This question sounds simple but turns out to have three separate parts: does the vendor attach provenance metadata at creation time, does that metadata survive normal sharing, and is there another signal such as watermarking when the metadata is gone?
As of May 11, 2026, the short answer is: sometimes yes at creation time, often maybe after distribution.
That distinction matters. People frequently inspect a reposted image and conclude too much from what is missing. The better question is not just “did the vendor add metadata?” but also “what path did this file take before it reached me?”
ChatGPT and OpenAI image outputs
OpenAI’s help article on C2PA in ChatGPT images says images generated with ChatGPT on the web and through the API serving the DALL·E 3 model include C2PA metadata. It also says people can use verification tools such as Content Credentials viewers to inspect those records.
OpenAI Help Center
That is the positive part.
The caveat is in the same article. OpenAI also says metadata like C2PA is not a silver bullet, can be removed intentionally or accidentally, and notes that many social platforms strip image metadata today. That means a file that no longer carries the record might still have originated from OpenAI’s image workflow.
So the right summary is:
- OpenAI says provenance metadata is attached at creation time for these flows.
- The absence of that metadata later does not, by itself, rule out OpenAI origin.
Adobe Firefly outputs
Adobe’s documentation is relatively specific here. The Firefly Content Credentials overview says Adobe automatically applies Content Credentials to assets where 100% of the pixels are generated with Adobe Firefly, such as Text to Image outputs.
Adobe Firefly Content Credentials overview
Adobe also says the record includes non-personal information such as:
- issuer
- issue date
- app or device used
- AI tool used
- general action information
Another notable detail is durability. Adobe says copies of Content Credentials for Firefly-generated assets may be stored in Adobe’s public Content Credentials cloud, where they can be recoverable with Adobe’s inspection tooling.
That does not guarantee recovery in every case, but it shows Adobe is treating provenance as something that may need support outside the file itself.
Gemini and Google AI outputs
Google’s public framing on the page we reviewed centers on SynthID rather than C2PA wording. DeepMind describes SynthID as an invisible digital watermarking system for AI-generated content, including images and video. It says the watermark is designed to be imperceptible and resilient to common modifications such as cropping, filters, and lossy compression.
Google DeepMind SynthID
The same page says users can ask Gemini to check whether an image, video, or audio clip was created or altered by Google AI by looking for a SynthID watermark.
So Google’s public story here is more watermark-forward in the source we used, while OpenAI and Adobe are more explicit about C2PA or Content Credentials. That does not mean Google has no provenance-related work elsewhere. It does mean you should avoid flattening all vendor approaches into one identical mechanism.
Why these answers are easy to misread
Here are the two most common mistakes:
Mistake 1: “The vendor adds metadata, so I will always see it.”
No. Distribution paths matter. Screenshots, exports, social platforms, and editing tools can all disrupt or strip attached records.
Mistake 2: “I do not see metadata, so the image cannot have come from that vendor.”
Also no. OpenAI explicitly warns that metadata can be removed, and Adobe’s own durability model suggests that in-file records alone are not enough.
Metadata, Content Credentials, and watermarking are not interchangeable
This is the part that keeps getting compressed in social media summaries.
- Metadata is the broad category.
- Content Credentials are a provenance-oriented, tamper-evident implementation within the C2PA ecosystem.
- Watermarking is a different mechanism that can help identify or reconnect content even when embedded metadata is not directly available.
That distinction is not academic. It changes what you can conclude from a positive or negative result.
What you should actually do with a file
If you are evaluating an image that might have come from one of these systems:
- preserve the original file if possible
- inspect for provenance or Content Credential information
- consider whether the file may have been reposted or screenshotted
- if Google-origin is plausible, understand that watermark-based checks are a different layer than metadata inspection
- if the stakes are high, combine origin checks with visual review and detector output
This is slower than “metadata yes/no,” but it is much more faithful to how these systems are documented.
What this can and cannot tell you
What it can tell you
- what OpenAI publicly says about C2PA in ChatGPT and DALL·E 3 image workflows
- what Adobe publicly says about Firefly Content Credentials
- what Google DeepMind publicly says about SynthID watermarking
- why later absence of metadata is not a clean negative signal
What it cannot tell you
- that every redistributed file will preserve those signals
- that watermarking and metadata are the same thing
- that one vendor’s approach can be generalized to all AI image systems
- that any one signal settles whether a visual claim is truthful
The realistic takeaway
Yes, major vendors now publicly describe provenance or watermark mechanisms for generated images. That is progress. But the internet’s distribution layer is still rough on file history.
So if you are trying to verify a suspicious image, think in terms of creation-time support versus what survived long enough for you to inspect.
If you want another layer after provenance checks, run the file through a local scan on the Detectiks home page.
Last reviewed
May 11, 2026.
