What the EU AI Act Means for Your Image Files

What changed on 2 August 2026

Article 50 of the EU AI Act sets transparency obligations, and its requirements applied from 2 August 2026. The part that reaches your image files is short: providers of AI systems generating synthetic image, audio, video or text must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.

Machine-readable is the operative phrase. A visible watermark in the corner does not satisfy it, because the point is that other software can check the file automatically. The obligation is also framed with a feasibility qualifier — solutions must be effective, interoperable, robust and reliable as far as is technically feasible, accounting for the state of the art — which is why implementations differ between vendors rather than converging on one method.

In practice the industry settled on two layers: a signed C2PA manifest attached to the file, and for some vendors an invisible watermark embedded in the pixels. The first is metadata and comes off. The second is part of the image and does not.

This is why your tools changed

If you have noticed provenance data appearing on exports where it did not before, this is the cause. Vendors have been moving ahead of the deadline rather than waiting for it.

Adobe made Content Credentials mandatory for workflows touching generative features and removed the in-app option to disable them, with photographers reporting credentials on exports more broadly than they expected. OpenAI signs DALL-E and ChatGPT image output. Google output carries both Content Credentials and SynthID.

None of that is a change in what the tools can do. It is a change in what they write down, driven by a compliance deadline rather than by anything about your workflow.

The December date almost nobody mentions

The August date is not the end of it, and this is the part worth planning around.

The AI Omnibus provisional agreement reached in May 2026 gives generative systems that were already on the market before 2 August until 2 December 2026 to meet the machine-readable marking requirement under Article 50(2).

So the tools writing nothing today are not exempt — they are inside a grace period. If your current workflow produces unmarked files, the reasonable assumption is that it will start producing marked ones between now and December, without you changing anything.

That makes any list of which tool marks what a snapshot rather than a standing fact, including the ones on this site. Checking your own exports periodically is the only approach that stays accurate.

Who the obligation actually falls on

This is the question people get wrong most often, and the distinction matters.

The marking duty in Article 50(2) is placed on the provider of the AI system — the company shipping the model or the tool. It is a requirement that outputs leave their system marked. It is not framed as a duty on you to keep a mark that is already in a file you hold.

There is a separate obligation, in Article 50(4), on those who deploy AI systems that generate or manipulate deep fake content, requiring disclosure that the content is artificially generated or manipulated. That one attaches to the person publishing, and it is about disclosure rather than about metadata.

The practical consequence: removing a manifest does not discharge a disclosure duty if one applies to you. The two are separate mechanisms, and a clean file does not answer a question the law is asking about the content.

This is a description of how the provisions are structured, not legal advice. If your work falls near deep fake territory or you publish commercially into the EU, the disclosure question is worth asking someone qualified rather than inferring from a metadata guide.

What it means for photographers specifically

The awkwardness for photographers comes from a mismatch the law does not really address. The marking records that a generative feature was involved. It does not record proportion.

A photograph where a generative tool removed a power line is marked the same way as an image generated entirely from a prompt, and the platforms reading those marks mostly do not distinguish between them either. Retouching has removed distractions for decades; what changed is that the tool now writes it down.

Two things follow. Removing the manifest from files you publish is ordinary metadata removal and does not alter the picture. And the copies you keep are the ones that matter for proving authorship, so archive originals with their metadata intact and clean only what goes out.

What the marking cannot do

Provenance metadata makes a claim about origin. It does not establish that an unmarked file is a camera original, and this asymmetry is permanent.

A file can arrive unmarked for many innocent reasons: a screenshot, a format conversion, a platform re-encode, a tool inside its grace period, or a generator outside EU jurisdiction. Absence of a mark is not evidence of anything.

Marking also says nothing about truthfulness. An image can be entirely camera-captured and still be a lie about what it depicts, through framing, caption or context. Provenance answers how a file was made, which is a narrower question than whether it is honest.