If you feel like the ground is shifting, it is.
You are in more reviews where someone says “RAG” or “agent” or “guardrails” and expects you to know what that means for the interface. Not so you can train models. So you can design the part users actually touch: trust, limits, recovery, and what happens when the AI is wrong.
That is the job now. Judgment over jargon. But you still need a shared vocabulary, or every spec meeting turns into two parallel conversations.
These ten terms come from the AI UX glossary. They are the ones I keep coming back to when I am reviewing a feature, writing a brief, or arguing about what “done” looks like before anything ships.
Framer - The best code-free tool for designers to create beautiful websites. Use our special promotion code: partner25proyearly to get 3 months free yearly Pro subscription.
What it is: When an AI states something confidently that is false, outdated, or unsupported by its inputs.
This is a normal failure mode, not a rare bug. Trustworthy AI UX assumes hallucinations will happen and designs verification into the flow.
Why it matters for designers: Without citations, edit steps, and low-confidence states, users act on wrong answers in high-stakes workflows.
Example: A research assistant labels unsourced claims as “Unverified,” requires citations for factual bullets, and offers “Compare to source” hover on each sentence tied to retrieved snippets.
Deep dive: The full Hallucination entry includes three common mistakes (no recovery path after a bad claim is the one teams forget), plus links to adjacent terms like RAG, Guardrails, and Human-in-the-Loop. For spec work, jump straight to the related patterns: Confidence Indicators, Citations, and Citation Tooltips, each with interactive demos. If you are designing agents, pair this with the Agentic UX framework.
What it is: Retrieve relevant documents or records first, then ask the model to answer using that material. Search step + generation step.
This powers “chat with your docs,” in-app help, and support copilots that must stay tied to your knowledge base.
Why it matters for designers: RAG UX must show sources, gaps, and confidence when nothing relevant is found. Otherwise users trust fluent wrong answers.
Example: An internal policy bot answers “Can I expense rideshare?” with quoted handbook sections, links to the source PDF page, and “Not found in policy” when retrieval returns nothing.
Deep dive: On the RAG page you get the full search-then-generate breakdown, mistakes like hiding which corpus was searched, and links to Embeddings, Semantic Search, and Context Window. Ship-ready patterns: Citations, Citation Tooltips, and Confidence Indicators. For conversation-heavy products, the Chat UX framework shows how retrieval fits the full chat contract.
What it is: How an interface ties AI answers to verifiable sources: documents, URLs, files, or tool results. Retrieval fetches context; grounding UX shows it.
Why it matters for designers: Without grounding UX, even accurate RAG feels like magic or hallucination. Users need one-click paths from a claim to its source.
Example: Numbered citation chips open the exact paragraph. Clicking a chip scrolls the source list and highlights the span.
Deep dive: Grounding (UX) is where retrieval meets interface design: Perplexity-style examples, export/share pitfalls, and links to RAG, Hallucination, and Explainability. Three patterns to study together: Citations for inline numbering, Citation Tooltips for hover-to-verify, and Source Browser for the side-panel deep read.
What it is: The design of interfaces for software that plans and acts on a user’s behalf. Tools, files, APIs, multi-step workflows. Not only text replies.
The core questions are permission, reversibility, autonomy, cost, handoffs, and trust in what the agent actually did.
Why it matters for designers: A wrong chat answer is embarrassing. A wrong agent action can delete data or send email. Agentic UX makes oversight legible before and after execution.
Example: Before archiving 14 contacts, the agent shows a plain-language plan, lets the user edit scope, and writes each completed action to a readable log with undo.
Deep dive: The Agentic UX glossary entry connects to companion terms you will spec together: Intent Preview, Action Receipt, and Autonomy Slider. Patterns with demos: Scope Disclosure and Human in Loop. For the full territory map (permission, undo, autonomy, cost, handoffs), open the Agentic UX framework: 36 patterns across 7 territories.
What it is: A person reviews, approves, or corrects AI output before it affects users, records, or systems. The safety net when automation is fast but not yet trustworthy enough to ship without oversight.
Why it matters: HITL patterns need clear diffs, one-click approve/reject, audit logs, and SLA-aware queues. Especially for agents that chain multiple steps.
Example: An email agent drafts replies in “Pending review.” Support leads batch-approve five messages with inline edits before anything reaches customers.
Deep dive: Human-in-the-Loop covers checkpoint design, queue metrics, and links to Agent, Guardrails, Tool Use, and Workflow. Two patterns to prototype first: Human in Loop for single-step approval and Approval Workflows for batch review. Also see the Agentic UX framework for where HITL sits in the autonomy gradient.
What it is: Telling users, in plain language, what an AI feature can and cannot do before and during use. Visible statements of scope, limits, data access, and known failure modes.
Why it matters: Undisclosed limits drive support tickets, safety incidents, and churn when users expect ChatGPT-level generality from a narrow feature.
Example: A document Q&A feature states it only searches uploaded PDFs, shows when a question is out of corpus, and links to “What this can answer.”
Deep dive: Capability Disclosure includes copy placement mistakes (disclosure at signup but never at the action) and ties to Probabilistic UI and Hallucination. Patterns to reference in your spec: Scope Disclosure for in-product limits and Confidence Indicators for when the system is operating at the edge of its scope.
What it is: Deterministic systems return the same output for the same input every time. Stochastic systems, including most LLMs, sample from possible answers so results can vary run to run.
This distinction explains why AI buttons feel “alive” but also unreliable compared to traditional software.
Why it matters for designers: You design different UX for each mode. Confirmations and diffs for stochastic drafts. Strict validation for deterministic pipelines. Never promise spreadsheet precision from a sampling model without structure.
Example: “Regenerate summary” produces three variants users can compare. “Export to CSV” runs a schema-locked step with validation errors, not a free-form chat reply.
Deep dive: The Deterministic vs. Stochastic entry walks the full contrast with links to Structured Output, Hallucination, and Inference. Patterns that encode each mode: Streaming for in-progress stochastic output and Confidence Indicators for signaling when variation is expected vs. when output is locked.
What it is: Rules, filters, and policies that block unsafe inputs, limit risky outputs, and keep AI behavior aligned with product and brand standards.
Why it matters for designers: Users experience guardrails as refusals, redactions, tone shifts, and blocked actions. Copy and fallback UX must explain limits without feeling arbitrary.
Example: An image editor refuses NSFW prompts with specific guidance (”Try describing style without people”) and offers safe preset styles instead of a dead-end error.
Deep dive: Guardrails maps the full stack (system prompts, classifiers, validators, rate limits) and links to Moderation, System Prompt, Prompt Injection, and Human-in-the-Loop. Pattern: Human in Loop for review gates guardrails trigger. Prompt: Prompt Engineering Best Practices for writing policy-aware system instructions.
What it is: When AI produces live interface elements (forms, dashboards, cards, charts) from prompts or data, not just static text.
Why it matters for designers: GenUI shifts design work toward review, constraints, and design systems. You need design tokens, component allowlists, accessibility checks, and diff/review before merge. Otherwise you get pretty demos that break your system.
Example: A PM describes a metrics panel; GenUI renders cards bound to real data schema. The designer swaps auto-picked charts for approved components before release.
Deep dive: Generative UI (GenUI) covers preview-vs-production boundaries, version compare, and links to DESIGN.md, SKILL.md, Structured Output, and Workflow. Pattern with demo: Generative UI. For agent constraints on generated screens, browse DESIGN.md systems.
What it is: Adapting a base model to your domain, tone, or task by training on curated examples, beyond what a system prompt alone can reliably enforce.
Why it matters for designers: You feel fine-tuning as more consistent voice, format, and terminology. But also slower update cycles when brand or policy changes. Plan for versioning, evals, and rollback because behavior shifts in subtle ways.
Example: A healthcare portal fine-tunes on approved patient-facing phrases. The UI still shows “Draft: clinician review required” because tuning reduced but did not eliminate risk.
Deep dive: Fine-Tuning includes the prompts-vs-tuning-vs-RAG decision frame, versioning pitfalls, and links to Few-Shot Prompting, System Prompt, Model, and Personalization. Prompt: Prompt Engineering Best Practices for what to exhaust before you tune. Pair with Guardrails and Human-in-the-Loop when tuning reduces but does not eliminate risk.
You do not need to memorize definitions. You need enough shared language to ask better questions in review:
What happens when the model hallucinates?
Where does grounding show up in this flow?
Is this stochastic or deterministic, and did we design for that?
What guardrails will users hit, and do we explain them?
Is this GenUI going through our system, or around it?
That is how alignment gets faster. Not more frames. Clearer terms, then prototypes you can argue with productively.
Browse the full glossary: aiuxplayground.com/glossary
Related guides on AI UX Playground:
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.