Accessibility analysis
Type: Finding
Status: Active supporting research
Canonical for: Accessibility-specific implications and evidence gaps for agent-consumed or generated design-system output
Depends on: System profiles and preserved accessibility evidence in corpus/
Scope note: Accessibility guidance is central only when the research can show how an agent receives and applies it. Component accessibility or human-facing guidance alone satisfies neither contextual retrieval nor evaluation of agent decisions.
Where conventional DS guidance remains sufficient
Existing accessible primitives still address much of the visible interface:
- Semantic elements and ARIA.
- Keyboard operation and logical focus order.
- High contrast, color contrast and reflow.
- Responsive layouts.
- Non-color communication.
- Accessible dialogs, popovers, inputs, cards and error messages.
- Automated, keyboard and screen-reader testing of known component states.
Carbon demonstrates this directly by building AI Label on an accessible popover trigger and integrating it into tested controls. Microsoft explicitly directs agent teams to Fluent and tokens rather than bespoke UI. Atlassian’s accessible defaults become inherited constraints for generated use.
AI-specific or intensified problems
Streamed and changing output
Carbon implements streaming and careful scroll behavior, including user scroll-away and stream completion. Microsoft requires responsiveness and user control of pace. Public guidance is weak on live-region granularity, announcement batching, whether partial tokens should be announced, and how users return to the generated region. Implementation exists; accessibility rules are incomplete.
Focus under dynamic replacement
Carbon documents focus behavior for the AI Label and revert state. The corpus is much less explicit about focus when generated cards, panels or whole interfaces are replaced. Generative UI makes stable landmarks and predictable focus materially harder.
Long-running tasks and agent status
Microsoft suggests distinct sounds for availability, completion and error, but sound cannot be the sole signal. Tier 1 guidance does not provide a robust shared model for queued/running/waiting/blocked/canceled/partially complete states across assistive technologies.
Motion
Atlassian is developing semantic motion tokens meant to communicate intent to humans and AI. This may help agents select appropriate motion, but public evidence does not show AI-specific reduced-motion rules for generated or agent-driven transitions.
Generated content and provenance
Carbon’s label is keyboard accessible and localizable. Challenges remain when labels repeat at fine granularity, authorship is mixed, or generated content changes after editing. Excessive disclosure can create navigation and cognitive load.
Generated interfaces
SAP says accessibility must become a generation-time guardrail. This is directionally strong, but the corpus lacks a public enforcement schema and evidence that arbitrary compositions preserve landmarks, reading order, focus, zoom/reflow and understandable interaction.
Error recovery
Microsoft’s guidance to preserve workflows and allow revise/regenerate/edit is accessible in principle because it reduces restart burden. Consequential action rollback and partial failure need more specific patterns.
Content accessibility
Microsoft adds prompt simplicity, global comprehensibility, cognitive-load control and multimodal input/output. Carbon adds localization and a structured explanation template. Atlassian adds AI voice/tone. This shows content is not secondary metadata: it shapes whether an AI experience can be understood and corrected.
Evidence gaps
- No shared streamed-output announcement protocol.
- Little public user research on repeated AI labels with screen readers.
- Weak guidance for accessible plans, tool calls and background task timelines.
- No established a11y validation method for novel generated compositions before render.
- Little guidance for uncertainty/confidence communication across cognitive abilities.
- Few public test suites for agent lifecycle patterns.