AI Visibility Policy

Date: 23 March 2026

1. Principles

The Knowledge Hub’s design philosophy states: “AI is invisible infrastructure, not a visible product feature” and “Helping you get organised, not learning from you.” This policy codifies that principle into actionable rules for every surface of the application.

1.1 The Four Visibility Rules

Rule 1: AI processing is invisible. Classification, embedding, entity extraction, summarisation, quality scoring, freshness calculation, deduplication, and layer inference are backend operations. Users never see the mechanism. The platform “just works.”

No “AI-powered” badges
No “three-pass AI pipeline” descriptions
No model names (e.g. “claude-sonnet-4-6”) in user-facing text
No “AI-generated” labels distinguishing human from machine content
No Sparkles icons used to signify AI involvement

Rule 2: AI-derived outputs are presented as platform features. Quality scores, freshness states, classification confidence, summaries, and digests are presented as native platform capabilities. They are not labelled as “AI” any more than a spell-checker labels its suggestions as “AI.”

Quality score = “Quality: 78” (not “AI Quality Score”)
Summaries = “Summary” (not “AI-generated summary”)
Digests = “Content Digest” (not “AI-generated digest”)
Classification = “Domain: Health & Safety” (not “AI classified as…”)

Rule 3: Claude bridge actions are visible and honestly labelled. Buttons and prompts that take users TO Claude (via claude.ai, Claude Desktop, or Claude Code) are acceptable because the platform’s role is to bridge users into Claude products. These are integration touchpoints, not AI branding.

“Continue in Claude” / “Open in Claude” = acceptable
“Review with Claude” = acceptable (it opens Claude)
“Take action” with external link icon = acceptable
“AI-powered workspace” = not acceptable (AI branding, not a bridge)

Rule 4: No in-app chat sidebar. Knowledge Hub does not embed a generic chat surface. All AI-assisted interactions either run as invisible background processing (Rule 1) or route the user to Claude via the ClaudePromptButton bridge (Rule 3). The CopilotKit sidebar that previously implemented Rule 4 was fully removed in S109.

1.2 Decision Framework

When adding a new feature, ask:

Does this feature use AI internally? Keep it invisible. Label the output, not the mechanism.
Does this feature send the user to Claude? Label it as a Claude bridge action. Use the ClaudePromptButton component with an action-oriented label.
Does this feature display AI-generated content alongside human content? Present both as “content.” Do not distinguish provenance. If editors need to know, use a subtle metadata footnote (not a badge).
Does this feature expose AI confidence or scoring? Present as a platform metric. “Quality: 78” not “AI confidence: 78%.“

2. Future Guidance

When Adding New Features

Never use the Sparkles icon to indicate AI involvement
Never label content as “AI-generated” or “AI-powered”
Never expose model names to users
Always present AI-derived outputs as platform features
Always use action-oriented labels on bridge buttons (“Take action”, “Review”, “Create”) rather than “Ask Claude”
The word “Claude” is acceptable only when describing a destination (“Open in Claude”, “Continue in Claude”) or in integration settings

When Claude Branding IS Appropriate

Settings/Integrations page: configuring the Claude connection
Bridge buttons that redirect to Claude: “Open in Claude”, “Review with Claude”
Documentation about the MCP integration

Admin-only Provenance surface: model names, token counts, cost pills, and classification reasoning may be rendered on /provenance tabs (Per-item, Pipeline Health, Audit, Cost, Disputes) because these exist for auditing and debugging, not user experience. The same data MUST NOT appear on any surface reachable by viewers, editors, or anonymous users. P0-3b is the companion change that removes these fields from user-facing item headers — with this carve-out, the removal is understood as a relocation, not a deletion.

Editor+admin Source Information surface: The classification_confidence metric may be rendered as a plain text percentage inside the Source Information accordion on /item/[id] for users with admin or editor roles. It MUST NOT be shown to viewers or anonymous users, MUST NOT be accompanied by AI branding (no “AI confidence” label, no Sparkles icon, no colour-coded badge), and MUST NOT be accompanied by AI-mechanism fields (model names, reasoning, tokens, cost). This is a narrow, scoped amendment of the P0-3b removal, introduced for editor/admin quality review use cases. See .planning/.archive/.specs/source-information-spec.md §9 for the decision record.

AI acceptance criteria are internal policy, not UI: The five criteria (automated, scheduled, invisible, measurable, actionable) are principles documented here for implementation guidance. They are not rendered anywhere in the product UI — the banner that previously displayed them was explicitly removed from scope.

Note: A direct “Claude” link previously lived in the site header as an allowed exception under this section. It was removed in April 2026 once the MCP connector setup in Settings → Connections provided a higher-quality bridge path, and once the header needed space for a Sign out control. The header exception is retired — do not reintroduce it without a fresh discussion against the four principles in §1.1.

When Claude Branding IS NOT Appropriate

Section headings (“Take Action with Claude”)
Feature descriptions (“AI-powered workspace”)
Content badges (“AI-generated”)
Processing descriptions (“three-pass AI pipeline”)
Button labels that could be action-oriented instead (“Ask Claude” vs “Take action”)

3. AI Framing — Avoid / Use Instead

The product’s framing rule is: “Helping you get organised, not learning from you.” AI is invisible infrastructure — like electricity. The product never announces it.

Avoid	Use Instead
”Learning from your documents"	"Setting up your content library"
"AI is analysing…"	"Extracting…” / “Finding…"
"AI-powered”	Do not label it
”Intelligent” / “Smart”	Do not label it

This table is the canonical AI framing reference for all UI copy. It was extracted from ads-v1.md §7.1 on 7 April 2026 (S151 Task 17b) before that document was archived. For the broader UX framework see docs/reference/ux-principles.md.