Disclosure

Framing

The problem

AI-generated content circulates outside the system as if it were human-authored — recipients cannot calibrate trust or detect synthetic provenance.

The pattern

Mark every AI-touched surface or output with a visible, persistent disclosure that survives copy-paste, screenshot, and downstream sharing.

Why chat breaks here

Chat transcripts have no surface for marking the assistant contribution once it leaves the chat — the labelling has to come at the artifact level.

Risks

Heavy-handed disclosure becomes legal-disclaimer noise users learn to tune out; weak disclosure satisfies regulators without informing recipients.

Avoid when

The output is internal-only and disclosure adds friction without recipient value.

Use when

AI-touched outputs leave the system into untrusted contexts where recipients need to know AI was involved.

DOPE evaluation

Directability

Users can expand any disclosure to see what AI contributed to the artifact

Observability

AI presence is visible without expansion or hover — recipients see it at a glance

Predictability

The disclosure surface is consistent across the product so users learn to recognise it

Explainability

Each disclosure links to the model, prompt context, and time that produced the contribution

In the wild

• Microsoft 365 Copilot · "Created with Copilot" labels (Microsoft) — Documents drafted with Copilot get a header label and an in-doc footer note. Survives copy-paste into another doc; visible to anyone the file is shared with. Loosely couples to C2PA where the file format supports it.
• Google SynthID (Google DeepMind) — Invisible cryptographic watermark embedded in Gemini-generated images, audio, and video. Robust to cropping and re-encoding — detection happens at the model side. Banner-level signal without visible chrome.
• C2PA Content Credentials (C2PA · Adobe / OpenAI) — Open standard: cryptographic provenance metadata attached to images and video at generation. Sora, Adobe Firefly, Leica cameras all sign with C2PA. Disclosure as portable signal, verified outside the originating product.
• Apple Intelligence · Image Playground provenance metadata (Apple) — Images generated with Image Playground carry AI-provenance metadata written into the file itself, so the disclosure travels with the artifact when it is saved or shared. Apple's real disclosure mechanism — at the file level, not the editing surface.

FAQ

When should I use the Disclosure pattern?

AI-touched outputs leave the system into untrusted contexts where recipients need to know AI was involved.

When should I avoid the Disclosure pattern?

The output is internal-only and disclosure adds friction without recipient value.

What problem does Disclosure solve?

AI-generated content circulates outside the system as if it were human-authored — recipients cannot calibrate trust or detect synthetic provenance.

Why is chat the wrong fit for this?

Chat transcripts have no surface for marking the assistant contribution once it leaves the chat — the labelling has to come at the artifact level.

Related patterns

• Alternative to: AI Provenance — Banner-level visibility at a glance vs per-fragment marking inside the document.
• Often paired with: Confidence Signals — Disclosure says it is AI; Confidence Signals say how sure it is. Together they let recipients calibrate trust.
• Often paired with: Footprints — Disclosure addresses the recipient; Footprints addresses the auditor — different audiences, same lineage.