Make AI involvement understandable.

Existing industry patterns

The Shape of AI’s Disclosure pattern distinguishes AI actors, actions and outputs. It is part of a collection already represented in the atlas’s curated AI references. Its Watermark pattern also considers content-origin signals. This atlas treats visible disclosure, embedded watermarking and provenance as separate mechanisms.

Google PAIR’s Explainability + Trust chapter connects explanations to capabilities, data and context. Microsoft’s Human-AI Interaction guidelines help frame expectations and correction. Together these inform the decisions below; they do not validate a faith product’s labels or benefits.

Four layers to keep distinct

1. Identity: is the speaker an AI system or a real person?
2. Operation: did AI generate, edit, summarise, translate, retrieve, recommend or narrate?
3. Artifact origin: which text or media came from a publisher, a person or a model?
4. Evidence and review: what sources support a claim, what was checked, and by whom?

A citation does not substitute for authorship disclosure. “Human-reviewed” needs a real reviewer and an accurately stated review scope. Model branding and technical metadata can sit in details; they do not explain all four layers.

Place information where it is needed

Before use: name the operation on the entry control and state relevant input scope before personal data is requested. During use: retain sender identity and show whether generation or a real human handoff is happening. On output: keep the label adjacent to its content, including in a saved-item library. On export: preview the artifact as a recipient will encounter it.

Use readable text with accessible names. Avoid colour alone, unexplained sparkle symbols and hover-only details. On audio-only surfaces, provide a concise audible introduction where origin could be misunderstood; preserve origin information in playback details and downloaded artifact information.

Visible labels, watermarks and provenance

A visible label helps the current reader interpret origin. A watermark encodes a signal in the media. Provenance records creation or modification information. C2PA’s specification describes tamper-evident asset provenance, not a certification of the content’s truth. This entry reviews its introduction and scope rather than its full implementation requirements.

Use supported machine-readable mechanisms alongside meaningful visible or audible disclosure. Their absence does not prove human authorship. A destination may strip metadata; copied text and screenshots may also lose information. Private prayer intentions, prompts, recordings and conversation history should not become public provenance by default.

Faith-specific interpretation and authority

An AI label should neither suggest divine origin nor quietly reduce a verified translation to a model output. Name generated reflection as reflection, prayer assistance as an editable draft, and source-based synthesis as an interpretation. Keep authored prayers and human teaching attributed. Do not use clergy titles, likenesses or conversational voices to imply accountability that the service does not provide.

Disclosing AI is separate from permission to use personal data and from providing safe, accurate content. “AI can make mistakes” does not repair invented Scripture, unsupported citations or harmful pastoral claims. Read the legal research map for the separate role-dependent assessment of transparency and data obligations.

What the app evidence establishes

VSB’s verse-study capture shows an AI-generated summary label and numbered source markers. These are distinct visible signals; citation accuracy still needs testing. Lumen’s onboarding offers engagement choices, but that image does not demonstrate output disclosure. Product setup and artifact attribution should be evaluated separately.

Test understanding, not badge recognition

1. Show a first-time reader a mixed Scripture and explanation view: can they identify each origin?
2. Ask what the label says AI did and what a reviewer actually checked.
3. Repeat through a deep link, shared card, saved draft, screen reader and audio-only entry.
4. Test an edited artifact, an unavailable source and a metadata-stripping destination.
5. Check whether the reader can inspect sources, correct the output, choose an offered non-AI route or end the experience.

The two pattern families below include original wireframes and decision conditions. Their effectiveness remains to be evaluated with relevant readers.