Images from ChatGPT and the OpenAI API carry a C2PA provenance record signed by OpenAI and, since May 2026, the invisible SynthID watermark. The checker above shows the C2PA record, while only OpenAI can check the watermark, at openai.com/verify.
If a ChatGPT image in your store is a deep fake, Article 50(4) of the EU AI Act still requires a visible label (Commission guidelines para. 117).
What ChatGPT writes into the file
This applies to images from ChatGPT, Codex and the OpenAI API, including the gpt-image and DALL·E 3 models.
| Trace | Status October 2026 | Source |
|---|---|---|
| C2PA (Content Credentials) | Yes Signed record with issuer and AI-use details, in ChatGPT and the API since February 2024. The C2PA list of conforming products includes the “OpenAI Media Service” since 28 January 2026. | ↗ ↗ C2PA |
| IPTC Digital Source Type (XMP) | Not documented OpenAI documents no separate IPTC field outside the C2PA record. | none found |
| Invisible watermark | Yes SynthID from Google DeepMind, since May 2026 in images from ChatGPT, Codex and the API. Can be checked at openai.com/verify. | ↗ ↗ ↗ |
| Prompt and settings in the file | Not documented OpenAI documents no prompt fields in the file. | none found |
| Visible mark on the image | Not documented OpenAI describes no visible watermark for images. | none found |
C2PA (Content Credentials) is a signed provenance record in the file, IPTC Digital Source Type a metadata field on how the image was made, for example trainedAlgorithmicMedia for fully AI-generated images. An invisible watermark sits in the pixels, and only the vendor can read it.
How to check a ChatGPT image
- Drag the original file from ChatGPT into the box above. It is read in your browser, not uploaded. Screenshots and copies from stores or social networks usually no longer carry the data.
- A manifest then appears under “Content Credentials (C2PA)”. The “Certificate (organisation)” row names the issuer, which OpenAI gives as “OpenAI OpCo, LLC”.
- Images from Microsoft Azure carry a record issued by Microsoft, with “Azure OpenAI ImageGen” or “Azure OpenAI DALL-E” as the program.
- Without a record, SynthID may still be in the image. To check it, upload the image at openai.com/verify.
More detail in the Content Credentials viewer and AI image checker.
The checker is not an AI detector. It only shows what the file says. An empty result does not prove a real photo, not least because providers of AI systems placed on the market before 2 August 2026 have until 2 December 2026 to add machine-readable marking (Article 111(4) AI Act).
What survives uploads and edits
According to OpenAI, editing, converting or sharing a file can remove its metadata. SynthID, by contrast, is part of the image and may survive some changes.
A program without C2PA support drops the record or breaks its signature. So check the download before you crop it or cut it out.
Labelling duty in your store under the EU AI Act
If you use ChatGPT to create or edit images for your own store, you are a deployer, even when an agency or an employee works on your instructions (Article 3(4) AI Act, Commission guidelines para. 12, 14). If the image is a deep fake, meaning realistic and falsely appearing authentic, you have had to label it visibly since 2 August 2026 (Article 50(4) AI Act). The labelling check tells you whether that applies, and the guide to the labelling duty explains the rules.
The traces in the table do not replace that label, because people cannot see them straight away (guidelines para. 117). Still, keep them in the file, as the guidelines recommend. They document where the image came from.
Typical store uses are lifestyle shots with your own product, AI models and product photos with a new background. The need for a label depends on what the image shows, not on the tool.
How ChatGPT labels its own output
For checking, OpenAI offers the openai.com/verify page and a developer API. Both only detect signals from OpenAI's own tools.
According to OpenAI, a “not detected” result does not rule out that the image came from OpenAI, for example if the metadata was stripped, the watermark degraded or the image came from an older model.
Frequently asked questions
Does every ChatGPT image say it was made with AI?
The original file does, through a C2PA record since February 2024 and SynthID since May 2026. According to reports at the time, older images carry no record, and after a screenshot or upload it is often gone.
Why does the checker find nothing in my ChatGPT image?
Usually the file has been changed by a screenshot, crop, cut-out, compression or upload. SynthID may still be in the image, but only OpenAI can check it.
Do I have to label a ChatGPT image in my store?
If it is a deep fake, yes, visibly on the image (Article 50(4) AI Act). A clearly fantastical motif is usually not a deep fake according to the guidelines (para. 116).
Does this also apply to DALL·E images?
Yes. DALL·E 3 ran in ChatGPT and the API, and both have added the C2PA record since February 2024.
Sources
Para. refers to the paragraph number in the European Commission's guidelines on Article 50, C(2026) 5054.
- OpenAI: Advancing content provenanceOpenAI · Platform · 19 May 2026
- OpenAI Help: C2PA und SynthID in OpenAI-BildernOpenAI · Platform · 6 Feb 2024
- OpenAI: Bildprüfung (verify)OpenAI · Platform · 19 May 2026
- Google DeepMind: SynthIDGoogle DeepMind · Platform · 11 Oct 2026
- Microsoft Learn: Content Credentials in Azure OpenAIMicrosoft · Platform · 11 Oct 2026
- C2PA: Verzeichnis konformer ProdukteC2PA · Standard · 11 Oct 2026
- Regulation (EU) 2024/1689 (AI Act)EUR-Lex · Law · 12 Jul 2024
- Commission Guidelines on the transparency obligations under Article 50 AI Act, C(2026) 5054 finalEuropäische Kommission · Commission · 20 Jul 2026
- Regulation (EU) 2026/1744 (Digital Omnibus on AI)EUR-Lex · Law · 24 Jul 2026
