Removing DALL-E Metadata: My Step-by-Step Process

I use DALL-E for almost every blog post I write. Whether it's a header image, an illustrative diagram, or a decorative thumbnail, DALL-E 3 and the newer GPT-Image model have become my go-to tools for generating blog visuals. They're fast, the quality is consistently good, and integration with ChatGPT makes it easy to iterate on concepts.

But there was a problem I didn't notice for months: every image I generated and posted on my blog was carrying hidden metadata that identified it as AI-generated. Not only that, but the metadata included information about when each image was created and which OpenAI tool produced it.

I only discovered this when a reader mentioned that some of my blog images were showing "AI-generated" labels when shared on social media. That led me to investigate what DALL-E was embedding in my images — and how to clean it.

What DALL-E Metadata Contains

DALL-E 3 and GPT-Image embed C2PA Content Credentials into every generated image. When I inspected my first DALL-E image with a metadata viewer, here's what I found:

C2PA manifest containing:

• claim_generator: "OpenAI" or a similar identifier
• actions: An assertion that the image was generated by AI
• timestamp: The exact UTC timestamp of generation
• software_agent: Information about the specific DALL-E model version
• ingredients: Sometimes references to source materials used in the generation

XMP metadata containing:

• Creator tool information referencing OpenAI
• Generation timestamps
• Sometimes model identifiers

EXIF data (minimal):

• Software fields that may reference the generation tool

The C2PA manifest is the most significant piece. It's cryptographically signed, meaning it's designed to be tamper-evident. You can't edit it — you can only remove it entirely.

Why This Metadata Matters for Bloggers

If you're using DALL-E images on your own blog, the metadata doesn't directly affect your readers. Browsers don't display C2PA information, and most visitors won't inspect your image files.

But here's where it matters:

Social media sharing. When readers share your blog posts on Facebook, LinkedIn, or other platforms, those platforms may scan the images for C2PA metadata and apply "Made with AI" labels. I had blog posts shared on LinkedIn where the header image — generated with DALL-E — showed up with an AI label, which undermined the credibility of the article.

Content aggregation. Some content aggregators and RSS readers now scan for C2PA metadata. Your images might be labeled or flagged without you knowing.

Future compliance. As more platforms implement AI content labeling, having C2PA metadata in your images means they could be affected by future policy changes.

Privacy. The generation timestamp in DALL-E metadata reveals exactly when you created each image. If you're publishing under a schedule or working on time-sensitive content, this can expose your workflow timing.

My Step-by-Step Removal Process

Here's the exact process I use for every DALL-E image before adding it to a blog post:

Step 1: Generate and Download

I generate the image in ChatGPT or the OpenAI API and download it to my computer. DALL-E 3 images typically download as PNG files, while GPT-Image outputs can be PNG or WebP.

Step 2: Strip DALL-E Metadata

I use the Remove DALL-E Metadata tool, which is specifically designed to handle the metadata OpenAI embeds:

1. Open the tool in my browser.
2. Upload or drag-and-drop the DALL-E image.
3. The tool identifies and removes:
   • C2PA Content Credentials (JUMBF manifests)
   • XMP packets with OpenAI tool references
   • EXIF software fields
4. Download the cleaned image.

The process takes about 3-5 seconds per image. For PNG files, it removes the C2PA data from the caBX chunk. For JPEG exports, it strips the APP11 markers containing JUMBF data.

Step 3: Handle ChatGPT-Specific Metadata

If I generated the image through ChatGPT (rather than the API directly), there's sometimes additional metadata. In those cases, I run the image through the Remove ChatGPT Image Metadata tool as well, which targets metadata specific to ChatGPT's image generation pipeline.

Step 4: Verify

I verify the cleaned image by running it through the Remove AI Metadata tool's inspection mode. This shows me a complete metadata breakdown so I can confirm everything is gone. I also check at contentcredentials.org/verify for an independent verification.

Step 5: Optimize and Upload

With the metadata stripped, I resize and compress the image for web (using TinyPNG or similar), then upload it to my blog's CMS. The cleaned image is indistinguishable visually from the original — same quality, same colors, same everything. The only difference is the absence of hidden metadata.

Handling Multiple DALL-E Images

When I'm writing a blog post with several DALL-E images, I process them as a batch. The Remove AI Metadata tool supports batch uploads, so I can drop 10-15 images at once and clean them all together. This saves significant time when I'm preparing a long-form article with multiple visuals.

Tips for DALL-E Users

Clean before resizing. I used to resize images first, then clean them. But some image processing tools can inadvertently preserve or re-encode C2PA data during resize operations. Now I clean first, then resize.

Watch out for ChatGPT's edit feature. When you ask ChatGPT to edit a DALL-E image (like "make the sky more blue"), the edited version gets a new C2PA manifest. Each edit adds a new link to the provenance chain. Clean the final version, not intermediate ones.

PNG vs WebP. GPT-Image sometimes outputs WebP files. C2PA in WebP uses a different storage mechanism (RIFF container metadata) than PNG. The removal tools handle both, but if you're manually inspecting files, know that the metadata location differs.

Keep your originals. I keep a folder of unedited DALL-E originals with full metadata. If I ever need to prove that I generated an image (for copyright purposes or client work), the original C2PA manifest is useful evidence.

Why I Don't Just Use Stock Photos

People have asked me why I don't skip the whole AI image thing and just use stock photos. The answer is simple: DALL-E images are more specific to my content. When I write about a technical concept, I can generate an illustration that exactly matches what I'm describing. Stock photos are generic; DALL-E images are custom.

The metadata issue is a minor inconvenience solved by a 5-second cleaning step. The creative flexibility DALL-E gives me is worth far more than the small cost of running each image through the Remove DALL-E Metadata tool before publishing.

My Final Workflow

For reference, here's my complete DALL-E image workflow from start to finish:

1. Generate image in ChatGPT or via API.
2. Download the image (PNG preferred).
3. Run through Remove DALL-E Metadata (or Remove ChatGPT Image Metadata for ChatGPT-generated images).
4. Verify with Remove AI Metadata inspection.
5. Resize and compress for web.
6. Upload to blog CMS.

Six steps, about 30 seconds total per image. And I never have to worry about "Made with AI" labels appearing on my content when it gets shared. That peace of mind is worth the small extra effort.