ux-research checked

Guidelines for Human-AI Interaction

Research-backed guidelines for AI experiences, including capability disclosure, uncertainty, efficient dismissal, correction, and control.

Open source

Pattern Decisions This Source Supports

Pattern Supported decision Required contract Claim note
Adaptive defaults Choose adaptive defaults when the system proposes a starting value inside the current task based on context or learned behavior. Opening the task shows the proposed value and why it was chosen before the user submits, exports, sends, schedules, or applies anything. Microsoft Research supports contextual relevance, correction, dismissal, and user control for AI-infused behavior.
Agent plan preview Choose agent plan preview when an AI agent's proposed multi-step execution needs review before it starts. The preview names the objective, plan version, source context, model or workflow instructions, planned tools, and expected output before run. Supports user control, correction, expectations, uncertainty, and capability boundaries.
Agent progress trace Choose agent progress trace when an agent or automation run has started and users need live status across multiple steps or tools. The trace belongs to one run ID, reviewed plan version, objective, actor, start time, current state, and final state. Supports user control, feedback, correction, and expectation-setting during AI-assisted experiences.
AI agent acts without approval Flag this anti-pattern when an AI agent or automation executes a high-impact side effect without showing and requiring approval for the exact action and payload first. The agent distinguishes read-only steps, draft steps, reversible local changes, and external side-effect steps before execution. Supports user control, capability disclosure, correction, and human oversight in AI experiences.
AI confidence shown as fake precision Flag this anti-pattern when an AI or automated surface shows a precise score without calibration scope, decision threshold, freshness, uncertainty reason, or safe next action. The interface distinguishes raw internal scores from user-facing calibrated confidence. Supports AI uncertainty communication, capability limits, correction, and user control.
Confidence / uncertainty display Choose confidence / uncertainty display when users need to judge prediction reliability before acting on AI or automation output. The display names the task, prediction, source of the confidence estimate, calibration scope, and last calibration or update time when available. Supports capability boundaries, uncertainty communication, correction, feedback, and control in AI experiences.
Correction feedback Choose correction feedback when users need to correct AI output, source use, assumptions, recommendations, safety behavior, or answer quality after an AI response is shown. A correction feedback action preserves answer ID, response version, claim span, source ID, user reason, expected correction, submitter, timestamp, and chosen scope. Supports user correction, feedback, control, and calibrated trust for AI behavior.
Dangerous-action review Choose dangerous-action review when the user is about to execute a high-impact action and needs to inspect the exact payload, risk, evidence, and side effects before it leaves the safe preview state. The review is bound to a specific action ID, payload version, target, actor, permission scope, source context, evidence set, and policy trigger. Supports human control, correction, uncertainty handling, and capability-boundary communication for AI-assisted high-impact actions.
Editable AI output Choose editable AI output when the primary object is generated output after creation and users need to revise, review, save, copy, or apply that output. The generated draft, user edits, tracked changes, source mappings, citations, review status, and final output state are modeled as distinct data rather than one mutable text blob. Supports user correction, feedback, control, and capability-boundary disclosure in AI experiences.
Escalate to human Choose escalate to human when users need a route from AI, automation, chatbot, self-service, or failed recovery to a human channel or queue. The escalation action names the human destination and whether it is live, asynchronous, specialist, supervisor, emergency, or review-only. Supports user control and appropriate human-centered recovery when AI is uncertain or inappropriate.
Human approval gate Choose human approval gate when automation is paused at runtime and cannot execute the next step until an eligible human authorizes it. The gate belongs to a specific automation run, step ID, payload version, model or workflow version, target object, and approver rule. Supports human control, correction, uncertainty, and capability-boundary guidance.
Prompt suggestions Choose prompt suggestions when the user needs help discovering useful AI requests, not when the task can be captured by a fixed form. Selecting a suggestion populates or inserts editable prompt text instead of silently running the request. Human-AI guidelines support capability clarity, user control, correction, dismissal, and appropriate reliance.
Recommendations Choose recommendations when the system ranks a small set of likely useful items from behavior, context, similarity, popularity, rules, or model output. Opening a recommendation navigates to that item or action and records the event according to the disclosed analytics and personalization rules. Microsoft Human-AI guidelines support capability disclosure, uncertainty, efficient dismissal, correction, and user control.
Regenerate / retry Choose regenerate / retry when the task is rerunning an AI response attempt, creating a new answer version, continuing a failed generation, or recovering from failed AI source or tool work. Retry same prompt repeats the submitted prompt and visible context unless the UI explicitly shows which context changed. Supports user correction, control, feedback, and appropriate reliance in AI experiences.
Scope clarification Choose scope clarification when the AI needs a missing boundary before producing a reliable answer, plan, retrieval query, or tool action. The system identifies what boundary is missing before asking the user to clarify. Supports clarification, control, and avoiding inappropriate AI assumptions.
Tool-use visibility Choose tool-use visibility when users need to inspect exact tool names, purposes, inputs, outputs, permissions, side effects, failures, or redactions. Each tool-use item identifies one tool call or tool-call attempt with stable call ID, run ID, step ID, tool name, status, timestamp, and permission scope. Supports user control, correction, feedback, and capability-boundary disclosure in AI experiences.

Evidence Role

This source is treated as ux-research evidence. Use it to validate the decision rules above, not as a visual style reference.

Publisher: Microsoft Research. Last checked: .