OpenAI says DALL-E 3 will add watermarks to images with C2PA metadata but acknowledges the metadata can “easily be removed either accidentally or intentionally”

OpenAI's image generator DALL-E 3 will add watermarks to image metadata as more companies roll out support for standards …

The Verge 2024-02-07 Emilia David

Context & Ripple Effects

OpenAI is attaching C2PA provenance data to DALL-E 3 outputs after researchers had already shown that AI-image watermarks can be removed or falsely applied to human-made images, underscoring the gap between labeling and reliable authentication researchers showed watermark labels could be stripped or spoofed.

The move matters as a baseline disclosure mechanism, not as a durable enforcement tool. Related coverage later shows OpenAI supplementing provenance efforts with a detector for unaltered DALL-E 3 images, reflecting the need for multiple signals when metadata does not survive editing or reposting.

First-order effects

DALL-E 3 images will carry C2PA metadata, giving compatible tools a standardized way to display their AI provenance.
OpenAI explicitly limits the claim: ordinary editing or deliberate removal can sever that provenance record, so the label cannot reliably follow every copy.

Second-order effects

Publishers, platforms, and developers that want to use the signal must implement C2PA reading and display flows; interoperability gaps will determine how often users actually see it.
Because metadata can disappear, downstream moderation and verification systems cannot treat a missing C2PA record as evidence that an image is human-made; they need complementary checks.

Third-order effects

If model providers converge on provenance standards while accepting their fragility, AI-image disclosure is likely to become a layered system of metadata, detection, and verification rather than a single watermark test.
The central competitive question shifts from whether a provider labels outputs to whether its provenance signals remain interpretable across creation, editing, and distribution tools.

The trend: This is one step toward a synthetic-media control plane in which provenance metadata provides useful disclosure but must be combined with other verification mechanisms.

Discussion

@marypcbuk.bsky.social Mary Branscombe on bluesky
The point of C2PA isn't to have metadata you can't remove, it's to have content origin data be so universal that removing the metadata is suspicious of itself [embedded post]
@openai @openai on x
Images generated in ChatGPT and our API now include metadata using C2PA specifications. This allows anyone (including social platforms and content distributors) to see that an image was generated by our products. https://help.openai.com/... [image]
@mer__edith Meredith Whittaker on x
And if I screenshot the image? <ducks>
@gdb Greg Brockman on x
Provenance metadata for images generated by our products:
@af_mada @af_mada on x
Awesome to see this initiative going forward. Huge step to see both Meta & OpenAI following C2PA protocol. @X when is your turn ? It would be a great addition to community note, specifically before US election !
@iuliabaltoi Jules on x
It's great to see broader adoption for media provenance protocols like C2PA in an effort to fight misinformation!
@slyon66 Santiago Lyon on x
Fantastic to see @OpenAI implementing Content Credentials on Generative AI images from @dalle_openai - all part of the work of the @C2PA_org and @ContentAuth !!
@henkvaness @henkvaness on x
So, @OpenAI now supports metatags, which is also beneficial for researchers, #OSINT (Open Source Intelligence) practitioners, and verification specialists. OpenAI claims that, on social platforms, it's now possible to identify images generated by their products through these...
@orobicoak @orobicoak on threads
I'd be curious to see where they'll draw the line, there'll be cases where digital art will be labelled incorrectly. For example, many artists paint over AI pictures, or make pieces that combine AI imagery with other media. It's hard to look at this matter in black and white. …
@skoskie @skoskie on threads
AI images need an encoded tag so they can be easily recognized. Otherwise AI will begin to consume its own images which will begin to distort its own reality. In other words it will begin to think six-fingered people are actually common, and produce more images of people with s…
@ralfskirr Ralf Skirr on threads
Meta will soon start labeling AI images on Facebook, Instagram, and Threads with “Imagined with AI” labels. I love AI generated images, but I think it's really important that they are identifed as such. It's already almost impossible to know what is a real photograph and what i…
@emilybell Emily Bell on x
This is a key commitment by Meta. In the past - particularly around its ‘fact check’ flags etc - Meta's labeling or flagging of content has been hopelessly inconsistent . Can it do better with AI?
@alexeheath Alex Heath on x
Talked to Meta's @nickclegg about why they are starting to label AI-generated photos. He said realistic video/audio made with AI also needs to start being disclosed by users or it could be taken down. Also: “I think it's really unlikely that you're going to get a video or audio..…
@mer__edith Meredith Whittaker on x
The election year focus on ‘deep fakes’ is a distraction, conveniently ignoring the documented role of surveillance ads—or, the ability to target specific segments to shape opinion. This's a boon to Meta/Google, who've rolled back restrictions on political ads in recent years.
@nickclegg Nick Clegg on x
Content from partners incl Google, OpenAI, Microsoft, Adobe, Midjourney, & Shutterstock will be marked with labels denoting if something was created w AI. While this isn't a perfect solution & won't catch everything, it's an important first step in an evolving process.
@nickclegg Nick Clegg on x
Today, we're announcing that we will label AI-generated images from industry partners that get posted to Facebook, Instagram, and Threads. https://about.fb.com/...
@metanewsroom @metanewsroom on x
Is it AI? The line between AI and human-made content can be blurry, but labeling it shouldn't be. We're rolling out industry-leading practices that will identify AI-generated images across @facebook, @instagram and @threadsapp__. https://about.fb.com/... [video]
@andymstone Andy Stone on x
Since its launch, we've labeled photorealistic images created using Meta AI so people know what they're looking at. In the coming months, we'll also label images created elsewhere when we can detect industry standard indicators that they're AI-generated. https://about.fb.com/...

Chronicles