Skip to content

sergiovazquezdml/Base-de-datos-aluminio / .agents/skills/agentic-ai-generative-ux

Agentic AI & Generative UX

Verified

Design patterns for agentic AI systems, multi-agent orchestration, generative UI, RAG interfaces, AI safety guardrails, trust calibration, and the paradigm shift from tool-based to agent-based interfaces. Covers AI copilot frameworks, prompt engineering UX, and AI anti-patterns.

references/ai-safety-trust-guardrails.md

View on GitHub
# AI Safety, Trust Calibration, and Guardrail Design Patterns

## Authoritative Reference for Trustworthy AI Interface Design

This reference defines the design patterns, psychological frameworks, and guardrail architectures required to build AI systems that users trust appropriately, recover from gracefully when they fail, and interact with safely across risk domains. It consolidates guidance from NNG Group's State of UX 2026, Microsoft's HAX Toolkit, Google PAIR, OpenAI's design guidelines, and the EU AI Act's UX implications into a unified, actionable framework for practitioners.

---

## AI Trust Calibration

### Trust as the New Benchmark

NNG Group's 2026 research elevates trust to the defining metric for AI product success. Accuracy alone does not determine adoption. Users who do not trust the system will verify every output manually, eroding the efficiency gains AI promises. Users who over-trust the system will propagate errors, hallucinations, and biased outputs into critical decisions. Neither state is acceptable. Calibrated trust, where users trust the AI exactly as much as its actual reliability warrants, is the design goal.

### The Trust Calibration Problem

Users arrive at an AI system with preexisting trust levels shaped by media narratives, prior experiences with other AI products, domain expertise, and personality traits. These initial trust levels almost never match the actual reliability of the specific system they are using. The calibration problem manifests in two failure modes:

- **Over-trust:** The user accepts AI output without verification. They copy AI-generated text directly into reports, act on AI recommendations without checking sources, and delegate high-stakes decisions without oversight. This is dangerous because every AI system has failure modes, hallucination rates, and blind spots.
- **Under-trust:** The user verifies every output manually, re-checks every calculation, and second-guesses every recommendation. This defeats the purpose of the AI system entirely and often leads to abandonment after the novelty wears off.

### The Trust Lifecycle

Design for trust as a dynamic process, not a static state:

1. **Initial Trust Formation:** The user's first interaction establishes a trust baseline. Onboarding, product framing, and the first few outputs disproportionately influence this baseline. Set realistic expectations during onboarding. Show the AI succeeding and failing during guided walkthroughs so users form an accurate mental model from the start.
2. **Trust Through Experience:** Each interaction either reinforces or adjusts the user's trust level. Consistent accuracy builds trust incrementally. A single catastrophic failure can collapse trust entirely. Design for graceful failure before optimizing for peak performance.
3. **Calibrated Trust:** The mature state where the user has an accurate mental model of when the AI is reliable and when it is not. They verify selectively based on domain, confidence indicators, and personal stakes. This is the target state.

### Factors That Influence AI Trust

| Factor | Trust-Building Design | Trust-Eroding Design |
|---|---|---|
| **Accuracy** | Show accuracy metrics by domain; surface known limitations upfront | Hide error rates; present all outputs with equal certainty |
| **Transparency** | Explain reasoning; cite sources; show confidence levels | Black-box outputs with no rationale |
| **Consistency** | Same input produces reliably similar output; behavior is predictable | Outputs vary wildly between identical prompts |
| **Controllability** | User can adjust, override, undo, and constrain AI behavior | AI acts autonomously without user controls |
| **Accountability** | Clear feedback mechanisms; errors acknowledged and addressed | No way to report errors; failures are silent |

---

## Confidence Visualization Patterns

### Color-Coded Confidence

Map AI confidence scores to a three-tier color system with explicit thresholds:

| Confidence Range | Color | Label | Recommended Action |
|---|---|---|---|
| 90-100% | Green (`#22C55E`) | High Confidence | Auto-action permitted in low-risk contexts |
| 60-89% | Amber (`#F59E0B`) | Moderate Confidence | Present as recommendation; user approves |
| Below 60% | Red (`#EF4444`) | Low Confidence | Advisory only; require human judgment |

Never rely on color alone. Pair every color indicator with a text label, icon, or pattern for accessibility. Use a checkmark icon for high confidence, a caution triangle for moderate, and a question mark or warning icon for low confidence.

### Mode Switching by Confidence

Implement dynamic behavior that adjusts the interaction model based on confidence:

```
if confidence >= 0.90 and risk_tier == "low":
    execute_action()          # Auto-execute, log for review
    show_notification("Done: filed expense report for $47.50")
elif confidence >= 0.60:
    present_recommendation()  # Show recommendation, await approval
    show_dialog("I recommend categorizing this as 'Travel'. Approve?")
else:
    present_advisory()        # Show analysis, require human decision
    show_advisory("I'm not sure about this item. Here are 3 possible
                   categories with my reasoning for each.")
```

### Linguistic Qualifiers

Match the language the AI uses to its actual confidence level. Never present uncertain outputs with confident language.

| Confidence Level | Linguistic Pattern | Example |
|---|---|---|
| High (90%+) | Declarative statements | "This transaction is a duplicate of invoice #4521." |
| Moderate (60-89%) | Hedged assertions | "This appears to be a duplicate. Verify against invoice #4521." |
| Low (below 60%) | Explicit uncertainty | "I'm not sure, but this might be related to invoice #4521. Please review." |
| Very Low (below 30%) | Refusal to assert | "I don't have enough information to classify this. Here are some possibilities." |

### Quantitative Confidence Display

For numeric predictions, show uncertainty ranges rather than point estimates:

```
Revenue Forecast: $2.4M
                  ├── Likely range: $2.1M - $2.7M (80% confidence)
                  └── Full range:   $1.8M - $3.1M (95% confidence)
```

Use confidence bars for binary or categorical predictions. Display the bar with a fill proportional to the confidence percentage, a numeric label, and the threshold markers for action tiers.

### Probabilistic Output Display

When the AI has multiple plausible answers, show them ranked with likelihoods rather than collapsing to a single "best" answer:

```
Diagnosis Suggestions:
  1. Seasonal allergies     — 72% confidence
  2. Common cold            — 18% confidence
  3. Sinus infection        —  7% confidence
  4. Other                  —  3% confidence
```

Present the top answer prominently but keep alternatives visible and selectable. This pattern is critical in medical, legal, and financial domains where the second-most-likely answer may be the correct one.

---

## Trust Psychology Framework

### Three States of Trust

Design interventions must address all three trust states:

**Calibrated Trust** is the target. The user checks AI output selectively, based on the domain and stakes. They trust the AI for tasks where it has proven reliable and verify outputs in domains where it has not. Design interventions for this state focus on maintaining calibration: continue showing confidence indicators, surface accuracy metrics, and gently flag when the user is operating in a domain where the AI has lower reliability.

**Over-Trust** is the most dangerous state. Indicators include: the user never edits AI outputs, never checks citations, auto-accepts all recommendations, and has disabled confirmation dialogs. Design interventions:

- Introduce periodic "verification nudges" that ask the user to spot-check a random output.
- Show accuracy statistics by domain so users see that reliability varies.
- Present occasional "known wrong" test outputs during onboarding to demonstrate fallibility.
- Make the cost of unverified errors visible through case studies or contextual warnings.

**Under-Trust** is the most wasteful state. Indicators include: the user re-does every task the AI completed, ignores AI suggestions, and takes longer with the AI than without it. Design interventions:

- Show the AI's track record for the specific task type the user is verifying.
- Offer A/B comparisons between AI output and the user's manual work.
- Progressively increase autonomy as accuracy is demonstrated over time.
- Celebrate verified accuracy milestones to reinforce confidence.

### Trust Repair After Failure

When the AI fails, the trust repair process determines whether the user recovers to calibrated trust or collapses into permanent under-trust:

1. **Acknowledge the error immediately.** Never hide or minimize failures. "I made an error in the calculation. Here's what went wrong."
2. **Explain why it happened** in terms the user can understand. "The source data contained a formatting inconsistency that caused an incorrect total."
3. **Show the corrected output.** Do not just apologize; provide the fix.
4. **Demonstrate what changed** to prevent recurrence. "I've flagged this data pattern for manual review in future runs."
5. **Offer increased oversight temporarily.** "Would you like me to show my work for the next few calculations?"

### Trust Variation by Domain

Trust requirements differ dramatically across domains. Apply stricter guardrails and more conservative confidence thresholds in high-stakes domains:

| Domain | Trust Sensitivity | Recommended Approach |
|---|---|---|
| Medical / Health | Extreme | Always advisory; never auto-action; mandatory source citations |
| Financial / Legal | Very High | Advisory with mandatory review for transactions above threshold |
| Creative / Writing | Moderate | Higher autonomy acceptable; focus on preference alignment |
| Data Analysis | High | Show methodology and intermediate steps; enable spot-checking |
| Customer Service | Moderate-High | Escalation paths to humans; never impersonate human agents |

---

## Uncertainty Visualization for LLM Outputs

### Highlighting Low-Confidence Passages

Apply inline visual markers to LLM-generated text that distinguish high-confidence content from uncertain content. Use a subtle background highlight (light amber or a dashed underline) on passages where the model's token-level confidence falls below a threshold. Provide a hover tooltip explaining: "This passage has lower confidence. Consider verifying independently."

### Showing Alternative Completions

For critical outputs, offer a "Show alternatives" expansion that reveals two or three alternative completions for the same prompt. Display them in a tabbed or stacked layout with a brief label indicating how each differs. This empowers the user to select the most appropriate version rather than accepting a single output blindly.

### Confidence Ranges for Factual Claims

When the AI makes factual assertions, annotate them with inline confidence and source availability:

```
The company was founded in 1987 [High confidence, 3 sources].
Revenue exceeded $4.2 billion in Q3 [Moderate confidence, 1 source].
The CEO announced a restructuring plan [Low confidence, unverified].
```

Use a consistent annotation syntax throughout the product so users learn to read confidence signals at a glance.

### Visual Differentiation Between Certain and Uncertain Content

Establish a clear visual language that separates AI-confirmed content from AI-uncertain content:

- **Certain content:** Standard text rendering with no additional markers.
- **Uncertain content:** Lighter text weight or opacity reduction (0.7 opacity), dashed left border, or a subtle background tint.
- **Unverifiable content:** Italic rendering with an explicit label: "[Unverified]".

Never mix these visual treatments within the same paragraph without clear demarcation.

---

## XAI Design Integration

### Feature Attribution Visualization

Show users what influenced the AI's decision. For classification tasks, display a ranked list of the top contributing factors with their relative weights. Use horizontal bar charts or a ranked list with percentage contributions:

```
Decision: Loan Application APPROVED

Top factors:
  Credit score (780)           ████████████████░░  42%
  Income-to-debt ratio (3.2)   ██████████░░░░░░░░  27%
  Employment tenure (8 yrs)    ██████░░░░░░░░░░░░  16%
  Savings balance              ████░░░░░░░░░░░░░░  10%
  Other factors                ██░░░░░░░░░░░░░░░░   5%
```

### Counterfactual Explanations

Provide "what if" explanations that show how changing an input would change the output. These are among the most intuitive explanation types for non-technical users:

- "If your credit score were below 650, this application would likely be denied."
- "If the document had included a signature on page 3, confidence would increase to 95%."
- "Removing the outlier data point in row 47 changes the trend from positive to neutral."

### Example-Based Explanations

Ground explanations in concrete examples drawn from the training distribution or prior cases:

- "This email is classified as phishing because it closely resembles 847 confirmed phishing emails in our database, particularly those impersonating financial institutions."
- "This image is flagged as potentially AI-generated because it shares visual artifacts with known synthetic images, specifically in the hand region."

### Progressive Explanation Depth

Implement a three-tier explanation system that serves users at every expertise level:

1. **Summary (default):** One sentence explaining the decision in plain language. "Approved because of strong credit history."
2. **Detail (expandable):** Factor-by-factor breakdown with visualizations. Accessible via a "Why?" or "Show reasoning" button.
3. **Technical (advanced):** Model internals, feature importance scores, confidence calibration curves, and raw probability distributions. Accessible via a "Technical details" link or developer mode.

### When NOT to Explain

Resist the impulse to explain everything. Explanations add cognitive load and can create new problems:

- **Explanation overload:** Providing detailed explanations for every trivial decision trains users to ignore explanations entirely, including on high-stakes decisions where they matter.
- **Explanation gaming:** Detailed feature attributions can teach adversarial users how to manipulate inputs to achieve desired outputs (e.g., gaming a loan approval system).
- **False confidence from explanation:** A plausible-sounding explanation does not make the output correct. Users may trust explained wrong answers more than unexplained right answers.

Show explanations by default only for decisions above a defined risk threshold. For low-stakes decisions, make explanations available on demand but do not surface them proactively.

---

## UX Guardrail Framework

### Consolidated from Microsoft HAX, Google PAIR, and OpenAI Guidelines

Organize guardrails into three temporal phases:

### Pre-Action Guardrails

Apply before the AI executes any action:

- **Input validation:** Check user inputs for ambiguity, missing context, and potential misinterpretation. "You asked me to 'delete the old files.' Which folder? Files older than when?"
- **Intent confirmation:** For ambiguous or high-stakes requests, restate the interpreted intent and confirm. "I'll send this email to all 2,400 subscribers. Confirm?"
- **Constraint checking:** Verify the requested action falls within the user's configured guardrail boundaries (spending limits, scope restrictions, permission boundaries).

### In-Action Guardrails

Apply during execution:

- **Progress monitoring:** Show real-time execution progress with the ability to pause or cancel.
- **Anomaly detection:** If the agent encounters unexpected conditions during execution (unusually large data sets, unexpected error rates, contradictory information), pause and escalate.
- **Human-in-the-loop triggers:** Define specific conditions that halt autonomous execution and require human input before continuing.

### Post-Action Guardrails

Apply after execution:

- **Output review:** Present results for user review before they become final. Stage outputs rather than auto-committing.
- **Undo and rollback:** Provide single-click undo for every AI action. For irreversible actions, implement a "cooling off" delay (30 seconds to 5 minutes depending on severity) during which the action can be cancelled.
- **Feedback collection:** Offer lightweight feedback mechanisms (thumbs up/down, inline correction, "flag this") on every output.

### Guardrail Severity Tiers

| Tier | Severity | Guardrail Pattern | Example |
|---|---|---|---|
| 1 | Low | Auto-execute with log entry | Sorting a list, reformatting text, generating a summary |
| 2 | Medium | Recommend and await approval | Sending an email, modifying a database record, making a purchase under $50 |
| 3 | High | Mandatory multi-step confirmation | Deleting data, sending to a large audience, financial transactions over $50 |
| 4 | Critical | Requires secondary authentication | Changing security settings, accessing sensitive records, irreversible bulk operations |

---

## Risk-Tiered Approval Patterns

### Informational Tier (Low Risk, Reversible)

The agent acts autonomously and records the action in a reviewable log. Use for routine, reversible, low-impact operations. The user does not need to approve each action but can review a summary at any time.

**UI Pattern:** A collapsible activity feed or notification badge showing recent agent actions. Each entry includes timestamp, action description, and a one-click undo button. Example: "Organized 47 emails into folders. [Undo] [View details]"

### Advisory Tier (Medium Risk)

The agent recommends an action, explains its reasoning, and waits for user approval. Use for actions that have moderate impact, affect other people, or involve spending.

**UI Pattern:** A modal or inline card showing the proposed action, the reasoning, and Approve / Modify / Reject buttons. Include a "Always allow this type" checkbox for users who want to promote recurring actions to the informational tier. Example: "Recommend replying to John's email with the attached draft. [Approve] [Edit] [Reject]"

### Mandatory Tier (High Risk, Irreversible)

The agent cannot proceed without explicit, deliberate user approval. Use for irreversible actions, significant financial impact, actions affecting many people, or security-sensitive operations.

**UI Pattern:** A full-screen or prominent confirmation dialog requiring the user to type a confirmation phrase or provide biometric confirmation. Display a clear summary of consequences. Include a mandatory delay (at least 5 seconds) before the confirm button becomes active to prevent reflexive clicking. Example: "This will permanently delete 2,400 customer records. Type 'DELETE' to confirm. This action cannot be undone."

### Classifying Actions Into Tiers

Evaluate every agent capability across two dimensions:

| | Reversible | Irreversible |
|---|---|---|
| **Low Impact** | Tier 1 (Informational) | Tier 2 (Advisory) |
| **High Impact** | Tier 2 (Advisory) | Tier 3 (Mandatory) |

Add one tier for actions that affect other people, involve financial transactions, or touch sensitive data.

---

## Guardrail Configuration Interfaces

### Autonomy Level Controls

Provide a global autonomy slider with three named presets and a custom option:

- **Conservative:** Agent recommends everything; never auto-executes. Suitable for new users and high-stakes domains.
- **Balanced:** Agent auto-executes low-risk actions, recommends medium-risk actions, requires approval for high-risk actions. The default for most users.
- **Aggressive:** Agent auto-executes low and medium-risk actions, recommends high-risk actions. For experienced, trusting users in low-stakes domains.

### Per-Domain Permission Settings

Allow users to set different autonomy levels for different action domains:

```
Email:        [Conservative ▾]
Calendar:     [Balanced ▾]
File Management: [Balanced ▾]
Purchases:    [Conservative ▾]
Code Changes: [Aggressive ▾]
```

### Spending and Action Limits

Expose configurable limits with clear defaults:

- Maximum single transaction amount: $50 (default)
- Maximum daily aggregate spending: $200 (default)
- Maximum number of emails sent per hour: 10 (default)
- Maximum file deletions per action: 5 (default)

### Audit Log Access

Provide a searchable, filterable audit log showing every action the agent took, including:

- Timestamp and action description
- Which guardrail tier applied
- Whether the user approved, and when
- Input data and output result
- Confidence score at time of action

---

## AI UX Anti-Patterns

### AI Fatigue

NNG Group's State of UX 2026 documents widespread AI fatigue among users. The symptoms: users dismiss AI suggestions without reading them, disable AI features proactively, and express frustration at AI-powered products that previously worked well as deterministic tools. Combat AI fatigue by making AI features optional, disableable, and subordinate to the user's primary workflow rather than interrupting it.

### AI Slop

The term describes low-quality AI-generated content that floods products and platforms. AI slop degrades the user experience through generic text, hallucinated details, repetitive phrasing, and content that is technically grammatical but substantively empty. Prevent AI slop by implementing quality gates: minimum confidence thresholds for publishing, human review for user-facing content, and quality scoring that rejects outputs below a defined standard.

### Unnecessary AI

Adding AI to features that function perfectly well without it. A settings toggle does not need "AI-powered personalization." A sort function does not need a "smart sort" mode. Before adding AI to any feature, ask: does this feature have ambiguity or complexity that deterministic logic cannot resolve? If the answer is no, do not add AI. The NNG 2026 report specifically calls out products that replaced reliable deterministic features with less reliable AI alternatives.

### AI Washing

Labeling non-AI features as "AI-powered" for marketing purposes. This erodes trust in genuine AI features across the entire product. Reserve AI labeling for features that actually use machine learning models. Use specific, honest descriptions: "Generated by AI" not "AI-enhanced."

### Forced AI

Requiring users to interact with AI when they prefer direct manipulation. Never gate a core workflow behind an AI interaction. Always provide a non-AI path to accomplish the same task. If a user wants to manually set a reminder for 3 PM, do not force them through a conversational AI flow to do so.

---

## AI Error Recovery and Hallucination UX

### Hallucination Context

Current LLM hallucination rates vary by model and domain but remain non-trivial. Industry benchmarks as of 2025-2026 place hallucination rates between 3-15% for leading models on factual tasks, with higher rates for niche or specialized domains. Design every AI interface under the assumption that some outputs will be incorrect.

### Visual Indicators for AI-Generated Content

Label all AI-generated content explicitly. Use a consistent indicator (an icon, a subtle badge, or a header line) that communicates: "This content was generated by AI and may contain errors." This is not optional. It is increasingly a legal requirement under the EU AI Act and a trust-building practice universally.

### Feedback Loops

Implement three levels of user feedback on every AI output:

1. **Binary feedback:** Thumbs up / thumbs down on the overall output.
2. **Inline correction:** Allow users to select and correct specific portions of AI-generated text, with the correction submitted as training signal.
3. **Flag inaccuracy:** A dedicated "Report error" or "Flag inaccuracy" action that captures the specific claim, the user's correction, and optional context.

### Graceful Degradation

When the AI fails entirely (API timeout, model error, safety filter triggered), degrade gracefully:

- Show a clear, non-technical error message: "AI features are temporarily unavailable."
- Preserve all user work and context so nothing is lost.
- Provide a manual fallback path to complete the task without AI.
- Offer a retry mechanism with a reasonable delay.
- Never display raw error codes, stack traces, or technical jargon to end users.

### Recovery Patterns That Maintain Trust

After an error, the system's next output disproportionately affects trust. Apply extra validation and lower confidence thresholds on the first output after a failure. Surface the recovery explicitly: "I've re-checked this result using an alternative approach. Here's the corrected output with sources."

---

## Human-in-the-Loop Workflow Patterns

### Escalation Triggers

Define explicit conditions that escalate from AI to human:

- Confidence drops below the domain-specific threshold.
- The user explicitly requests human assistance.
- The task involves sensitive categories (medical advice, legal guidance, financial planning above a threshold).
- The AI detects conflicting information in its sources.
- The action would affect a number of people above a defined limit.
- Three or more consecutive user corrections indicate the AI is performing poorly on this task.

### Parallel Review

Route AI outputs through human review before they reach the end user. Display the review status clearly: "AI-generated draft, pending human review" transitions to "Reviewed and approved by [Name]." This pattern is essential for customer-facing content, medical summaries, and legal documents.

### Review Queues

For high-volume AI decisions, implement a batched review queue where human reviewers audit a sample or review flagged items. Design the queue interface to show: the AI's decision, its confidence, the key inputs, and approve/reject/modify actions. Optimize for reviewer speed by pre-selecting the AI's recommended action and requiring minimal clicks for approval.

### Handoff UX Between AI and Human Agents

When transitioning from AI to human support, preserve full context. Transfer the conversation history, the AI's assessment, and any actions already taken. Never force the user to repeat information. Display the handoff explicitly: "I'm connecting you with a human specialist who has the full context of our conversation."

### Collaborative AI

Design for simultaneous human and AI contribution to the same artifact. Use cursor presence indicators (similar to collaborative editing in Google Docs) to show where the AI is working. Allow the user to accept, reject, or modify AI contributions in real-time. Provide clear undo for AI contributions that does not affect the user's own edits.

---

## Ethical AI UX

### Bias Disclosure

Communicate known AI limitations honestly and proactively:

- During onboarding, disclose the training data's scope and known gaps.
- When the AI operates in a domain where bias is documented (hiring, lending, medical diagnosis across demographics), show a contextual notice.
- Provide links to model cards or bias assessments for users who want detail.

### Content Filtering UX

Implement content safety filters with user-appropriate controls:

- Show a clear message when content is filtered: "This response was blocked by safety filters." Never silently omit content.
- Provide an appeal or feedback mechanism: "If you believe this was filtered incorrectly, let us know."
- For professional contexts, offer configurable sensitivity levels with clear descriptions of what each level permits and blocks.

### Data Privacy in AI Context

Make the AI's data usage transparent and controllable:

- Show which user data the AI accesses for each feature.
- Provide per-feature data sharing toggles.
- Disclose whether conversations are used for model training, and provide an opt-out.
- Implement data deletion that actually removes user data from AI training pipelines, not just from the user-facing interface.

### Opt-In Design for AI Features

Never force AI on users. Apply progressive opt-in:

1. **Awareness:** Show that AI features exist without activating them.
2. **Trial:** Let users try AI features with a single interaction before committing.
3. **Adoption:** Allow users to enable AI features permanently with clear controls to disable.
4. **Customization:** Let users configure AI behavior, autonomy, and scope.

### EU AI Act UX Implications

The EU AI Act creates concrete UX requirements for AI systems deployed in the EU:

- **Transparency obligation:** Users must be informed when they are interacting with AI. Display a persistent, unambiguous indicator.
- **High-risk system requirements:** AI systems in hiring, credit, education, and healthcare must provide explanations of decisions, human oversight mechanisms, and accuracy documentation.
- **Right to explanation:** Users affected by AI decisions have the right to understand the basis. Build explanation interfaces as a core feature, not an afterthought.

---

## AI Feature Accessibility

### Screen Reader Compatibility with Streaming Output

Streaming AI responses create a challenge for screen readers. Implement the following:

- Use an ARIA live region (`aria-live="polite"`) for the response container so screen readers announce new content without interrupting the user's current focus.
- Buffer streaming tokens and announce complete sentences rather than individual words or tokens to avoid overwhelming the screen reader.
- Provide a "Response complete" announcement when streaming finishes.
- Offer a "Read full response" button that triggers a clean, complete read of the final output.

```html
<div role="region" aria-live="polite" aria-label="AI response">
  <p>The quarterly revenue increased by 12% compared to...</p>
</div>
<div aria-live="assertive" class="sr-only" id="ai-status">
  <!-- Announce: "AI is generating a response" / "Response complete" -->
</div>
```

### Keyboard Navigation for Copilot Patterns

Ensure all inline AI suggestion patterns are fully keyboard-accessible:

- **Tab** to accept an inline suggestion. **Escape** to dismiss.
- **Alt+]** / **Alt+[** to cycle through alternative suggestions.
- Provide a keyboard shortcut to invoke the AI copilot (e.g., **Ctrl+Space**) and another to dismiss it entirely.
- Never trap keyboard focus inside an AI suggestion popover. The user must always be able to continue typing to dismiss the suggestion implicitly.

### Managing Focus with Dynamic Content

When AI generates new UI elements (generative UI, expandable explanations, confirmation dialogs):

- Move focus to the new content only if the user initiated the action that triggered it.
- For background-generated content (agent completing a task), do not steal focus. Use a non-intrusive notification and let the user navigate to the result when ready.
- Implement focus restoration: when a user dismisses a dynamically generated element, return focus to the element that triggered it.

### Alternative Text for AI-Generated Images

Every AI-generated image must include descriptive alternative text. Since the image is generated programmatically, generate the alt text from the same prompt or model metadata:

- Include the subject matter, style, and key visual elements.
- Do not use "AI-generated image of..." as the alt text prefix. Describe the content, then note AI generation separately via a visible label or `aria-describedby`.
- For decorative AI-generated images, use `alt=""` with a visible "AI-generated" label elsewhere.

### Accessible Confidence Indicators

Never communicate confidence through color alone. Use a multi-channel approach:

| Confidence | Color | Icon | Text | ARIA Label |
|---|---|---|---|---|
| High | Green | Checkmark | "High confidence" | `aria-label="High confidence: 94%"` |
| Moderate | Amber | Caution triangle | "Moderate confidence" | `aria-label="Moderate confidence: 72%"` |
| Low | Red | Question mark | "Low confidence" | `aria-label="Low confidence: 38%"` |

Ensure confidence meters and progress bars have programmatic values via `aria-valuenow`, `aria-valuemin`, and `aria-valuemax`.

### Voice Interface Accessibility

For AI features accessed via voice:

- Provide verbal confidence indicators: "I'm fairly confident that..." maps to the moderate tier.
- Allow voice commands to request explanations: "Why did you recommend that?"
- Support interruption during long AI responses without losing context.
- Offer speech rate controls for AI voice output.
- Provide a transcript of all voice interactions that is accessible via screen reader.

---

## Agentic AI Control Patterns (2026)

### The Control / Consent / Accountability Triad

Smashing Magazine's February 2026 framework for agentic AI design identifies three non-negotiable pillars that every agent-driven interface must implement. These pillars address the fundamental trust challenge of autonomous systems: users must remain in command even as agents gain capability.

**Control: The User Can Always Intervene.** The user must be able to pause, redirect, or cancel an agent at any point during execution. This is not a graceful shutdown — it is an immediate halt. Control means:

- A persistent, always-visible "Stop" or "Pause" control during any agent operation. This control must not be obscured by loading states, progress overlays, or modal dialogs.
- The ability to redirect an agent mid-task: "Stop doing X and do Y instead" without requiring a full restart.
- No agent action should proceed to a point of no return without passing through an explicit consent gate first.
- Control extends to scope: users must be able to constrain what the agent can access, modify, and communicate with. An agent granted email access should not be able to access the filesystem without separate, explicit permission.

**Consent: Explicit Approval Before Consequential Actions.** Consent means the user explicitly approves every action that has real-world consequences before it executes. This goes beyond simple confirmation dialogs:

- Approval gates must clearly describe what will happen, what data will be affected, and whether the action is reversible.
- Batch consent is acceptable for low-risk, homogeneous actions: "The agent wants to rename 47 files following the pattern you specified. [Approve All] [Review Each]."
- Consent must be informed — the agent must disclose its confidence level, the data it used to make the recommendation, and any alternatives it considered.
- Consent cannot be irrevocably delegated. A user who previously said "always approve this type of action" must be able to revoke that standing consent at any time.

**Accountability: Full Audit Trail.** Every agent action must be logged, timestamped, and attributable. Accountability means:

- A complete, searchable audit trail that records every action the agent took, the input data it used, the decision it made, and the outcome.
- Each log entry identifies whether the action was auto-executed (Tier 1), approved by the user (Tier 2/3), or escalated to a human reviewer.
- Audit logs must be exportable in standard formats for compliance and review.
- When something goes wrong, the audit trail must make it possible to trace the exact chain of decisions that led to the failure, including what the agent "saw" at each decision point.

### Multi-Step Agent Oversight UX

Agents that execute multi-step plans require specialized oversight patterns that go beyond single-action approval:

**Progress Reporting.** Display a live, structured plan showing the agent's intended steps, current progress, and upcoming actions. Each step shows its status (pending, in progress, completed, failed) and allows the user to inspect details, skip steps, or insert additional steps.

**Rollback on Failure.** When a step in a multi-step plan fails, the agent must be able to undo all previously completed steps in the current plan and restore the system to its pre-plan state. Display the rollback progress clearly: "Rolling back step 3 of 5... Step 2 of 5... Restored to original state." If full rollback is not possible, disclose exactly what was changed and what could not be reverted.

**Approval Batching.** For plans with many low-risk steps and a few high-risk steps, batch the low-risk steps for blanket approval while gating high-risk steps individually. Present this as a plan review screen: "Steps 1-4 are low risk and can be auto-approved. Step 5 (sending email to 200 recipients) requires your explicit approval."

> For implementation patterns including dialogue flows, confirmation UI components, and agent state management, see `conversational-ai-dialogue-patterns.md`.

---

## Cross-References

- For foundational AI interface patterns (chat, copilot, streaming), reference `ai-spatial-voice-ux/references/ai-native-interface-patterns.md`.
- For dark pattern avoidance and ethical design foundations, reference `ux-ethics-content-strategy/references/dark-pattern-avoidance.md`.
- For WCAG compliance and ARIA pattern details, reference `accessibility-inclusive-design/references/wcag-aria-patterns.md`.
- For design token systems that support confidence color coding, reference `ui-visual-design-system/references/design-token-architecture.md`.
- For agentic AI orchestration and generative UI context, reference the parent skill `agentic-ai-generative-ux/SKILL.md`.

## Key Sources

- NNG Group: State of UX 2026 (Trust Calibration, AI Fatigue, AI Slop)
- NNG Group: AI Trust Lifecycle Research
- Microsoft HAX Toolkit: Guidelines for Human-AI Interaction
- Google PAIR: People + AI Guidebook
- OpenAI: Best Practices for Safety and Trust in AI Applications
- EU AI Act: Regulation (EU) 2024/1689 (Transparency and Explainability Requirements)
- ACM CHI 2025: Trust Calibration in Human-AI Interaction
- Gartner: Predicts 2026 — AI Trust, Risk, and Security Management
- NIST AI Risk Management Framework (AI RMF 1.0)
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 323kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall