AI limitation onboarding

Problem Context

• The feature uses AI, machine learning, recommendations, prediction, natural-language generation, tool use, or agent behavior that users may misunderstand.
• Users need to know what the AI can do, what it cannot do, what data it has and lacks, what it may get wrong, and what responsibility remains with them.
• The product is anthropomorphic, conversational, autonomous, personalized, safety-sensitive, or likely to invite Ask me anything expectations.
• The AI may fail because of missing data, low confidence, unsupported input, unstable context, permission limits, outdated sources, policy filters, or unavailable tools.
• The onboarding must be short enough to preserve momentum while still giving users a mental model they can use during the real task.

Selection Rules

• Choose AI limitation onboarding when the user needs first-use or changed-capability orientation about AI benefits, limits, data scope, uncertainty, feedback, human responsibility, and fallback.
• Use regular onboarding when the experience teaches product setup or first value without needing AI-specific capability and limitation calibration.
• Use inline message when a specific screen or field needs contextual AI guidance after onboarding, not the initial mental model.
• Use warning text when the user is about to take one risky action and needs consequence copy near that action.
• Use confidence / uncertainty display when one AI output needs reliability context; use AI limitation onboarding when users need broad expectations before using the feature.
• Use source grounding display when one answer needs evidence coverage; use AI limitation onboarding when users need to understand source scope and missing data before asking.
• Use scope clarification when the AI must ask a question before acting on an ambiguous request; use AI limitation onboarding to teach why the AI may ask and what it cannot infer.
• Use human approval gate when execution must pause for approval; use AI limitation onboarding to explain ahead of time which future actions require review.
• Do not hide capability limits in terms of service, release notes, or help docs when users must understand them before using the AI.
• Do not present limitations as blame-shifting disclaimers; pair each limit with safe next steps, examples, feedback, or fallback.

Required States

• First-run welcome state with user benefit, AI role, and plain-language capability boundary.
• Can do and cannot do state with examples matched to the user's task.
• Data scope state showing included sources, excluded sources, permission limits, and stale data risk.
• Uncertainty state explaining low confidence, missing data, unsupported input, and when output may be wrong.
• Human review state naming high-impact outputs, approval needs, and user responsibility.
• Safe trial state with reversible sample task, sandbox data, or preview-only action.
• Feedback and correction state showing how users can correct, opt out, or improve future behavior.
• Manual fallback state that completes the task without AI when the model cannot help.
• Re-onboarding state for changed AI capabilities, new data use, new autonomy, or degraded capability.
• Skipped, later, remembered, expired reminder, mobile compact, and accessibility states.

Interaction Contract

• The onboarding explains what the AI can help with and what it is not able to do before the first risky use.
• The AI role is not framed as a human, expert, teammate, or authority unless the limits and user responsibility are equally visible.
• Users can skip or defer noncritical teaching, but required safety boundaries remain discoverable before high-impact use.
• Safe trial actions do not commit production changes, send messages, expose private data, or teach the model without consent.
• Data-use disclosures distinguish selected user input, connected sources, historical activity, personalization signals, and unavailable data.
• Users can find correction, feedback, human review, and manual fallback without leaving the task.
• Re-onboarding appears when capability, data access, autonomy, policy, or failure behavior changes enough to affect user expectations.
• The experience supports keyboard navigation, readable status text, non-color-only signals, and focus return after details or trial actions.

Implementation Checklist

• Inventory AI capability, unsupported tasks, data sources, missing data, permissions, policy filters, confidence behavior, review needs, feedback use, personalization, retention, and manual alternatives.
• Write benefit-led copy that says what users can accomplish, then name concrete limits and recovery paths in the same flow.
• Provide a short safe trial using sample data or preview-only output so users can learn behavior without production risk.
• Separate AI limitation onboarding from privacy consent, notification banners, release notes, and legal disclaimers.
• Add re-onboarding triggers for capability expansion, data-source change, model behavior change, automation increase, policy restriction, or repeated user failure.
• Track viewed, skipped, reminder scheduled, trial completed, feedback sent, fallback used, and re-onboarding acknowledged events without treating view count as success.
• Test with new and returning users for over-trust, under-trust, misunderstanding of source scope, failure recovery, and whether users know when to apply judgment.

Common Generated-UI Mistakes

• Using charming human-like copy that implies broad expertise while hiding actual AI limits.
• Showing a one-time disclaimer that users dismiss before they understand what changes in the task.
• Listing broad safety policy text without task-specific examples, safe trial, feedback, or fallback.
• Treating confidence scores, citations, or model names as a replacement for explaining capability boundaries.
• Forcing a long tour that delays first value and teaches features users do not need yet.
• Failing to re-onboard after the AI gains autonomy, loses a data source, or changes behavior.

Critique Questions

• Can a first-time user say what this AI can and cannot do for their task?
• Does the onboarding explain what data the AI has, what data it lacks, and how that affects answers?
• Can users try the AI safely before relying on it?
• Do users know when to review, correct, approve, or switch to manual work?
• Is limitation copy paired with recovery instead of only legal caution?
• What re-onboarding appears after capability, data, or autonomy changes?