Spec — AI generated content disclaimer

Compliance: EU AI Act art. 50 (entrando em vigor 2026+); LGPD art. 9 (transparency); GDPR recital 71. Pattern reference: Microsoft Teams AI label spec; MIT Sloan labels research; UK ICO AI guidance.

Princípios

1. *lwayslabeled*— todo AIgenerated content em produto distribuído carrega disclaimer. Não há "trusted enough to skip".
2. *isk-proportional*— tier escala com risco da decisão induzida pelo content.
3. *18n mandatory*— copy disponível em todo locale shipado; fallback en-US.
4. *udit-trail*— dismissal decisions persistidas para forensics.
5. *ditorial copy owner-curated*— IA não edita os textos do disclaimer autonomamente (memory feedback_kds_owner_curated_content).

R1 — Three tiers

R1.1 — Tier 1: Labelonly (default, lowrisk)

Chip subtle abaixo do bubble:

🤖 Generated by AI — verify

• Color: text-muted (per themes/color-roles.kmd).
• Typography: label-small (per themes/typography.kmd R1).
• Click: expands tooltip explicativo (key ai.disclaimer.tooltip.low).

Applied to: text responses, suggestions, casual content (chat, notes, drafts).

R1.2 — Tier 2: Label + modal (medium-risk)

Same chip, mas firsttimein-session opens explanatory modal:

┌────────────────────────────────────────────┐
│  About this AI-generated content           │
│                                            │
│  This response was generated by an AI     │
│  model. It may contain inaccuracies or    │
│  bias. Please verify critical information  │
│  before acting on it.                      │
│                                            │
│  Source model: Claude Opus 4.7             │
│  Limitations: knowledge cutoff 2026-01    │
│                                            │
│  [Don't show again]   [Got it]             │
└────────────────────────────────────────────┘

Applied to: factual/technical responses, code suggestions, research summaries.

Dismissal persistida per (koder_user_id, workspace_id, tier); "Don't show again" suppresses future modals for that tier — chip ainda presente.

R1.3 — Tier 3: Label + banner + confirmation (high-risk)

Banner persistente ACIMA do bubble + checkbox antes do user agir:

┌─────────────────────────────────────────────────┐
│  ⚠ HIGH-RISK AI OUTPUT                          │
│  This response may affect important decisions   │
│  (deploy / financial / medical / legal). You    │
│  MUST verify it independently before acting.    │
│                                                 │
│  [ ] I understand this is AI-generated and     │
│      will verify before acting                  │
└─────────────────────────────────────────────────┘

Required actions on the bubble (Deploy, Pay, Sign, Submit) blocked until checkbox toggled.

Applied to: deployment actions, financial transactions, medical/legal advice, code that touches production.

R2 — Risk derivation

risk field comes from gateway response. Derivation:

Product can override via gateway request annotation koder_risk_tier_min: "high" (e.g., Kortex production-deploy view sempre forces tier 3).

Cross-link: gateway response includes _meta.koder.risk: "low"|"medium"|"high".

R3 — Dismissal persistence

Schema (kdb-kv):

key: ai_disclaimer_dismissal:<koder_user_id>:<workspace_id>:<tier>
value: {
  dismissed_at: ISO8601,
  expires_at: ISO8601 | null,
  granted_by: <koder_user_id>
}

• Tier 1 (label): NÃO dismissable. Sempre presente.
• Tier 2 (modal): dismissable per tier-2 globally; modal hides; chip stays.
• Tier 3 (banner+confirmation): NÃO dismissable globally; checkbox per-message obrigatório.

Autoexpire: Tier 2 dismissal expires após 90 dias (redisclosure refresh).

R4 — Audit log

Toda dismissal + toda Tier 3 confirmation emit audit event:

{
  event_type: "ai.disclaimer.acknowledged" | "ai.disclaimer.dismissed",
  tier: 1 | 2 | 3,
  koder_user_id: ...,
  workspace_id: ...,
  conversation_id: optional,
  message_id: optional,
  model: <source model>,
  timestamp: ISO8601
}

Retention: respeita policies/identity-data-retention.kmd R2 (auth_events 24m success / 12m failure).

R5 — Risk derivation override (product opt-in)

Products que sabem ser highriskby-default declarar via gateway request:

# Kortex deploy view
gateway.complete({
  prompt: ...,
  koder_risk_tier_min: "high",  # forces tier 3 for ALL responses in this view
})

Once set, ALL bubbles in the session inherit min tier 3 regardless of gateway-derived tier.

Cross-link: products specify this in their product spec (e.g., Kortex deploy ratifies tier 3 default).

R6 — Surface bindings

R7 — i18n

Editorial copy. NÃO editable por IA autonomamente per feedback_kds_owner_curated_content.

R8 — Acessibilidade

• Label chip: role="status" (não interactive em tier 1; interactive em tier 2 quando clickable).
• Modal: role="dialog" aria-modal="true".
• Banner tier 3: role="alert" (announces immediately on render).
• Checkbox tier 3: <input type="checkbox" aria-required="true" aria-describedby="disclaimer-body">.
• Action buttons (DeployPaySign) que dependem do checkbox: aria-disabled="true" antes; aria-disabled="false" após.
• Screen reader: tier 3 banner anuncia "High-risk AI output. Verification required before action."

R9 — Per-preset variation

Style do disclaimer respeita preset (color e shape), MAS conteúdo text + tier behavior é fixed.

T-suite

• *1*Tier 1 label: assistant bubble com risk low → chip "Generated by AI — verify" presente.
• *2*Tier 1 not dismissable: spec validation prevents disabling tier 1.
• *3*Tier 2 modal firsttime: first mediumrisk response → modal aparece; dismiss → modal hides; chip stays.
• *4*Tier 2 dismissal persisted: navegar away + voltar → modal NÃO reaparece (within 90d).
• *5*Tier 2 expiry: avançar clock 91d → modal reaparece on next medium response.
• *6*Tier 3 banner: high-risk response → banner ABOVE bubble + checkbox.
• *7*Tier 3 action gate: Deploy button starts aria-disabled; check checkbox → button enables.
• *8*Tier 3 NOT globally dismissable: dismiss attempt rejected.
• *9*Audit log emits: T3T5T7 each emit ai.disclaimer.* events.
• *10*Product override: product sets koder_risk_tier_min: "high" → ALL bubbles in view tier 3.
• *11*i18n switch: change locale to pt-BR → all keys translated.
• *1*Missing disclaimer in assistant bubble: violates R1.1 → static check / lint warning OR runtime assertion.
• *2*Tier 3 action without checkbox: attempt to Deploy without check → action blocked + screen reader announce.

Cross-link

• Companion policy: policies/ai-content-disclosure.kmd — comportamento (quando + onde aplicar)
• Consumer: chat-message-bubble.kmd R2 hosts the label
• Colortypography: `themescolor-roles.kmd, themes/typography.kmd`
• Retention: policies/identity-data-retention.kmd
• Storage: policies/multi-tenant-by-default.kmd
• Compliance basis: EU AI Act art. 50 · LGPD art. 9 · ICO AI guidance · MS Teams AI label spec