What OpenAI documents
OpenAI says images generated with ChatGPT, the OpenAI API and Codex carry provenance signals: C2PA Content Credentials, which are signed metadata attached to the file, and a SynthID watermark, an invisible signal embedded in the image itself through a partnership with Google. OpenAI also previews a public verifier at openai.com/verify that checks a supported upload for a SynthID watermark or a trusted OpenAI C2PA manifest. Coverage can vary by product, export route, file format and creation date, so read OpenAI's current guidance before relying on either signal.
Start with the original export
Ask for the file exactly as it was downloaded from ChatGPT, not a screenshot or a copy pasted into a chat app. A screenshot records what was on the screen and drops the original file structure, including any Content Credentials. Note how the file reached you and whether it was cropped, edited or re-exported. Each of those steps can remove metadata, so a missing manifest on a re-shared copy says very little about where the image started.
Read the signals separately
A trusted C2PA manifest, a SynthID result, an ordinary metadata field and a detector score are different kinds of evidence. Plain-text fields such as a Software tag can be typed by anyone, so a field mentioning OpenAI deserves follow-up rather than belief. A signed manifest is stronger only after its signature and signer have been validated. A classifier score is an estimate whose reliability depends on how the classifier was tested. Look for signals that agree, and write down the ones that conflict.
Generated and edited are different
A photograph edited with a generative tool can contain real captured material alongside synthetic additions, so the question “Was this made with AI?” may not have a simple yes-or-no answer. Ask the creator what changed and whether an unedited capture exists. Do not use a detector label to accuse a student, reject a portfolio or identify a person. Image Evidence performs no face recognition and does not determine identity.
What to do next
For an image that matters, compare the supplied original with the claimed source and ask for its creation history. If the file may be an OpenAI export, check it with OpenAI's verifier after reviewing its terms, and keep a written distinction between a signal that was found and a claim that was verified. Image Evidence has not published a ChatGPT-specific benchmark; the accuracy page explains what evidence a performance claim would need.
Can you tell if an image was made with ChatGPT?
Sometimes. If the original file still carries OpenAI's Content Credentials or SynthID watermark, a supported verifier can report it. Many copies lose those signals: screenshots, crops and re-uploads to social platforms can strip metadata, and heavy edits may weaken a watermark. Visual cues such as unusually smooth textures or odd lettering are not reliable on their own, because current models produce convincing detail and real photos can look strange.
Does ChatGPT add a watermark to images?
OpenAI says supported images from ChatGPT, its API and Codex include C2PA metadata and an invisible SynthID watermark. Neither is a visible logo. The C2PA manifest lives in the file and can be lost when the image is re-saved or screenshotted. The SynthID signal is embedded in the pixels and is designed to survive some common edits, but no watermark survives every transformation.
Is there a free ChatGPT image detector?
OpenAI previews a verification tool at openai.com/verify for checking supported files for its provenance signals; review its current availability and terms before uploading. General-purpose AI image detectors estimate likelihood rather than reading OpenAI's signals, and their accuracy depends on their test data. The Image Evidence checker is in development and does not connect to OpenAI's verifier.