Overview
This series asks what an AI-content label actually tells a reader. It first separates EU AI Act duties by actor, output, context, exceptions, and timing; then explains how C2PA preserves declared provenance and how validation limits its claims. A U.S. experiment with text policy messages shows labels can be believed without significantly changing persuasion. The practical answer is layered disclosure, provenance, fact-checking, and audience measurement—not a label treated as truth or completeness.
Learning outcomes
- The EU AI Act ties disclosure to the actor, interaction or output, context, exceptions, and timetable—not to AI involvement alone.
- Human-facing notices, machine marks, provenance, and independent verification solve different problems and may need to coexist.
- A valid or trusted C2PA record establishes integrity and signer status under a trust model, not truth, completeness, or expertise.
- In a 1,601-person U.S. experiment, participants often believed authorship labels, but labels did not significantly change persuasion.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
When does the EU AI Act require AI-generated content to be disclosed?
It depends on the actor, the interaction or output, the context, the exceptions, and the staged timetable; the Act does not impose one blanket label on every AI-assisted file.
What can C2PA provenance establish about an AI-made file?
It can record declared actions, inputs, signatures, and content bindings and let a validator check specified integrity and trust conditions. It cannot by itself prove the content is true or the history complete.
Are human-facing labels, machine marks, and provenance records interchangeable?
No. A human-facing disclosure informs people, a machine-readable mark helps software detect output, and provenance records declared history; they can coexist but do not replace fact-checking.
Do AI authorship labels reduce persuasion?
Not necessarily: in a 1,601-person U.S. experiment using text messages about four policy proposals, labels were often believed but did not significantly change support, confidence, perceived accuracy, or stated sharing intention.