What the answer means
Every result on this page opens with an answer about OpenAI markers: a field whose job is to say which tool made or signed the file (Content Credentials’ generator and signer, the Software or Creator tool field, generator data, the IPTC AI system field) and that names ChatGPT, a GPT image model, DALL·E or OpenAI. Under it you get the same evidence report as the AI image detector, with every signal and every check listed.
- ✓ OpenAI marker found: the file names an OpenAI tool as its maker or signer. Signed Content Credentials are the strong case: their signature is checked too, and the Content Credentials checker shows the full history. A Software field is unsigned, and the box says so.
- Another generator named: the file points to a different tool, such as Stable Diffusion, and the box names it.
- A mention only: a caption or keyword mentions ChatGPT. That is not counted as a marker.
- Only a weak hint: no marker, but the image has a size ChatGPT and other generators often output, such as 1024 × 1024 or 1536 × 1024. Plenty of ordinary pictures share those sizes, so this alone says little.
- No marker: nothing in the metadata points to OpenAI. Read the next section before you conclude anything.
Two signals, and which one this page can read
OpenAI’s help center, “Provenance signals in OpenAI-generated content” (read October 25, 2026), says images generated with ChatGPT, Codex or the OpenAI API include both:
- Content Credentials (C2PA): metadata that records which tool created the file and when. It is detailed, but OpenAI notes it “can be lost through everyday actions, such as taking a screenshot or converting a file.” This page reads it.
- SynthID: an invisible watermark from Google DeepMind, embedded in the pixels. OpenAI says it is designed to survive light edits, cropping and screenshots, while heavy cropping or compression can make it undetectable. This page can’t read it: only OpenAI’s checker at openai.com/verify and its Content Provenance API can.
So the order that makes sense: check here first, privately. If there’s no marker and the question matters, upload the original file to openai.com/verify, keeping in mind that it shares the image with OpenAI (OpenAI says files checked there are only processed to look for its signals, aren’t stored unless legally required, and aren’t used to train its models). OpenAI recommends the original exported file, not a cropped or converted copy.
When neither finds anything
Neither check can prove an image is not from ChatGPT. Images made before the signals were added, from an unsupported model or export path, or passed through apps that strip metadata and recompress pixels can carry no trace at all. For a viral picture, a reverse image search for the earliest copy often settles more than any detector. Images from Google’s Gemini carry SynthID too; the SynthID checker walks through Google’s own check.