OpenAI Image Verification: What OpenAI's Checker Can Confirm

OpenAI image verification is OpenAI's own check for content made with its tools. You can upload a file at openai.com/verify in your browser, and developers can send images and audio to OpenAI's Content Provenance API. Either way it looks for two signals OpenAI adds to content from its tools (not all of it carries them):

• C2PA Content Credentials. Signed metadata that names the issuer, says AI was used and, when available, the model and the time the content was made. Editing, converting or sharing a file can remove it.
• SynthID. A watermark embedded in the image or audio itself. OpenAI says it may survive some transformations, even when the metadata is gone.

Each signal gets its own result, detected or not detected; in the API there is no overall verdict, so read the two separately. The C2PA result only counts as detected when a trusted or valid manifest names OpenAI as its issuer and records an AI-generation action. A manifest from another company, an invalid one, or none at all comes back as not detected. In other words, OpenAI's verification answers one question: does this file carry OpenAI's own signals? If OpenAI's API asked you to verify your organization before using its image models, that is a separate account check, not this tool.

Upload the original file whenever you can. OpenAI's own advice is to verify the original, because compression, cropping, screenshots, metadata removal and format conversions can erase or weaken a signal. For the API, OpenAI lists PNG, JPEG and WebP images and MP3, Opus, AAC, FLAC, WAV and PCM audio, one file per request, up to 50 MiB, with audio no longer than 60 seconds. Text is a separate check, open only to approved organizations such as AI research and academic institutions, and video isn't on the list.

Screenshots are where people get stuck. A screenshot is a new file, so the original's Content Credentials never come along: none of the 10 screenshots we made of AI images carried any. SynthID is the signal that still has a chance, because it sits in the pixels, but OpenAI names screenshots among the things that can weaken it. So a not detected result on a screenshot tells you very little either way.

Not detected means the checker didn't find OpenAI's signals. OpenAI is explicit that this isn't proof a person made the picture, and lists the reasons content made with its tools can still come back clean: its metadata was stripped or shows signs of tampering, its watermark was degraded, it came from a legacy model, or it was made before provenance signals existed.

It can't speak for anyone else either. OpenAI says the tool doesn't currently detect content from another company's AI model and isn't a general-purpose AI detector. In our own 100-image test, 40 of the 50 AI images came from other companies' generators, from Google's Nano Banana to Midjourney. We didn't run them through OpenAI's tool, but by OpenAI's own description it detects none of them, including the 16 originals signed with their makers' own AI credentials and the 12 Midjourney and Stable Diffusion XL originals that carried no credentials at all.

Each company's checker confirms its own tools. The table above lines them up: OpenAI's for OpenAI content, Gemini's for Google's. Gemini reads Google's SynthID watermark, shows a summary of Content Credentials, and also takes short videos.

A detector like the one on this page answers a different question: does this image look AI-generated, whoever made it? It reads Content Credentials from any signer on the C2PA trust lists first, which settled all 8 of our OpenAI originals without calling a model, then sends everything else to a detection model that judges the pixels. That gives you a likelihood with reasons, not a signature. The checks work best together:

• You suspect ChatGPT or another OpenAI tool. Run OpenAI's verifier. A SynthID detection there is evidence we can't give you, because our checker doesn't read SynthID.
• You don't know which tool made it, or OpenAI says not detected. Check the image here and read the reasons.
• It looks like a Gemini or Nano Banana image. Ask Gemini, which reads Google's SynthID (more on Nano Banana images).

For pictures from ChatGPT in particular, the ChatGPT image detector page shows what we found in 10 GPT images, screenshots included.

A detected result is evidence of one signal, not the file's full history. OpenAI tells developers not to infer a prompt, an account or an individual creator from a verification result, and to pair automated decisions with human review when the stakes are high. Visible watermarks and labels are separate from the signals it checks, so a logo in the corner is neither confirmed nor ruled out by it. If you don't know where a picture came from at all, start with is this real or AI, which covers what each answer means and how to handle screenshots and videos.

Not in the usual sense. OpenAI says it isn't a general-purpose AI detector: it looks only for OpenAI's own Content Credentials and the SynthID watermark in OpenAI content, and doesn't detect images from other companies' models. A detector judges whether any image looks AI-generated.

You can upload one, but the original's credentials won't be in it, and OpenAI says screenshots can weaken the watermark too. A not detected answer on a screenshot means little. Find the original file if you can, or check the screenshot here, where a detection model judges the pixels.

OpenAI's documentation covers images and audio. Text checks are open only to approved organizations, and video isn't listed. For a short clip made with Google AI, the Gemini app can check videos for SynthID.

No. OpenAI says it doesn't currently detect content from other companies' models. None of the Midjourney files in our test carried Content Credentials (a few had an unsigned AI label anyone could write), so for its pictures a detection model does the real work; see the Midjourney detector.

They describe the Content Credentials in the file. A trusted or valid manifest that names OpenAI and records an AI-generation action counts as detected. OpenAI says not to treat an invalid manifest as reliable evidence, and not present means the image has no available C2PA manifest. The issuer can still be shown when the outcome is not detected.