By Shelby Britton, VP Product Marketing, Cloudinary
While we swelter in endless heatwaves, it’s hard to even fathom the chill of December 2, 2026. But that is precisely when the grace period runs out for certain use cases to comply with the EU AI Act’s Article 50 transparency provisions that were enacted on August 2. The compliance burden will fall on brands publishing AI-generated or AI-manipulated content in the EU. Failure to do so may result in hefty fines.
Article 50 mandates that providers and deployers make users aware when they are interacting with AI or viewing synthetic, manipulated, or specific biometric-driven content. That means a brand using generative AI in its marketing must use machine-readable formats and technical labeling to mark audio, image, video, or text outputs as “artificially generated”.
This means AI compliance has to get real. It can’t live exclusively in a legal policy document, and, as a priority, must become an organic part of your entire media pipeline. That means for many racing against the deadline clock, AI transparency just became an urgent operational problem. Fortunately, solutions are at hand.
Embedding transparency into the workflow
Resource-strapped creative teams are using AI to transform existing assets in ways that would have taken hours – or a small army of Photoshop wizards – just a few years ago. But if AI touches an asset, someone will now have to be on top of the right transparency information, wherever it goes.
For visual content, there is a big distinction in the regulation between machine-readable transparency and human-facing disclosure. In a nutshell, machine-readable information allows software and platforms to understand that AI was involved and, potentially, how the asset was created or modified, while human-facing labels are designed for situations where viewers could otherwise be misled (think, deepfakes).
That matters, because not every AI edit needs to have a glaring “TOUCHED BY A ROBOT” badge slapped on it. But which ones do?
Let’s say an enterprise visual or video asset starts life by either being uploaded or generated directly in a DAM using an AI tool. Then a designer crops it, a localisation team creates five regional variants, the ecomm guys merrily go off and create a dozen responsive sizes, and finally the CDN delivers those assets to websites, apps, marketplaces and social channels.
Here’s our problem: at which step did the compliance happen? If the answer is “someone on the marketing team added a label in Photoshop,” you have a scaling problem. And this is where it gets hairy. The original asset might be compliant, but by the strict terms of the Act the derivative might not be. And as things progress, more noise could enter the signal; someone might download a clean version, but someone else might inadvertently remove metadata during processing. Six months on, good luck telling an auditor which version involved AI.
This is why AI transparency needs to be treated as an asset-management and delivery problem rather than a manual labeling exercise. The only practical approach is to embed machine-readable provenance into the workflow so that compliance information travels as a permanent companion to the asset.
Really, you need to get yourself set up so that you can always answer the following AI image questions:
- Where does AI enter our workflow?
- Where is that information recorded?
- Does provenance survive transformation?
- What happens when one of our assets becomes a derivative?
- Can disclosure be applied automatically at delivery?
- Can marketing and legal teams govern these rules centrally?
C2PA – a useful standard to base your work on
This is where—if you’re not familiar with its work—the C2PA (Coalition for Content Provenance and Authenticity) becomes particularly interesting. What people I talk to like about it is that C2PA provides a nice, practical standardised framework for attaching verifiable provenance information to digital media. Usefully, its Content Credentials can capture information about an asset’s origin and subsequent transformations, including whether AI was involved.
The EU AI Act doesn’t mandate C2PA specifically, but it’s emerged as one of the leading practical approaches for interoperable, machine-readable provenance, and it is increasingly supported across the technology ecosystem.
However you decide to move ahead, your colleagues in both Legal and Marketing should be able to take comfort that AI image compliance is baked into your workflow.
Ideally, that should translate to a behind-the-scenes process where
- if the DAM knows an asset was AI-generated, that information should always be available to downstream transformations and delivery systems
- when a derivative is created, the provenance should be preserved or appropriately updated
- and when the asset is delivered, the relevant disclosure can be applied automatically.
Like it or not, AI transparency compliance is becoming part of the media supply chain. The teams that operationalise it inside their DAM and delivery pipelines will be in a much better position to react to whatever the market and the technology decide to tempt us with next.
