DALL-E image watermark detector

Three eras, three different signals to check for.

DALL-E's watermark history spans a visible signature, a ChatGPT-era corner mark, and C2PA metadata. Check a file for all three signals right here.

of DALL-E watermark history: 3 eras
current metadata standard: C2PA

DALL-E detector

What the DALL-E image watermark detector does

DALL-E has carried different kinds of evidence at different points in its life as a product: a visible signature on the standalone DALL-E 2 tool, a corner mark on some images generated through ChatGPT, and a C2PA metadata manifest on more recent output. This page checks a file for all three and tells you which one it finds.

Drop your image into the box on this page and our media workflow opens to run the check. It reads for a visible mark and for an embedded manifest, and reports both in the same session.

The problem

Checking a DALL-E image for a watermark, in plain terms

The standalone DALL-E 2 product added a small multicolor signature to the corner of its output as a matter of course, a visible feature of that specific tool rather than something buried. Once DALL-E's image generation moved inside ChatGPT, that particular signature wasn't the default anymore, and the evidence shifted toward a corner mark on some releases and, more consistently, a signed metadata manifest.

So the honest answer to "does DALL-E watermark its images" depends heavily on which DALL-E and which era you mean. A file from the original standalone tool and a file from ChatGPT's image generation today can carry entirely different kinds of evidence.

How it works

How to check a DALL-E image for a watermark in three steps

What you check has to be the actual file, not a description of what DALL-E usually does.

Step 1: Get the original download, not a screenshot. Only a download can carry a C2PA manifest, and it preserves a visible signature or corner mark at full quality if one is present.
Step 2: Drop it into the box above. Our media workflow reads for any manifest and confirms any visible mark.
Step 3: Weigh the result against when the file was made. A pre-ChatGPT DALL-E 2 image, a ChatGPT-era image, and a current image can each carry a different combination of evidence.

Reading the result

What a missing watermark actually tells you

A visible signature or a verified metadata manifest is solid evidence of a specific tool and, often, a specific era of it. It's hard to add either of those convincingly to a file after the fact.

An empty result tells you much less. Cropping removes a visible signature outright, and metadata is stripped by screenshots, re-saves and most social platforms as routine handling. Treat an absent result as unresolved, not as proof the image never touched DALL-E.

Why it's like this

Why DALL-E's evidence has changed so much over time

DALL-E started as a standalone consumer product with a distinctive visible signature, then became one generation path inside ChatGPT, where OpenAI's broader, evolving approach to image marking took over. Each stage reflects a different product and a different set of choices, not one fixed policy applied throughout.

That's also a different kind of signal from an in-pixel scheme like Google's SynthID, which is a statistical mark in the pixels rather than a visible logo or a metadata record. DALL-E's evidence, across every era, has lived in the places you can see or the places you can check with a reader, not somewhere invisible to both.

Straight answer

Does this prove an image came from DALL-E?

A verified signature or manifest is strong evidence for that specific file. It says nothing conclusive about every DALL-E image lacking one, since cropping and ordinary sharing routinely remove both kinds of evidence without anyone intending to.

Use a check result as one input alongside where the file came from and what era it's likely from, not as a single answer that settles the matter.

Who uses it

Who checks a DALL-E image for a watermark, and when

A handful of recurring, ordinary situations.

Newsrooms and fact-checkers: Establishing whether a circulating image shows signs of DALL-E before it runs anywhere, and roughly when it was made.
Platform moderation: Reading a visible signature or C2PA metadata as one signal within a broader synthetic media policy.
Stock and licensing review: Confirming whether a submitted file declares a generative tool chain.
Anyone curious about one image: Checking a single file out of plain curiosity about which era of DALL-E made it.

FAQ

Questions about the DALL-E image watermark detector

Everything below is answered in full, nothing to click open.

Does DALL-E put a watermark on its images?

It depends on the era. The original standalone DALL-E 2 tool added a visible signature by default; images generated through ChatGPT have carried a corner mark on some releases and, more consistently, C2PA metadata.

How do I check if an image came from DALL-E?

Drop the file into the box on this page. Our media workflow reads it and reports what it finds.

Is my image uploaded when I run a check?

Yes, to our own media workflow, so it can be read. It isn't published or shared.

What did the original DALL-E watermark look like?

The standalone DALL-E 2 tool added a small multicolor square signature to a corner of the image, a visible and deliberate feature of that specific product.

Is that the same as the ChatGPT image watermark?

Not necessarily. Once DALL-E moved inside ChatGPT, the watermarking approach shifted toward OpenAI's broader, evolving policy, which has included a corner mark on some releases and metadata more consistently.

Does a screenshot carry the same evidence as a download?

No. A download can carry a C2PA manifest and preserves a visible mark at full quality. A screenshot carries neither reliably.

Does cropping remove the evidence?

Cropping out a visible signature or mark removes that specific evidence. It has no effect on metadata, which lives in the file's data rather than the visible picture.

Can the metadata be faked?

The signature behind a C2PA manifest is cryptographic, so a convincing forgery is difficult to produce.

Is this tool affiliated with OpenAI?

No. This is an independent detector, not made by, endorsed by, or connected to OpenAI.

What's the difference between this and the ChatGPT image detector?

They overlap closely and often point at the same underlying evidence. This page focuses on DALL-E's own history, including the standalone tool's signature, which predates ChatGPT's image generation.

Is there a limit on how many images I can check?

The checker handles ordinary photo sizes without trouble. A very large file just takes a moment longer.

Does this detector check video?

No. This route covers images. Video watermark checks have their own route on this network.

If there's no visible mark and no metadata, does that mean the image is real photography?

Not reliably. Both signals can be missing through ordinary handling rather than because the file was never made with DALL-E.

Can I check an image I don't have the rights to use?

Checking provenance is separate from reuse. Make sure you have the rights to whatever you actually do with the file.

Is the checker free?

Yes. There's no charge to run a check through this page.

Does it work on phones?

Yes, the upload control and the check both work in a phone browser.

Does this predict what a third-party AI detector would say?

No. This page reports what's actually in the file, not what a separate detection service might conclude.

How do I know which era my DALL-E image is from?

Look at how you generated it. Anything from the standalone dalle.com tool predates ChatGPT's integration; anything generated inside ChatGPT is the later era.

Does this cover Bing Image Creator too, since it also uses DALL-E?

No, Bing runs its own separate deployment with its own undocumented behavior. That has its own dedicated route on this network.

Check your DALL-E image right here.

Drop the file into the box above. Our media workflow reads it and hands back what it finds.

Check my image