On 2 August 2026, the transparency layer of the EU AI Act stopped being a future obligation.
Article 50 now applies. The Verge covered the arrival of the labels themselves, and the practical weight sits in three requirements that most marketing departments have not read.
Providers of generative systems must mark output in a machine-readable format so it is detectable as artificially generated. Deployers producing deepfakes must disclose that the content was artificially generated or manipulated. Systems built to converse with people must make clear that the person is talking to a machine.
This lands on marketing, not on legal
Read the obligations again and notice who actually produces the artifacts. Product photography variants. Voiceovers. Synthetic presenters. Ad creative generated in bulk. Support chat widgets.
That is a marketing inventory, not a compliance inventory. It is also the same asset class I wrote about when synthetic likeness became a licensable asset. The legal team will own the policy and the marketing team will own every single asset the policy describes.
The exposure is not symbolic either. Infringement of the transparency obligations can reach 15 million euro or 3 percent of total worldwide annual turnover, whichever is higher.
For a mid-market company, 3 percent of global revenue is not a fine you absorb. It is a budget line that did not exist last quarter.
The delay is real and it is a trap
There is a complication worth understanding precisely. The AI Omnibus proposal contemplates transitional relief specifically for the marking and detection obligation in Article 50(2), with the Council pointing at a revised deadline of 2 December 2026.
That relief has not been formally adopted. So the honest status is that one sub-obligation may move, the rest applied on 2 August, and the final answer arrives later.
I have watched enough regulatory cycles to know what most teams will do with that ambiguity. They will read the word December, file the whole thing under next quarter, and do nothing.
That is a mistake for a reason that has nothing to do with law. The work required to comply is the same work required to make your content trustworthy to machines, and that second reason pays regardless of what Brussels decides.
Build the provenance pipeline, not the label
Treating this as a labeling task produces a label. Treating it as an architecture task produces something useful.
The underlying requirement is provenance: for any asset you publish, you should be able to answer where it came from, whether a model generated it, whether a human reviewed it, and when.
Most content operations cannot answer those questions today. Assets move from a generation tool to a designer to a scheduling platform, and the origin metadata falls off at the first handoff.
So the fix is boring and structural. Record generation origin at creation, carry it through your asset management, and make human review an explicit recorded step rather than an assumed one.
The marking requirement itself is not a visible sticker. It asks for output to be detectable as artificially generated in a machine-readable way, which in practice means embedded metadata and watermarking standards that survive a re-export.
That word survive is doing heavy lifting. Metadata dies when a file passes through a compression step, a screenshot, or a social platform that strips it, and most marketing workflows contain all three.
The chatbot obligation is the one I expect to catch companies off guard, because it is the cheapest to fix and the easiest to forget. If a support widget on your site is answering in natural language, the person on the other end needs to know they are talking to a machine, stated clearly rather than buried in a tooltip.
Do that and the label becomes a rendering detail. Skip it and you will be reconstructing the history of ten thousand assets under deadline, which is the expensive version of this project.
The second-order effect is competitive
Here is the part I find more interesting than the compliance question.
Once artificial origin is machine-readable at scale, every platform that ranks content gains a cheap new signal. They do not have to guess anymore. Snapchat has already started demoting wholly AI-generated video, and it will not be the last. This is the AI content bill coming due, just billed by a regulator this time.
Which means the labeling regime quietly becomes a distribution regime. The mark you attach for a regulator in Brussels is readable by a recommendation system in California.
Companies treating this as paperwork will comply and lose reach. Companies treating it as a provenance upgrade will be able to prove what a human made, at the exact moment proving it starts to matter commercially.
The teams that start now will be doing metadata hygiene. The teams that start in November will be doing archaeology on their own asset library, under a deadline, with a legal review attached.
The label is not the point. The record behind it is.