Do AI images have metadata? What each generator puts in the file, and what ties it to you
The Morse teamPublished 9 min read
Yes, most AI images have metadata as of November 2026. Images from ChatGPT, the Gemini app, Adobe Firefly, and Microsoft Paint carry a signed record called a C2PA manifest, or Content Credentials, that says an AI tool made them. ChatGPT and Gemini images also carry an invisible SynthID watermark in the pixels, Meta AI images carry IPTC metadata and an invisible watermark, and every Midjourney download hides an ID tag. None of these marks names you by default, though Midjourney's file names carry your display name, and Adobe lets you add your name and social accounts to the credentials on your Firefly images. Two of the marks, Midjourney's tag and an ID a researcher found in Microsoft Paint's images, point to a record on the company's servers, made from a signed-in account.
We read each company's own help and policy pages on November 16, 2026, along with the C2PA standard's explainer and a security researcher's August 2026 teardown of Microsoft Paint. OpenAI deserves credit for answering the privacy question in one sentence on its provenance page: "Provenance signals do not include details about the user, organization, or prompt" (OpenAI).
What metadata is in an AI image?
Usually a C2PA manifest, and sometimes a watermark you can't see. The C2PA explainer calls a manifest "a cryptographically bound structure that records an asset's provenance," made of statements about the file's origin, "modifications (i.e., what happened using what tools) and use of AI" (C2PA). It's signed, so an edit by a tool that doesn't update it shows up as tampering.
OpenAI's developer docs show what a check reads from one of its manifests: who issued it, which model generated the image, and when (OpenAI). Adobe lists the same kinds of fields for Firefly, plus a thumbnail of the output for images made with Text to Image on the Firefly website (Adobe, archived).
A watermark is a pattern woven into the pixels themselves, so there's no field to delete. Google says a SynthID watermark "will usually still exist even if the image, video, or audio is re-scaled, re-colored, compressed or altered in other ways" (Google).
Does ChatGPT add metadata to images?
Yes, two kinds. OpenAI's help center says images "generated with ChatGPT, Codex, or the OpenAI API include both" Content Credentials and SynthID watermarks (OpenAI). OpenAI began adding Content Credentials in 2024, with DALL-E 3, and later added SynthID through a partnership with Google (OpenAI).
OpenAI's API output is covered "except for models planned for imminent deprecation, including gpt-image-1, gpt-image-1.5, and sora-2," so an app still running one of those older models may hand you an unmarked file (OpenAI).
Neither mark names you, by OpenAI's account. Its help page says provenance signals carry no details about "the user, organization, or prompt," and that its checker "does not identify who created the content or why it was created" (OpenAI).
What does each AI image generator put in the file?
Read on November 16, 2026, on each company's own pages, except where a cell says otherwise. "Not described" means the pages we read don't say.
| Generator | Visible label | C2PA manifest | Invisible watermark | Anything tied to your account | Source |
|---|---|---|---|---|---|
| ChatGPT and the OpenAI API | Not described | Yes | SynthID | None, says OpenAI: no user, organization, or prompt details | OpenAI |
| Gemini app | Yes, unless you switch it off | Yes, signed "Info by Google LLC" | SynthID, always | Not described | |
| Adobe Firefly | Not described | Yes, for fully generated images, and may be copied to Adobe's public cloud | Not described | Your name and social accounts, only if you add them | Adobe, archived |
| Meta AI | Visible markers on photorealistic images | Not described; Meta uses IPTC metadata | Yes | Not described | Meta |
| Midjourney | Not described | Not described; a hidden "ID tag" points to a real Midjourney image | Not described | Your display name, in the download's file name | Midjourney |
| Microsoft Paint and Photos | Optional Copilot logo | Yes | A per-image ID, per a researcher's analysis | Per the researcher, that ID is issued by Microsoft's server, which gets your prompt from your signed-in account | Microsoft, Xusheng Li |
| Grok Imagine | Not described | Not described | Not described | Not described | SpaceXSI |
| Claude (files it creates) | Not described | Yes, on supported files such as PNG or JPEG | Not described for files | Not described | Anthropic |
Gemini's visible mark is a setting. Open Settings, then Media Watermark, and switch it off. In India, South Korea, and Vietnam the switch appears only with an AI Ultra subscription. Google says the setting "doesn't affect SynthID watermarks or Content Credentials" (Google).
Meta's newer help page puts it more loosely. Content made or edited with Meta's AI tools and shared to Facebook, Instagram, or Threads "may be automatically labeled as AI content or in some cases, feature a visible watermark" (Meta). Meta's 2024 announcement called the label "Imagined with AI." The current help page calls it "AI content."
Grok is made by SpaceX Super Intelligence LLC, or SpaceXSI, according to its consumer FAQ (SpaceXSI). SpaceXSI's pages don't describe any mark on Grok images, but its acceptable use policy, effective August 14, 2026, bans "Stripping, altering or circumventing embedded provenance metadata or watermarks" (SpaceXSI).
Claude makes files rather than pictures from a prompt, but Anthropic's help page says that when Claude generates "a supported file type such as a PNG or JPEG, it will attach signed provenance metadata," which its free Claude Content Checker can read (Anthropic). Claude's text gets a separate watermark, covered in Claude's text watermark.
Can AI images be traced back to you?
Not from the C2PA standard alone, but an ID in the file can lead back to an account. The C2PA explainer says the core specification "does not directly address the topic of human or organizational identity," which "ensures respect for user privacy." It adds that extensions can put identity into a credential, and warns that "users should be aware of the privacy implications" (C2PA).
Adobe is the clearest case of opting in. Its Firefly credentials hold only "non-personal information" unless you've added more through Photoshop, Lightroom, or Adobe Content Authenticity, where you can "Add your name and social media accounts to verify your identity" (Adobe, archived). LinkedIn then shows a credential to everyone who sees the post, and its list of fields includes "Created by – The individual who generated the image" (LinkedIn). If you added your name to your Content Credentials, a LinkedIn post can show it.
The other route is an ID that means nothing on its own but matches a record. Midjourney says the tag in every image "points to a real Midjourney image," and its Profiles article says your gallery lives in your account, with an Archive tab to "Browse all of your public creations." The same article says your display name is used "in the file names of your downloaded images and videos" (Midjourney). We cover the rest of Midjourney's settings in is Midjourney private, and set its tag beside Adobe's Content Credentials in our Midjourney alternatives comparison.
What did a researcher find in Microsoft Paint's AI images?
An ID in the pixels, issued by Microsoft's server before the image is made. On August 20, 2026, reverse engineer Xusheng Li published an analysis of Paint and Photos on Windows. Microsoft hasn't confirmed it on the pages we read, so treat what follows as one researcher's findings.
Li found that even when Paint makes an image on the PC's own chip, it first sends your prompt to a Microsoft moderation server. The server sends back a "watermarkId," a 16-byte ID unique to that request, and Paint writes it into the image as an invisible watermark. In a PNG he saved, the signed C2PA manifest held the same value under a soft binding called "com.microsoft.invismark.1." Paint's visible-watermark setting, he wrote, "does not control this invisible watermark." Photos' Image Creator and Restyle Image features call the same watermark code.
Microsoft's own support page covers part of this. It says images from Image Creator, "or any other AI feature in Paint, will contain C2PA manifest," that you have to sign in with a Microsoft account, and that "Microsoft collects attributes such as device and user identifiers, along with the user prompts, to facilitate abuse prevention and monitoring." Li found no Microsoft page explaining the ID in the pixels (Microsoft).
Does a screenshot remove AI metadata?
It removes the C2PA manifest, but often not the watermark. OpenAI says Content Credentials "can be lost through everyday actions, such as taking a screenshot or converting a file," while SynthID is designed to withstand "light edits, copying and pasting, cropping, or screenshots." Heavy cropping and file compression can still make a watermark undetectable (OpenAI).
A stripped manifest can sometimes be found again. The C2PA explainer says soft bindings, made with "invisible watermarking or fingerprint lookup," let a credential "be discovered in cloud-hosted storage even if it is removed from the asset" (C2PA). Adobe says Firefly credentials "may be stored in Adobe's public Content Credentials cloud," where its Inspect tool can recover them (Adobe, archived).
Do social media sites strip AI metadata?
Many do, and some read it first. Midjourney says "Screenshots, edits, and most social media sites remove" its tag (Midjourney), and OpenAI says "platforms, editing tools, and file conversions may remove" metadata (OpenAI).
Meta reads the signals when you upload. Its help page says content with signals showing it was created with AI "will be labeled AI content," while content only modified with AI gets the note in the post's menu instead (Meta). LinkedIn reads credentials and marks any image or video signed with them with a C2PA icon that every viewer can click (LinkedIn).
Who can check an image for these marks?
Mostly anyone, with a few limits:
- OpenAI's verify page at openai.com/verify checks images and audio for OpenAI's signals. OpenAI says uploads are "not stored unless legally required" and aren't used for training (OpenAI).
- Gemini checks for SynthID and Content Credentials if you're signed in, with a quota of "approximately 10 image checks" in any 24 hours. Google says Gemini can only recognize content made by Google's AI tools (Google).
- Google's SynthID Detector checks for content "made with AI from Google or our partners, including OpenAI, NVIDIA, Kakao, and soon, Apple" (Google DeepMind).
- Midjourney's verify page is free, needs no account, and Midjourney says "we don't keep your image" (Midjourney).
OpenAI says its checker doesn't identify who made an image, and none of the other checkers' pages says theirs does. Who made it is in the generator's own records, under an account.
Where Morse keeps your chats and images
Morse keeps your chat history in your browser, encrypted, not on Morse's servers, and Studio, Morse's image generator, keeps the images you make in your browser, encrypted.
Your settings are stored the same way, and all of it is locked when you sign out. Your Morse account holds your sign-in details, plan, balance, and a usage ledger with no content. Morse doesn't train AI on user content, sell your data, or show ads (privacy policy), and because we don't store chats, we usually have very little to give authorities.
If you work out a prompt in chat first, that chat can hold the person, place, or idea behind the picture. Pick the mode for it:
- Private, to work out an image idea with hosts that have agreed to keep nothing. It's the default for a new chat. Open-weight models like DeepSeek, Kimi, and GLM, plus Grok, run here, and Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
- Anonymous, to ask GPT, Claude, or Gemini for prompt ideas without the provider knowing it's you. Morse's proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider sees the message without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy. Leave real names out of the text.
- E2EE, for writing you want only a verified enclave to read. Your browser checks the secure enclave's proof that it's genuine, then encrypts the message so only that enclave can open it, and Morse's servers relay text they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE has no web search or image generation, and it comes with Pro, at $19 a month.
Clearing your browser deletes your chats and Studio images, and Morse can't recover them, so keep a copy of any chat you need. For what ChatGPT does with the metadata in photos you upload, see how ChatGPT knows your location, and for how other image apps handle what you make, our ranking of private AI image generators.
Start free and make 3 Studio images a day on Z-Image Turbo, with no card needed.
Common questions
Does removing the metadata make an AI image undetectable?
Not for ChatGPT or Gemini images. Both also carry a SynthID watermark in the pixels, which OpenAI says is designed to survive light edits, cropping, and screenshots, though heavy cropping or compression can defeat it.
If I turn off Gemini's visible watermark, is the image still marked?
Yes. Google's help page says the Media Watermark setting controls only the visible mark and doesn't affect the SynthID watermark or the Content Credentials in the file.
Is it against the rules to strip AI metadata?
On Grok it is. The acceptable use policy of SpaceXSI, the company behind Grok, effective August 14, 2026, bans stripping, altering, or circumventing embedded provenance metadata or watermarks.
Does checking an image on OpenAI's verify page save it?
OpenAI says uploads to its verify page aren't stored unless the law requires it and aren't used for training.