You see an AI label under a photo, article, or video. The temptation is to treat it like a score: machine-made, human-made, trustworthy, suspicious.
Pause. The label has a smaller job.
This smaller job now matters in law as well as daily life. European Union transparency rules that began applying August 2 require machine-readable marks on certain AI-generated outputs and visible disclosures in some situations. The rules have specific scopes and exceptions. The practical distinction travels well beyond Europe: a mark can tell you something about how material was processed. It cannot grade the work for you.
That distinction becomes clearer when you separate the two signals. Machine-readable marks may sit inside a file as metadata or an embedded watermark. Visible labels sit where a person can see them. European Commission testing found that its optional icon worked better when paired with explanatory text. The Commission also says using the icon alone does not establish legal compliance.
Provider documentation draws an equally useful boundary. Anthropic says a detected Claude mark indicates that the material may have been processed by Claude. That does not settle who created the underlying idea or whether the claim is accurate. OpenAI says uploads, screenshots, resizing, and format changes can remove or break provenance metadata.
And a missing mark leaves the question open too. Older models, editing, translation, short passages, unsupported file types, and platform conversions can all weaken the trail.
Read the label twice
When you encounter one, ask:
- What did it detect? Look for plain language explaining whether the signal is a visible disclosure, metadata, or a watermark.
- What happened afterward? Editing, conversion, screenshots, and resharing may have changed the evidence before it reached you.
- Who checked the claim? A processing mark says nothing about whether somebody verified the facts and accepted responsibility for publishing them.
The label is worth reading. Then inspect the work behind it.