PTN—29

Content provenance

AI-generated content carries a human-readable origin marker that survives export and records the human’s edits.

AI content outlives the session that made it. The draft becomes a doc, the doc gets pasted into a deck, the deck gets forwarded — and three hops later, a reader is weighing words with no way to know a model wrote them. Self-identification solves the conversation; provenance solves the artifact. It is the label that travels with the content: who made this, with what, when, and how much a human has touched it since.

The marker must be readable by people, not just parsers. Cryptographic metadata like C2PA is the right substrate — tamper-evident, machine-verifiable — but a signature only software can see does nothing for the colleague reading the memo. Good provenance has both layers: a durable technical record underneath, and a plain-language surface on top — “AI-drafted, March 2026” — that a human encounters without installing anything. And it must survive the export path, because a label that dies on copy-paste exists exactly where it matters least: inside the app, before the content travels.

Provenance is a state, not a stamp. “AI-generated” is accurate for ten seconds; then a person rewrites the intro, corrects two figures, and signs off — and the honest label is now “AI-drafted, human-edited.” Recording that transition matters in both directions: it stops human judgment from being dismissed as machine output, and it stops machine output from borrowing a human byline it never earned. A provenance system that cannot represent the mixed case mislabels nearly everything, because nearly everything shipped is the mixed case.

Keep the surface calm and the depth inspectable. The everyday reader needs one quiet line; the auditor needs the model, the date, and the generation details on demand. A provenance panel behind the marker serves both without turning every document into a disclosure form — the label carries the honesty, the details carry the proof.

Anatomy

Fig. 01 — Content provenanceLive specimen

Document · renewal letter

Gold membership renewal

AI-drafted

Hi Dana — your Gold membership renews on August 1.

Renewing keeps your two seasonal tune-ups, priority scheduling, and 15% off repairs — $189 for the year, same as last year.

A document with a human-readable provenance marker that tracks edits and opens to full generation details.

1Readable marker. A plain-language origin label a person can read unaided — the metadata is the substrate, not the interface.
2Survives export. The marker travels with copy, download, and export — provenance that dies in transit protected nothing.
3Edit-aware state. Human edits update the label — “AI-drafted, human-edited” — so the mixed case is represented honestly.
4Inspectable details. Model, date, and prompt are available on demand for whoever needs the proof, not stapled to every view.

When to use

AI-generated content leaves the product — export, copy, download, publish.
Artifacts mix machine drafting with human editing and the ratio matters to readers.
The domain has disclosure obligations — journalism, marketing, legal, regulated communications.
Downstream readers will make credibility judgments based on who they think authored the content.

Design considerations

01Layer the marker: plain language on the surface, verifiable metadata underneath — each is useless without the other.
02Test the export paths — copy, PDF, download, print — and treat a label that strips in transit as a broken feature, not a limitation.
03Update provenance as humans edit, and grade it honestly: AI-generated, AI-drafted-human-edited, human-written-AI-assisted are different claims.
04Keep the everyday marker quiet and the audit trail deep — one line for readers, full details behind it for verifiers.
05Decide what the prompt discloses before exposing it; provenance details can leak more than the author intended.
06Adopt the open standard where one exists — a proprietary marker dies at your product’s boundary, which is exactly where provenance is needed.

Pitfalls

✕Metadata-only provenance: a C2PA signature no reader ever surfaces, technically present and practically invisible.
✕Labels that strip on export, so the content is marked everywhere it is safe and anonymous everywhere it travels.
✕The frozen stamp — content still labeled “AI-generated” after a human rewrote most of it, which misattributes their judgment and teaches people to discount the label.
✕Provenance as scarlet letter: styling the marker as a warning rather than a record, which pushes authors to strip it rather than carry it.

In the wild

Adobe Firefly: Attaches Content Credentials to generated images — C2PA metadata plus a human-inspectable panel showing how the asset was made, surviving export.
Google SynthID: Watermarks AI-generated images and text at generation time so origin remains detectable after the content leaves the product.
ChatGPT: Images from DALL·E 3 carry C2PA provenance metadata identifying them as AI-generated, verifiable by anyone with a credentials inspector.