GPT Image and ChatGPT image checker
Check if an image is AI-generated when it may have come from ChatGPT or the OpenAI image API. Newer ChatGPT images are GPT Image, not DALL·E 3.
- Files
- JPEG, PNG, WebP · up to 10 MB
- Model check
- General AI-image estimate, when available
- On this page
- ChatGPT and GPT Image exports
- Signatures
- Not verified by this site
Drop an image here
choose a file · paste with ⌘V / Ctrl V
How ChatGPT images leave the product
A file saved from ChatGPT, Codex or the OpenAI API is the one OpenAI’s provenance notes describe. A screenshot of the chat, a copy pasted into a doc, or an upload to a social app is a different file. OpenAI says Content Credentials can be removed by editing, conversion or sharing. The picture can still be theirs after that.
OpenAI’s help center says images from ChatGPT, Codex or the OpenAI API include both C2PA Content Credentials and a SynthID watermark, and that coverage can vary by product, model, export path, file type and date. The same page says API coverage excludes models it lists as planned for imminent deprecation, including gpt-image-1 and gpt-image-1.5. SynthID is in the pixels. openai.com/verify is the tool that looks for a SynthID watermark or a trusted C2PA manifest. This site has no SynthID decoder and does not verify signatures. Seeing the letters “c2pa” or “openai” in a header is an observation, not a validation.
DALL·E 3 is the older OpenAI image product. OpenAI has said Content Credentials started there in 2024, then continued on later image generation. If the file is a historical DALL·E export, use the DALL·E guide. If it is a current ChatGPT image, stay on this page. Do not call every ChatGPT image DALL·E.
The drawings are original illustrations of file paths. They are not outputs from GPT Image.
Metadata and Content Credentials
OpenAI documents two different signals. C2PA metadata can name the issuer, and on supported files can identify an OpenAI tool. It lives in the file container, so a re-save drops it. SynthID is a watermark in the media itself and is meant to survive some crops and compression. OpenAI also says a miss does not prove the image is real: the metadata may be gone, the watermark may be degraded, or the file may predate the signals.
Visible marks, when a product shows one, are a third thing. They are not the same as C2PA or SynthID, and they can be cropped. We do not promise that every GPT Image file has a visible mark. The authoritative check OpenAI offers is their own verify page, which this tool does not replace.
This site can notice source text in the first part of a file. It does not verify C2PA signatures, asset bindings or certificate trust, and it has no SynthID decoder.
OpenAI’s own verify tool
openai.com/verify accepts a supported image and reports whether it finds C2PA metadata, a SynthID watermark, or no supported signal. OpenAI says a hit means the file contains a provenance signal associated with OpenAI tools. It does not identify the person, the prompt, or whether the picture is true. A miss can mean the file predates the signals, came from an unsupported path, lost its metadata, or had the watermark degraded. Another company’s model is also possible.
Upload the original export there if the question is specifically OpenAI. Keep using this page when you want source text we can see and, when available, a general model score. The two results answer different questions and can disagree.
Visual tells, treated cautiously
GPT Image is discussed for following instructions and drawing text. Those strengths are not an identifier.
Readable text, still imperfect
Signs and labels fail less often than in older generators. Real photos contain readable text too, and GPT Image still misspells.
Instruction-shaped scenes
Diagrams, infographics and exact layouts are common because people ask for them. A designer can export the same layout.
No reliable skin tell
Portraits from current image models overlap. Do not use pores, eyes or lighting to name ChatGPT.
How this check handles the file
Source clues are checked in your browser first. When needed, file-header data and a smaller image copy may be sent to our backend and detection provider. Model copies have embedded metadata removed. Image URLs are retrieved by our server; see the privacy policy for processing and provider retention. Privacy details
Sources for the generator notes
Questions about GPT Image.
The answers below are about this generator’s files, not a promise to name the model.
Is this image from ChatGPT?
An original supported export may contain Content Credentials and a SynthID watermark. OpenAI’s verify tool is built to read those. We only report source text we can see and, when available, a general model score. A screenshot often has no metadata left.
Is GPT Image the same as DALL·E?
No. DALL·E 3 is the earlier product. Current ChatGPT image generation is GPT Image. Mixing the names hides which provenance note applies.
Does a missing watermark mean the photo is real?
No. OpenAI says a miss can mean the signal was stripped, degraded, or never added. Another company’s model is also possible. Missing evidence stays inconclusive.
Is the check free?
No account is required. Model checks depend on limits and availability.
Related reading
These notes explain what a file can and cannot show. They do not name GPT Image from the pixels.
Also inspect other generative models
Midjourney
Midjourney image checker for an original download or a later copy
Midjourney v7
Whether a file can be identified as Midjourney v7
Gemini
Gemini and Nano Banana image checker
Flux
Flux image checker across local and hosted exports
Flux.2
Flux.2 image checker for API and open-weight files
Grok Imagine
Grok Imagine image checker
DALL·E
DALL·E image checker for historical OpenAI exports
Stable Diffusion
Stable Diffusion image checker and PNG text reader
AI face images
AI face image checker for stills, not video
AI profile pictures
AI profile picture checker for avatars and headshots
Instagram photos
Instagram AI image checker for a saved photo
Reviewing profile photos or art entries?
Review LinkedIn candidate photos or competition artwork with context, and read our method and limitations.