AI Disclosure Policy Skill

Every company now ships AI-generated content somewhere — support replies, marketing images, chatbot conversations, synthetic voices — and most have no rule for when to say so. Meanwhile transparency regulation is arriving (the EU AI Act's transparency obligations for chatbots, synthetic media, and deepfakes being the headline example, with obligations phasing in through 2026–2027), and the trust cost of an undisclosed AI surface being discovered is higher than the disclosure ever was. This skill produces the policy: what you label, where, in what words — with the honest line that final regulatory judgment belongs to your lawyer, and this document is what makes that conversation short.

What This Skill Produces

• A surface inventory: every place AI-generated content reaches users or the public, with today's disclosure state
• A disclosure matrix: per surface — required (regulatory), expected (platform/industry norm), or chosen (trust) — with the reasoning
• Label copy ready to ship: UI strings, footer lines, image/video marks, chatbot self-identification wording
• The review triggers: what changes (new surface, new market, new regulation phase) forces a policy re-read, and who owns it

Required Inputs

Ask for (if not already provided):

• Where AI output ships today or soon: chatbots, support, marketing content, images/video/voice, code, docs — and which are fully automated vs human-reviewed
• Markets served (EU exposure changes obligations) and industry (regulated sectors add rules)
• Existing policy fragments ([[ai-usage-policy]] covers internal use — this skill covers outward disclosure; link them, don't duplicate)
• Risk posture: minimum-compliance or trust-differentiator

Process

1. Inventory before policy. List every AI-touching surface, then the ones the user forgot: auto-generated email, AI-assisted support macros, synthetic voices on calls, generated product imagery, auto-summaries in the product. For each: fully-AI, AI-drafted-human-approved, or AI-assisted — the disclosure answer differs by degree of human control.
2. Sort into required / expected / chosen. Required: where a regulation plausibly applies — chatbots that could be mistaken for humans, synthetic media, emotionally targeted content (flag these for counsel; cite the regulation family, not invented article numbers). Expected: platform rules and industry norms (ad platforms, app stores increasingly require labels). Chosen: where labeling is optional but discovery-risk or brand values argue for it. State the reasoning per row — a policy without reasons decays.
3. Write labels people won't hate. Honest, short, non-groveling: "AI-assisted, human-reviewed" beats a paragraph of throat-clearing. Chatbots self-identify at conversation start, not in a footer. Human-approved content can say so — the disclosure spectrum has two ends.
4. Decide the edge cases explicitly: AI-drafted-human-edited text (the big one — set a threshold and say it), internal content that leaks, user-facing personalization, A/B tests of the labels themselves (don't).
5. Wire the triggers. New surface, new market, automation-degree change, regulation phase-in dates → named owner re-reviews. Policy without a re-review trigger is a screenshot, not a policy.

Output Format

## Where AI ships today
| Surface | Degree (full / drafted / assisted) | Disclosed today? |

## Disclosure matrix
| Surface | Required / Expected / Chosen | Reasoning | Label |

## Label copy (ready to ship)
[Exact strings per surface type]

## Edge-case rulings
[The threshold decisions, stated plainly]

## Review triggers & ownership
[What forces a re-read, who owns it, standing counsel questions]

Quality Checks

• The inventory surfaced at least one AI surface the user didn't list
• Every matrix row carries reasoning; "required" rows name the regulation family and carry the flag-for-counsel marker — no invented article citations
• Label copy is shippable as-is: short, honest, located where users actually are (chatbot labels at the top, not the terms page)
• The AI-drafted-human-edited threshold is decided, not deferred
• The not-legal-advice line is present and the counsel-question list makes the legal review cheap

Anti-Patterns

• Do not assert specific legal conclusions ("Article X requires you to…") — identify plausibly-applicable obligations and route to counsel
• Do not write labels as apologies — disclosure done confidently is a trust feature
• Do not produce one blanket rule; the matrix exists because a support macro and a synthetic voice are different obligations
• Do not duplicate [[ai-usage-policy]] — internal use rules live there; this is outward-facing disclosure