# OSHUN V1 AI Disclosure Copy Guidance

Status: Accepted v1.0  
Owner: OSHUN Design + Product + Engineering  
Last Updated: March 26, 2026  
Implements: `V1-UX-039`  
Related:

- `docs/releases/v1/design/ux-principles.md`
- `docs/releases/v1/design/readability-standards.md`
- `docs/releases/v1/specs/web-app-admin-v1-prd.md`
- `docs/adr/ADR-0066-oshun-disclosure-and-synthetic-indicator-ui-behavior.md`
- `docs/adr/ADR-0069-oshun-assistant-persona-switching-ui-behavior.md`
- `apps/oshun/web/src/design-system/disclosure.ts`
- `apps/oshun/web/src/design-system/components/AssistantDisclosureStrip.tsx`
- `apps/oshun/web/src/design-system/components/AssistantMemoryState.tsx`
- `apps/oshun/web/src/components/assistant/AssistantPanel.tsx`
- `docs/releases/v1/specs/todos.md`

## 1. Purpose

This document defines the canonical copy rules for AI disclosure language across
OSHUN V1.

Disclosure copy exists to answer four questions quickly and truthfully:

- what is AI here
- what is synthetic here
- what memory or continuity scope is active
- where human review, if any, actually begins

The goal is not legalistic warning copy.

The goal is plain, timely, low-drift language that prevents mistaken human
attribution, hidden synthetic behavior, and vague memory claims.

## 2. Current Repo Signals, External Scan, And Gap

The current repository already has strong implementation signals:

- `disclosure.ts` defines one shared indicator set for `ai_assistant`,
  `memory_off`, `memory_session`, `memory_profile`, `synthetic_voice`,
  `synthetic_avatar`, and `synthetic_media`.
- The current assistant disclosure summary already says
  `Responses come from OSHUN Assistant, not a human operator.`
- `AssistantMemoryState.tsx` already resolves memory into explicit, legible
  states instead of abstract policy text.
- `ADR-0060` requires explicit identity, visible memory scope, visible synthetic
  modality signals, and plain-language companion copy.
- `ADR-0063` explicitly separates persona from disclosure.
- The PRD requires clear AI disclosure, memory-state visibility, synthetic
  modality indicators, and a user-accessible explanation of what is remembered
  and why.

External implementation research points in the same direction:

- EU AI Act Article 50 requires people to be informed that they are interacting
  with AI and says the disclosure must be clear and distinguishable at the first
  interaction or exposure.
- The European Commission's AI Act guidance frames these disclosures as a trust
  and anti-deception measure, especially for misinformation, impersonation, and
  consumer deception risks.
- OpenAI and Google Gemini both use short, front-door limitation copy such as
  "can make mistakes" and "check important info" rather than long legal
  paragraphs.
- FTC impersonation guidance reinforces the need to avoid copy that falsely
  implies a human, business, or authority connection.

Inference from those sources plus the repo state:

- OSHUN already has the visual and structural disclosure model.
- OSHUN does not yet have one canonical copy rule set explaining how disclosure
  should sound, what exact vocabulary it should use, and which phrases are
  disallowed.

## 3. Canonical Disclosure Thesis

OSHUN disclosure copy should be:

- immediate
- plain-language
- concrete
- calm
- non-defensive

OSHUN disclosure copy should not be:

- euphemistic
- theatrical
- buried in settings
- vague about memory
- vague about synthetic media
- falsely human-sounding

The canonical pattern is:

1. identify the AI system plainly
2. identify any active synthetic modality plainly
3. identify memory scope plainly
4. identify human review only when it is actually present

## 4. Core Rules

### 4.1 Say `AI` Directly

Use `AI assistant`, `AI-generated`, and `synthetic` directly when those things
are true.

Do not replace those terms with softer substitutes such as:

- `guide`
- `companion`
- `digital expert`
- `smart helper`
- `virtual human`

Those terms can describe product posture, but they cannot carry the disclosure
burden on their own.

### 4.2 Disclosure Must Happen At First Contact

AI identity and synthetic modality disclosure must appear at the first
interaction or first exposure on the active surface.

This mirrors both the PRD requirement and the external regulatory direction:
users should not need to infer they are interacting with AI after the fact.

### 4.3 Identity, Persona, Memory, And Grounding Are Separate Concepts

Do not collapse these into one sentence.

- identity answers whether the experience is AI-mediated
- persona answers which guidance lens or mode is active
- memory answers what continuity scope is active
- grounding answers whether evidence or sources shaped the output

Example:

- good: `AI assistant`, `Session memory`, `Grounded response`
- bad:
  `Nisaba is active, so your trusted scholar is using what it knows about you`

### 4.4 Memory Copy Must Describe Scope, Not Vibes

Memory language must tell the user what the system can use right now.

Avoid abstract phrasing such as:

- `personalized experience enabled`
- `continuity is on`
- `smart context available`

Prefer concrete scope language:

- `No saved or session memory is used beyond the current prompt.`
- `The assistant can use the current conversation while this session is open.`
- `The assistant can use approved saved memory alongside the current session.`

### 4.5 Synthetic Modality Copy Must Name The Modality

If voice is synthetic, say `Synthetic voice`.

If avatar motion or likeness is synthetic, say `Synthetic avatar`.

If the artifact itself is AI-generated, say `AI-generated media` or
`Synthetic media`, depending on the shared component contract.

Do not rely on fuzzy umbrella language such as:

- `enhanced media`
- `studio render`
- `digital presenter`
- `generated experience`

### 4.6 Human Review Must Never Be Implied

Do not imply a human reviewed, approved, authored, or monitored the output
unless that actually happened.

Avoid phrases such as:

- `our team thinks`
- `expert-backed answer`
- `reviewed for you`
- `we verified this`

unless the underlying workflow actually guarantees that claim.

When human review does exist, say exactly what it means:

- `Queued for human review`
- `Reviewed by an editor before publication`
- `Escalated to a human operator`

### 4.7 Limitation Copy Should Be Short And Useful

When the surface needs a limitation reminder, use direct lines similar to the
strongest current product patterns in the market:

- `AI can make mistakes. Check important info.`
- `Double-check time-sensitive or high-stakes details.`

Do not stack five warnings into every disclosure strip. Persistent disclosure
should stay compact, while deeper explanation can live in an explainer surface,
help state, or settings panel.

### 4.8 Do Not Sound Defensive Or Apologetic By Default

Disclosure is a product truth, not an apology.

Avoid copy like:

- `Please forgive any errors because this is AI`
- `We hope you understand this may not be perfect`

Use matter-of-fact language instead.

## 5. Canonical Vocabulary

### 5.1 Preferred Terms

- `AI assistant`
- `AI-generated`
- `synthetic voice`
- `synthetic avatar`
- `synthetic media`
- `memory off`
- `session memory`
- `profile memory`
- `current thread`
- `approved saved memory`
- `human review`
- `escalate to human operator`

### 5.2 Terms To Avoid

- `almost human`
- `feels human`
- `natural companion`
- `live expert` when the responder is AI
- `trusted advisor` as the sole disclosure mechanism
- `smart personalization` without a memory explanation
- `studio quality voice` without stating it is synthetic
- `verified` when the system only generated or summarized

## 6. Canonical Strings For Current Shared Indicators

These strings are the current canonical defaults because they already exist in
the shared disclosure contract and align with the PRD.

### 6.1 Identity

- Label: `AI assistant`
- Description: `Responses come from OSHUN Assistant, not a human operator.`

### 6.2 Memory Off

- Label: `Memory off`
- Title: `Prompt-only memory`
- Description: `No saved or session memory is used beyond the current prompt.`

### 6.3 Session Memory

- Label: `Session memory`
- Title: `Current thread only`
- Description:
  `The assistant can use the current conversation while this session is open.`

### 6.4 Profile Memory

- Label: `Profile memory`
- Title: `Approved memory available`
- Description:
  `The assistant can use approved saved memory alongside the current session.`

### 6.5 Synthetic Voice

- Label: `Synthetic voice`
- Description: `Voice replies are synthetic when voice output is enabled.`

### 6.6 Synthetic Avatar

- Label: `Synthetic avatar`
- Description: `Avatar motion, likeness, or voice is synthetic.`

### 6.7 Synthetic Media

- Label: `Synthetic media`
- Description: `The artifact itself is AI-generated media.`

## 7. Surface-Specific Rules

### 7.1 Assistant Shell And Persistent Strips

Use the shortest truthful phrasing possible.

Requirements:

- identity chip always present
- memory chip always present
- synthetic modality chip added when active
- summary sentence reads like a plain explanation, not a slogan

Good:

- `AI assistant`
- `Session memory`
- `Synthetic voice`

Bad:

- `Your guide`
- `Continuity on`
- `Enhanced voice`

### 7.2 Voice And Avatar Sessions

Voice and avatar sessions need stronger disclosure because mistaken human
attribution risk is higher.

Rules:

- the word `synthetic` must remain visible on the active surface
- the named persona or asset may complement the disclosure, but cannot replace
  it
- if calls or sessions are recorded, say so separately and plainly

Good:

- `Synthetic voice`
- `Synthetic avatar`
- `Calls may be recorded to improve the service.`

Bad:

- `Speak with Amara now`
- `Your host is live`
- `Studio voice enabled`

### 7.3 Generated Artifacts And Public-Facing Media

When the output itself is generated or manipulated, the artifact needs its own
disclosure and should not depend only on shell-level disclosure.

Rules:

- mark the artifact, not just the surrounding page
- use `AI-generated` or `artificially generated` when public-trust stakes are
  high
- do not hide the label in metadata drawers only

### 7.4 Memory Controls And Privacy Settings

Memory settings should answer:

- whether memory is on
- which scope is active
- what is remembered
- why it is remembered
- how to change it

Avoid:

- abstract privacy-branding language
- claims about personalization without naming the source of that personalization

### 7.5 Human Escalation And Review Flows

If a human enters the flow, say what changed.

Good:

- `Escalated to human operator`
- `Pending human review`
- `An editor reviewed this before publication`

Bad:

- `Handled with care`
- `Reviewed`
- `Safe to trust`

## 8. Do And Do Not Examples

### 8.1 Identity

- do: `Responses come from OSHUN Assistant, not a human operator.`
- do not: `Your support specialist is here.`

### 8.2 Memory

- do:
  `The assistant can use the current conversation while this session is open.`
- do not: `We remember what matters.`

### 8.3 Synthetic Voice

- do: `Voice replies are synthetic when voice output is enabled.`
- do not: `Voice mode uses a natural guide.`

### 8.4 Synthetic Avatar

- do: `Avatar motion, likeness, or voice is synthetic.`
- do not: `Meet your on-screen teacher.`

### 8.5 Generated Media

- do: `This artifact is AI-generated media.`
- do not: `This scene was assembled for you.`

### 8.6 Limitation Reminder

- do: `AI can make mistakes. Check important info.`
- do not: `Outputs are for entertainment purposes only unless otherwise stated.`

## 9. Definition Of Done

AI disclosure copy is done for a surface only when:

- AI identity is explicit at first interaction
- memory scope is explicit in plain language
- synthetic modalities are named explicitly when active
- human review is stated only when true
- persona language does not replace disclosure language
- artifact-level synthetic labels appear where the artifact itself is generated
- copy remains compact on primary surfaces and expandable on deeper explainer
  surfaces
