Content provenance is the attempt to answer "where did this file come from?" in a way that survives being shared. It covers two different mechanisms that often get discussed as one.
The first is metadata: a signed manifest travelling with the file recording the tool, the model, and the edit history. Readable, verifiable, and easy to strip — a screenshot removes it entirely.
The second is watermarking: a statistical pattern embedded in the pixels or the word choices themselves, invisible to a person but detectable by the right checker. Harder to remove, harder to verify independently, and detectable only by whoever holds the detection method. Anthropic's invisible watermarking work sits in this second category.
Why it matters at your desk. If you publish, provenance is becoming a disclosure question rather than a technical one. Platforms increasingly label AI-generated media automatically, and a marketer or designer shipping work with stripped-out credentials can find it labelled anyway, or flagged inconsistently across platforms. Knowing which of your tools write credentials, and which of your export steps destroy them, is now part of the workflow.
If you receive media, the limits matter more. Absence of a provenance signal proves nothing — it may have been generated without one, or cropped, or re-encoded. Google pulling an AI image tool after provenance concerns is a reminder that the ecosystem is still being built.
What to watch for: provenance tells you what made a file, never whether its claims are true. A verified-authentic photograph can still be staged, and a correct document can carry no credentials at all.