Since Article 50 of the EU AI Act started to apply on 2 August 2026, the question inside most content teams has shifted from whether to disclose to what exactly to write. "Made with AI" is not an answer. It covers a model that wrote every word and a model that suggested a headline, and those are not the same claim.
Four statements cover almost everything a publishing team produces. Here is what separates them.
Why one label is not enough
The obligation is about accuracy, not about having said something. A disclosure that overstates model involvement is inaccurate. A disclosure that understates it is inaccurate in the direction that costs you more, because that is the one somebody has a reason to challenge.
There is also a commercial argument that has nothing to do with the regulation. A blanket "AI was involved" flattens the work your team actually did. If a person researched, structured, wrote and edited a piece and a model helped somewhere in the middle, saying so precisely is a better description of the work than a generic warning label.
AI generated
What it means. A model produced the content, with little or no human authorship in the result. A person may have prompted it, selected between outputs, or corrected a typo. The substance is the model's.
Example. A product description generated from a spec sheet for two thousand catalogue items, spot-checked but not rewritten.
The common mistake. Teams avoid this label because it feels like an admission. If it is what happened, it is the accurate statement, and an accurate statement is not a weakness. The weakness is a piece labelled AI assisted that a reader can tell was generated.
AI modified
What it means. A human made the work, then a model altered it enough to change what it shows. The distinction that matters is whether the alteration changed the substance, not whether a tool touched the file.
Example. A photograph where a model removed an object from the background, or a recorded interview where a model altered phrasing in the transcript.
The common mistake. Treating every automated adjustment as a modification. Colour correction, noise reduction and grammar checking do not change what the work shows. Removing a person from a photograph does.
AI assisted
What it means. A human authored the work and used a model as a tool along the way. The model contributed to the process. The result is the person's work.
Example. An article a writer researched, structured and wrote, using a model to summarise sources, suggest an outline, or tighten a paragraph they then rewrote.
The common mistake. Assuming this is a softer version of AI generated and reaching for it when the truth is closer to generated. This is the label most working teams honestly are, which is exactly why it needs to be applied honestly rather than as a default.
Human authored
What it means. No model involvement worth declaring.
Why declare it at all. This is the one teams skip, because saying nothing feels equivalent. It is not. Silence is ambiguous, and it will become more ambiguous as more published content carries a declaration. A positive statement that a person wrote this, recorded at the time of publication, is worth more the more common AI content becomes.
The common mistake. Claiming it because no model was used in the final draft, while ignoring a model used in research or structuring. If a model shaped the work, the honest label is AI assisted.
How to decide in under a minute
Ask three questions in this order.
Did a model produce the substance of what a reader sees? If yes, AI generated. Stop.
Did a model change what the finished work shows, after a person made it? If yes, AI modified. Stop.
Did a model help at any point in producing it? If yes, AI assisted. If genuinely no, human authored.
The order matters. Working down from the strongest claim stops you settling on the most comfortable one and calling it accurate.
Declaring it so it can be checked later
Picking the right label is the first half. The second half is that the label is a statement you are making, and the party making a statement is the party asked to support it.
A label written on a page can be edited afterwards, including by you. A declaration bound to a hash of the exact file and fixed in time by a qualified timestamp cannot. That difference is invisible until a client, a platform or a counterparty asks you to substantiate a declaration you made months earlier, and then it is the whole thing. We set out what Article 50 asks for and where a sealed declaration fits on the Article 50 page.
Pick the accurate label. Then make it something you can still prove in six months.
