DALL-E image watermark detector

DALL-E's evidence has taken three distinct forms across its life as a product: a visible signature on the original standalone tool, a corner mark on some images generated through ChatGPT, and a signed C2PA metadata manifest more recently. Treating these as one fixed watermark rather than three separate historical signals is the most common error in reading a result.

This detector runs inside our own media workflow. The upload control on this page opens it: the file is checked against all three signals and the result reported back in the same session, alongside the era-by-era context below.

The standalone dalle.com product added a small multicolor signature to a corner as a default, visible feature. Once image generation moved inside ChatGPT, that specific signature was no longer the default behavior, and the applicable evidence became whatever OpenAI's broader, evolving image-marking policy was doing at the time.

A precise answer to whether a given DALL-E file is marked therefore requires knowing which product generated it and roughly when, not a single rule applied across the entire history of the name.

The result depends on matching the right signal to the right era of file.

Step 1: source the original file

A download, not a screenshot. Only a download reliably carries a C2PA manifest and preserves a visible signature or mark at full fidelity.

Step 2: run the read above

Our media workflow parses the manifest and confirms any visible mark in the same pass.

Step 3: attribute the result to an era

A pre-ChatGPT DALL-E 2 file, a ChatGPT-era file, and a current file carry meaningfully different evidence, and the result should be read against the right one.

A verified visible signature or metadata manifest is meaningfully strong evidence, since neither is trivial to add convincingly after generation. Either one narrows down the tool and, often, the era with real confidence.

An absent result carries far less weight. Cropping removes a visible signature outright, and metadata is a routine casualty of screenshots, re-encoding, and most social platforms. Read absence as unresolved, not as evidence against DALL-E involvement.

The asymmetry matters most here because DALL-E's history spans three eras: a negative result doesn't just mean no manifest, it could mean any of several different explanations depending on when the file was actually made.

DALL-E began as a standalone consumer product with a distinctive visible signature, a straightforward, if inflexible, approach. Its later life inside ChatGPT inherited OpenAI's broader and changing policy on image marking, which has moved between a visible mark, metadata, and both, across releases.

That's a different category of engineering choice from Google's SynthID, an in-pixel statistical mark built to resist crop, resize and re-encode by design. Across every era of its own history, DALL-E's evidence has sat somewhere visible or somewhere a metadata reader can check, never somewhere invisible to both.

For the specific file checked, a verified signature or manifest is good evidence. It doesn't generalize to every DALL-E image lacking one, since absence is exactly what cropping and ordinary sharing produce regardless of origin.

The honest use of this detector is as one input into a broader judgment about a file's likely origin and era, never a standalone verdict.

The recurring cases share a need to place a file in its correct era, not just confirm a tool name.

Editorial verification

Establishing whether a circulating image shows DALL-E's signals and roughly which era, before it runs in a story.

Platform policy enforcement

Reading a visible signature or C2PA metadata as one input into a broader synthetic media policy.

Licensing and stock review

Confirming whether a submitted asset declares a generative tool chain.

Legal and compliance review

Documenting a file's own claimed provenance as one part of a larger record.