carbon

IBM Carbon evidence profile

Type: System profile
Status: Active
Canonical for: Factual interpretation of preserved IBM Carbon and Carbon Builder evidence
Depends on: corpus/carbon/

Evidence labels: Documented, Implemented, Inferred, Emerging discussion.

What Carbon is changing

Carbon treats AI support as both a product-facing extension and an agent-facing delivery mechanism. These are related but separately implemented: Carbon for AI changes components, styling and interaction; Carbon MCP exposes existing system knowledge to software agents.

Relevance correction for the primary research question

Carbon for AI and the AI Label are evidence about designing AI experiences. They are not, by themselves, evidence that Carbon components are machine-consumable.

The central agent-consumption evidence is the public carbon-builder skill preserved in the Carbon MCP repository:

• Intent: explicit selection rules distinguish classification (Tag) from state (IconIndicator/ShapeIndicator) and define valid tab-family and Modal compositions.
• Structure: MCP results expose component IDs, framework, imports, variants, Storybook links and documentation fields; code and guidance are reconciled by canonical identity.
• Retrieval: the skill mandates Discover → Canonicalize → Target, code_search before imports, and targeted docs_search for usage/accessibility guidance.
• Constraint/evaluation: hard rules, forbidden substitutions, accessibility requirements and validation checklists constrain a compliant agent; public semantic-selection evaluations are not present.

Carbon Builder skill, selection and composition rules, query protocols, result validation.

AI product design

• Documented: Carbon for AI is formally described as an extension that gives AI instances a distinct visual and behavioral identity, centered on transparency and explainability. It includes stable AI style elements and AI variants, plus AI Chat. Preserved Carbon for AI guidance
• Implemented: AILabel, AILabelContent, and AILabelActions are public React exports with a typed AILabelProps surface. Web Components export CDSAILabel and an action button. React exports, web-component exports
• Implemented: AI presence is integrated into ordinary controls—not only net-new AI widgets. Preserved guidance names checkbox, form, select, data table, modal, tag, date picker, number input, text input, dropdown, radio button and tile. Code and stories show label/decorator integration in components such as Dropdown and TextInput. Carbon for AI component list, Dropdown story
• Documented: AI presence can be focused or broad. Placement rules differ for containers, inputs, table cells/rows/columns/tables, and forms. The label must not become a generic AI action trigger; regenerate uses a separate action. AI label usage
• Documented and implemented: Editing AI-suggested content removes AI styling and replaces the label with a revert action; APIs include onRevertClick, revertActive, and revertLabel. Guidance, public API snapshot
• Implemented: Carbon AI Chat has typed message, streaming, response, error, feedback and customization machinery. Its scroll controller explicitly distinguishes active streaming, stream completion, user scroll-away and Safari anchoring. AI Chat source, stream scroll controller
• Documented: AI Chat is backend-agnostic and offers React and web-component use. AI Chat overview

Evidence is strong for disclosure, explainability, generated-content states, correction/revert, chat and streaming. Public evidence is much weaker for citations, consequential approvals, agent plans, long-running background work, permission requests, memory and audit history.

Accessibility and content

• Implemented/documented: The AI label is a keyboard-focusable trigger whose popover can be opened and dismissed while focus remains on the trigger. When embedded in an input it adds a second tab stop; revert restores the prior AI value. Automated default/advanced-state tests and manual screen-reader testing are reported, while keyboard test data is marked unavailable for the stable label. AI label accessibility
• Documented: AI styling is constrained by contrast/readability requirements; glow spread is deliberately limited. Existing component accessibility pages retain keyboard, screen-reader and automated-test status. Carbon for AI accessibility
• Documented: The explainability popover has a content template: overview, supporting details, artifacts/resources and additional actions. AI abbreviations are localizable. AI label content
• Inferred: Carbon’s most rigorous accessibility evidence covers the label and modified conventional controls. AI Chat’s streaming and live-update accessibility require a dedicated follow-up audit; the presence of streaming code does not itself demonstrate screen-reader quality.

Architecture and machine readability

• Implemented: Carbon adds public AI design tokens (aiBorderStart, aiBorderEnd, aiPopoverBackground, caret and shadow tokens) to the normal elements/themes API rather than creating an unrelated token system. Elements API snapshot
• Implemented: The React package contains Code Connect mappings for AI Label variants and examples, linking named Figma components to actual React components. AI Label Code Connect
• Implemented: Component source, TypeScript props, Storybook stories, MDX and tests give machines exact existence, imports, props, example states and executable behavior.
• Documented and implemented: Carbon MCP supplies tools for components, tokens, icons, documentation questions and UI generation. Its repository is preserved separately. MCP overview, MCP repository
• Documented: Carbon publishes prompts, setup instructions, token-conservation guidance and a Figma Make prompt alongside the MCP overview. MCP documentation folder

What a machine can understand

Component identity, imports, props, many variants, token values/names, examples, AI-state integration, Figma-to-code mappings and human-authored usage guidance are discoverable. Carbon MCP reduces the retrieval burden.

What remains implicit

The system documents placement and use cases well for AI presence, but broader product intent—when an experience should use chat, when an explanation is adequate, how to compose agent workflows, or how risk changes approval—is not encoded in public component APIs. MCP access does not make these judgments executable.

Governance and discrepancies

• Implemented: Public API snapshots, component tests, stories and accessibility status provide multiple validation layers.
• Documented: Carbon warns that AI styling is not decoration and constrains the label’s behavior, a governance rule carried in documentation rather than a type-level prohibition.
• Discrepancy/history: Documentation records the rename from AI Slug to AI Label; React still exports compatibility aliases such as unstable__Slug and preview__Slug. This is evidence of migration support, not current conceptual ambiguity. React aliases
• Discrepancy: The high-level guidance calls AI Chat “preview,” while the discovery corpus also preserved a v1 announcement/repository at an active 2026 commit. Status should be checked per package release, not inferred from one page.

Evidence assessment

For the primary research question, Carbon Builder provides the strongest inspectable operational evidence in the corpus: it joins component discovery, canonical implementation, targeted guidance, explicit selection/composition rules and validation protocols. Coverage is selective, many semantic rules remain operational prose rather than a common schema, and the corpus does not show independent evaluation of novel design decisions. Carbon for AI remains valuable supporting evidence about AI-product patterns, not proof of machine consumability.