a user or system needs an answer before continuing.
The documented limits reinforce that narrow purpose. Each request accepts one file, with a maximum size of 50 MiB. Audio can be no longer than 60 seconds. The supported image formats are PNG, JPEG and WebP; supported audio formats are MP3, Opus, AAC, FLAC, WAV and PCM.
OpenAI also notes that content provenance checks are not eligible for Zero Data Retention. Rate limits are described as strict, with higher limits available by application. The documentation suggests that access may depend on the organisation: a request from an organisation without access can return a not-found response.
A missing mark is not an all-clear
The most likely misuse would be to treat “no signal” as proof of authenticity. A provenance marker can disappear through ordinary file handling, deliberate removal or degradation. Content from another provider is outside the tool's detection scope, and older OpenAI output may not contain the supported markers.
That means the checker is strongest when it finds something. A positive C2PA or SynthID result supplies a useful piece of evidence for fact-checkers, platforms and trust-and-safety teams. A negative result should remain one input among others, not the end of an investigation.
The launch gives OpenAI customers a direct way to interrogate signals that were previously less visible to end users. It also exposes the limits in plain terms. Provenance can help trace compatible media, but it cannot settle the much broader question of whether any arbitrary file was created or altered by AI.