The EU's AI Watermarking Deadline Already Passed. Does Compressing an Image Break Compliance?

Article 50 of the EU AI Act has been enforceable for almost a month now, and the compliance conversation so far has mostly been about which AI tools embed a disclosure credential at generation time — not about what happens to that credential the moment anyone actually uses the image afterward. That's the gap worth closing: C2PA Content Credentials, the mechanism the European Commission names as an example compliant approach, live in file metadata. Metadata is precisely what a huge share of ordinary image tools discard by default, because stripping it is one of the easiest ways to shrink a file.

What Actually Changed on August 2

Adobe Firefly, OpenAI's DALL-E 3 and Sora, and Google Imagen already ship C2PA credentials embedded in their output, which is genuinely good news — the biggest names in mainstream AI image generation were already positioned to comply before the deadline arrived. Midjourney is the clear exception: it remains the most prominent generator without default C2PA support, which means any Midjourney output reaching an EU user right now sits in direct tension with the law as written, unless that changes soon.

OpenAI went a step further back in May, licensing Google DeepMind's SynthID watermark as a second layer alongside C2PA metadata and previewing a public verification tool that checks both marks at once. That two-layer approach exists for a specific reason: an invisible, pixel-level watermark like SynthID can survive some transformations that a metadata tag can't, which is a meaningful hedge against exactly the problem this post is about.

The Metadata Problem the Guidance Already Half-Admits

The European Commission's own Code of Practice recommends a multi-layer approach — metadata plus invisible watermarking plus logging — rather than relying on any single mechanism alone. That combination isn't there for redundancy's sake; it's an acknowledgment that metadata-only credentials are fragile in practice. Instagram, X, and WhatsApp all strip C2PA data during their own recompression pipeline when a user uploads an image, which means the moment a compliant, C2PA-tagged AI image gets posted to any of the platforms where images actually circulate, the credential is already gone before another human being sees it.

That's a platform-level problem regulators have acknowledged but not solved. The part almost nobody is talking about sits one step earlier: the same fate applies to a completely ordinary, non-adversarial action — someone resizing an AI-generated image for their own website, or running it through a compression tool to make a page load faster. Most image-processing tools treat metadata as disposable weight to shed, precisely because dropping it is a quick, safe way to reduce file size without touching visible pixels. A credential embedded specifically to survive scrutiny can be gone after a completely routine, well-intentioned edit that had nothing to do with hiding anything.

What This Means If You Publish AI-Generated Images

1. Check whether your image pipeline preserves or strips metadata before you build a workflow around C2PA compliance. Many popular compression and resizing tools default to stripping it; some offer a metadata-preserving mode that isn't the default and has to be turned on deliberately.
2. Understand that "compliant at generation" doesn't mean "compliant at publication." The credential your AI tool embedded is only as durable as every step the image passes through afterward — your own CMS, your own image optimization step, any third-party hosting.
3. Don't assume a social platform repost carries the same disclosure the original file did. If disclosure matters for your use case, the visible watermark or in-image labeling approach survives platform recompression in a way metadata alone does not.
4. If you're building or choosing an image tool for a workflow that includes AI-generated content, ask specifically whether it has a metadata-preserving export option — this is not yet a standard checkbox most tools advertise clearly.

The Practical Takeaway

Article 50 put a real deadline on AI disclosure, and the major generators mostly did their part before it arrived. The open question now isn't whether the credential gets created — it's whether it survives the ordinary journey an image takes after that: a resize, a compression pass, a repost. That's a solvable problem, but only if it's treated as a real step in the pipeline instead of an assumption nobody checked.

Related reading:

• The EU AI Act Is Live: What Website Owners Must Know About AI Images — the original compliance guide from before the deadline arrived
• Compress AI-Generated Images: Midjourney and DALL-E, 2026 — practical compression guidance for AI images specifically
• X's Grok Everywhere: An AI Fake Detection Photo Guide — the detection side of the same authenticity problem