AI Provenance

Framing

The problem

In a document you co-write with AI, machine-drafted and human-written text read identically — so you cannot tell which parts still need a human to check them.

The pattern

Treat AI like tracked changes: mark the AI-drafted spans no human has reviewed, let the user accept or edit them span by span, and keep that review state attached to the document.

Why chat breaks here

Chat hands back one finished block with no record of what was reviewed; once it leaves the chat, AI-drafted and human-vetted text look exactly the same.

Risks

If everything is marked it is noise (the original mistake); if accepting is one bulk click it becomes rubber-stamping — the signal only works when unreviewed is the exception, not the default.

Avoid when

Throwaway or low-stakes text where no one will ever need to tell an AI draft from vetted prose.

Use when

When you co-write with AI, machine-drafted and human-written text read the same — so the useful signal is which parts a human has actually reviewed.

DOPE evaluation

Directability

Read, accept, or edit any AI-drafted span — the marker clears the moment a human has touched it

Observability

At a glance: which spans are AI-drafted and unreviewed vs. which a human has vetted — not a wall of identical "AI" markers

Predictability

Consistent everywhere: vetted text reads clean, unreviewed is marked, a claim that drifts from its source is flagged

Explainability

Open any marked span for what produced it — model, prompt, source-match — so you can vet it, not just trust it

In the wild

• C2PA Content Credentials (C2PA) — C2PA 2.1 ratified as ISO/IEC 22144 (2025). Shipping on Pixel 10, Galaxy S25, OpenAI image gen, DALL-E 3, Sora, Imagen, Adobe Firefly, Leica/Sony/Nikon/Canon. EU AI Act Art. 50 enforcement Aug 2026. Gold standard.
• Adobe Firefly Provenance (Adobe) — Firefly auto-applies Content Credentials to any 100%-AI-generated asset, including model + version, issuer, and editing actions. Per-asset provenance with editing chain visible.
• Microsoft 365 AI Watermarks (Microsoft) — Microsoft 365 AI watermark policy (2026): audio shipped, video rolling out March 2026, plus metadata embedding (model used, app, timestamp). Document-level provenance at scale.
• Google SynthID (Google) — Invisible watermark embedded in Imagen and Veo outputs, verifiable with the SynthID detector. Strong consumer-scale provenance — survives common edits and re-encodings.
• OpenAI Image Provenance (OpenAI) — All DALL-E 3 and Sora outputs carry C2PA credentials by default — provenance that travels with the artefact. The media world ships this at file level; the prototype shows the same instinct applied span-by-span to co-authored text, which is still emerging.

FAQ

When should I use the AI Provenance pattern?

When you co-write with AI, machine-drafted and human-written text read the same — so the useful signal is which parts a human has actually reviewed.

When should I avoid the AI Provenance pattern?

Throwaway or low-stakes text where no one will ever need to tell an AI draft from vetted prose.

What problem does AI Provenance solve?

In a document you co-write with AI, machine-drafted and human-written text read identically — so you cannot tell which parts still need a human to check them.

Why is chat the wrong fit for this?

Chat hands back one finished block with no record of what was reviewed; once it leaves the chat, AI-drafted and human-vetted text look exactly the same.

Related patterns

• Extends: Disclosure — Disclosure is the banner-level signal; Provenance is the per-fragment metadata that survives copy-paste.
• Alternative to: Citation Trail — Trace claims back to sources vs trace fragments back to authors. Same ground-truth-attribution instinct, different scopes.
• Often paired with: Action Audit & Undo — Audit logs every AI action; Provenance marks every AI fragment. Two halves of the trail.