No. OpenAI’s public verifier cannot identify every image created with artificial intelligence. It checks an uploaded file for supported provenance signals associated with OpenAI tools, including C2PA metadata and invisible SynthID watermarks. If it finds no signal, that does not prove the image is an authentic photograph or that no AI was involved.
This distinction matters when assessing images shared on social media. An AI-generated image may reach the verifier without metadata, with a weakened watermark, or after being produced by a different provider. This guide interprets documentation available as of October 24, 2026. No experimental files were uploaded, no audit of the service was conducted, and no independent detection accuracy was measured.
What OpenAI Verify actually checks
At openai.com/verify, users can upload a supported PNG, JPG or WEBP image and receive information about recognized signals. The service also supports eligible audio files, but this article focuses on images. Its documented scope covers content associated with ChatGPT, Codex and the OpenAI API, not every image-generation service on the internet.
C2PA stands for Coalition for Content Provenance and Authenticity. It is an open standard for recording the provenance of digital media. Its Content Credentials may identify the issuing tool and parts of the file’s history. Cryptographic signatures help reveal changes that do not match those records. Cameras and news organizations can also use C2PA, so the mere presence of credentials does not mean an image was generated by AI.
SynthID is an invisible watermarking technology developed by Google DeepMind and also used by OpenAI for supported content. The signal is embedded in the media and may survive certain edits that remove metadata. That resilience is not absolute. Also, different companies using the same underlying technology does not turn OpenAI’s verifier into a universal detector of signals issued by other providers.
In the Content Provenance API, the documentation reports C2PA and SynthID results separately. A third-party C2PA manifest, an invalid or missing manifest, or one that does not record an OpenAI AI-generation action is not counted as a positive OpenAI C2PA detection. This does not establish that the file is original or trustworthy.
Why an AI image may produce a negative result
A negative result may follow a sequence of file transformations. Screenshots, some sharing platforms, format conversions and editing software can strip C2PA metadata. Heavy cropping, recompression or substantial changes may weaken an embedded watermark until it is no longer recognized.
The documentation describes two further possibilities: the content may predate the relevant provenance signals, or it may come from a product, model or export method without compatible coverage. Images made by other companies also fall outside OpenAI Verify’s stated purpose. Missing signals therefore have several possible explanations, none of which can be selected based on absence alone.
Strictly speaking, a false negative occurs when independent ground truth establishes that a file belongs to the class a test is meant to identify, yet the test misses it. For an image of unknown origin, “no signal detected” is better described as inconclusive about AI generation: the verified ground truth needed to label it a detection error is missing.
A practical matrix for interpreting results
| Observed situation | Supported interpretation | What remains unproven |
|---|---|---|
| Supported OpenAI C2PA detected | A compatible provenance record attributed to OpenAI was found. | That the entire scene is true, remained unedited, or that its creator is known. |
| Supported SynthID detected | A watermark associated with supported OpenAI content was recognized. | How much was generated or edited, or whether the publishing context is correct. |
| No signal found | The uploaded file showed no recognized supported signals. | That it is a real photograph or was never processed by an AI generator. |
| C2PA missing or invalid | The credential is unavailable or cannot serve as valid provenance evidence. | That fraud occurred; transformation or incompatibility may also explain the status. |
| Image from another AI provider | A missing OpenAI Verify signal is consistent with the tool’s limited scope. | That the image is not synthetic or that another verifier could not detect its own signals. |
This matrix is an editorial interpretation of documented states, not the output of five executed tests. In particular, the API distinguishes C2PA manifest status, such as trusted, valid, invalid and not_present, from signal outcomes, detected and not_detected. Mixing the two can lead to incorrect attribution.
How to check an image responsibly
- Start with the original file. Prefer an export from the creator or a directly received file, without cropping, screenshots or conversion. If only a social-media copy exists, document that limitation.
- Upload one file at a time. Use OpenAI’s official verifier with a PNG, JPG or WEBP image. The webpage offers uploads; these instructions are documentation-based and do not imply hands-on testing.
- Check which signal was found. C2PA and SynthID convey different evidence. If issuer information is available, confirm whether it points to OpenAI rather than another participant in the C2PA standard.
- Treat a missing signal as missing evidence. Do not label the image human-made or genuine. Record the result as “provenance not confirmed by this tool”.
- Investigate context outside the verifier. Look for the source file, its date, the earliest known publication, other images from the same event, and independent confirmation. Those steps support reporting, but do not turn a provenance check into conclusive forensic examination.
A positive detection does not prove truth or authorship
OpenAI itself says a match indicates a supported signal, not whether an image depicts real events, has the correct caption, was later modified, or who created it. A file can contain synthetic material and circulate in a misleading context, or combine human-supplied material with automated processing. The tool does not quantify each party’s contribution.
C2PA attests to signed provenance statements, not the truth of a scene. Even technically valid credentials should be read in light of their issuer and recorded claims, rather than treated as complete historical, legal or journalistic proof. That distinction is especially relevant to allegations, fraud and claims that could harm someone.
What about the API, privacy and error rates?
Developers can also perform documented checks through POST /v1/content_provenance_checks. Its response can report C2PA and SynthID findings separately. An integration does not change the meaning of not_detected. OpenAI previewed the public tool on May 19, 2026, and announced support for eligible audio and API access on July 31, 2026.
Regarding uploads, OpenAI states that files are processed to find compatible signals, are not used to train its models, and are not retained except when legally required. This is the provider’s stated policy, not an independent audit of data handling.
For the scope of this article, the documents reviewed do not provide an independently validated overall false-negative rate across images from every generator, format and transformation. Inventing an accuracy percentage would be misleading. The supported conclusion is narrower: a recognized signal helps establish compatible provenance; no recognized signal leaves the origin unresolved.
