Imagine a dramatic photograph appearing in your feed: a flooded landmark, an unexpected product announcement, or a public figure in an unlikely situation. Before sharing it, you want to know where it came from. Looking closely might raise questions, but you still need evidence.
OpenAI Verify offers one starting point. It checks uploaded images and audio for supported OpenAI provenance signals, including C2PA credentials and SynthID watermarks. OpenAI Verify
The useful question is how this evidence should influence your next decision: publish, investigate further, or leave the claim unconfirmed.
What can OpenAI Verify check?
The service covers supported content from ChatGPT, Codex and the OpenAI API. Accepted image formats include PNG, JPG and WEBP; audio formats include MP3, WAV and several others. OpenAI Verify
For someone reviewing a suspicious post, this provides a specific lead to investigate. For an editor checking a submitted illustration, it offers another piece of evidence to discuss with the contributor.
How to check an image or audio file
- Open the verification page.
- Upload one supported file.
- Review which provenance signals, if any, were found.
OpenAI recommends avoiding image cropping or conversion before checking. For audio, it recommends clips lasting 10 to 60 seconds. These are usage recommendations, rather than a guarantee of detection. OpenAI’s instructions
For your own records, keep the file you checked alongside the post’s URL and caption. That makes it easier to distinguish evidence about the file from claims made by the person sharing it.
C2PA and SynthID: two ways to preserve provenance
Understanding the signals helps explain what a verification result can establish.
C2PA Content Credentials record information about a file’s origin and history using digitally signed data. Depending on what was recorded, they can describe the tools involved and subsequent editing steps. Cryptographic validation checks the relationship between that information and the asset. C2PA’s explanation
Think of them as a verifiable production record. Their usefulness depends on what the record actually contains: a valid entry should not be mistaken for a complete account of everything that ever happened to the image.
The C2PA standard also supports mechanisms that can help reconnect media with credentials after embedded metadata has been removed. Whether a particular viewer uses those mechanisms depends on its implementation. C2PA FAQ
SynthID, developed by Google DeepMind, embeds an imperceptible watermark within generated media. Google says its image watermark is designed to withstand transformations such as cropping, filters and lossy compression. This describes the technology’s intended resilience, rather than independent proof of OpenAI Verify’s performance on every edited file. Google DeepMind
The two approaches provide different kinds of evidence:
| Signal | What it contributes |
|---|---|
| C2PA Content Credentials | Signed information about recorded origin and editing history. |
| SynthID watermark | An embedded signal that a compatible detector can look for within the media. |
These descriptions follow the C2PA documentation and Google’s SynthID overview.
How to interpret the result
OpenAI’s documentation supports the following reading of the results. OpenAI Help Center
| Result | Reasonable interpretation | What remains unproven |
|---|---|---|
| A supported OpenAI signal is found | The file contains evidence associated with OpenAI generation or export. | Its factual accuracy, creator’s identity and later use. |
| No supported signal is found | The check found no usable OpenAI provenance signal. | Whether the file was made without AI. |
Missing signals can have several explanations: an unsupported model or export route, an older file, stripped metadata, or a watermark damaged by transformations. Audio may also be too short or heavily modified. OpenAI says the tool is not designed to identify content from other AI services. OpenAI’s coverage and limitations
Treat an inconclusive result as a reason to continue checking. It should never become a shortcut to labelling a disputed image authentic.
Why Google’s watermark does not mean universal coverage
SynthID’s presence in this workflow raises an understandable question: could the same check identify everything made with Google’s AI?
Google describes its own verification route through Gemini, where users can submit supported images, video or audio and ask whether Google AI generated or edited them. That is a separate service with its own scope. Google DeepMind
The practical implication is to check which origins a service actually supports. A shared underlying technology does not, by itself, establish that every verification interface checks the same content.
Three questions to ask before sharing
For everyday verification, DEMYSTIA recommends separating three questions:
- Where did the file come from? Look for provenance information, the earliest available publication and an identifiable source.
- What is the post claiming? Write down the alleged event, location and date. An image and its caption need separate scrutiny.
- What corroborates the claim? Seek an original announcement, another independent account or earlier appearances of the same image.
Consider two hypothetical cases. An openly labelled AI illustration accompanies a science explainer: identifying its production method says little about whether the explanation is accurate. Conversely, an ordinary photograph of a flooded street is reposted as evidence of a disaster somewhere else: the misleading element is the claim attached to it.
For publishers, the editorial opportunity is to preserve this distinction when explaining a result to readers. Say what evidence was found, what it supports and what remains unresolved. That is more informative than attaching a broad “real” or “fake” label to a file.
The decision to share should rest on the claim and its supporting evidence together.