Designing AI-Powered Interfaces

Chat assistants, copilots, generative tools, and agentic products share one defining property: probabilistic output. The system will sometimes be wrong, and unlike a broken button, it fails fluently and confidently. Most of AI UX is managing that fact — expectations, status, verification, and keeping the human in control.

Foundational guidelines (start here)

• Microsoft's 18 Guidelines for Human-AI Interaction (Amershi et al., CHI 2019) — the most-cited canon, validated against 20 AI products. Organized by phase: initially (make clear what the system can do, and how well), during interaction (time services on context, show contextually relevant info), when wrong (efficient dismissal, correction, graceful degradation), over time (learn from behavior, update cautiously, notify about changes). Operationalized as the HAX Toolkit (microsoft.com/en-us/haxtoolkit) with per-guideline patterns.
• Google PAIR People + AI Guidebook — six chapters: User Needs + Defining Success, Data Collection + Evaluation, Mental Models, Explainability + Trust, Feedback + Control, Errors + Graceful Failure; updated for generative AI (pair.withgoogle.com). Core stance: solve a real user need — don't add AI because you can.
• Apple HIG — Machine Learning plus a dedicated Generative AI page (developer.apple.com/design/human-interface-guidelines): keep people in control, make AI involvement identifiable, hold generated content to the same quality bar, surface features in context rather than as a separate "AI mode."
• NN/g runs an ongoing research stream on generative-AI chat UX (nngroup.com/topic/ai) — several findings below come from it. It's qualitative usability research, not controlled experiments.

Core principles

• Set capability expectations up front (HAX G1–G2; PAIR "Mental Models"). State what the system is good at, what it can't do, and how reliable it is — in onboarding, empty states, and inline ("AI can make mistakes" is now a near-universal convention). Overpromising is the root cause of most downstream trust collapse.
• Design for calibrated trust, not maximum trust. The automation literature long predates LLMs: people both misuse (over-rely on) and disuse (under-rely on) automation (Parasuraman & Riley 1997); good design targets appropriate reliance — trust matched to actual reliability (Lee & See 2004). Overreliance is the bigger product risk: users accept fluent wrong answers. Cognitive forcing functions (a pause, or asking users to answer first) measurably reduce overreliance, at a cost in speed and satisfaction (Buçinca, Malaya & Gajos, CSCW 2021).
• Explainability at the right grain. Users rarely want model internals; they want actionable explanation: what inputs mattered, why this suggestion here, what to change for a different result (PAIR "Explainability + Trust"). In chat this mostly means sources, the steps/tools an agent used, and letting users ask "why?"
• Make AI status visible. Thinking/generating/retrieving/waiting are distinct states — show which is happening (extends heuristic #1). For agents, "working" without detail reads as hung after a few seconds.
• Graceful uncertainty beats confident wrongness. Hedging where the model is genuinely unsure ("I couldn't verify this"), declining unanswerable questions, and offering verification paths outperform bluffing for long-run trust. Caution: LLM self-reported confidence is often miscalibrated (a consistent finding in LLM evaluation research) — prefer qualitative framing and sources over "87% confident" theater. Related failure: sycophancy — agreeing with users instead of being accurate — is a documented pattern users don't notice (NN/g).

Interaction patterns

• Choose the surface deliberately: chat vs. embedded vs. proactive. Open chat suits exploratory, multi-turn, hard-to-formalize tasks. Embedded assistance (inline rewrite, autocomplete, "summarize" buttons) suits known intents inside an existing workflow — chat is not the default answer (NN/g "AI Chat Is Not (Always) the Answer"). Proactive suggestions carry the highest interruption cost: time them on context (HAX G3–G5), make dismissal one action, never steal focus.
• Solve the blank-canvas problem. An empty prompt box transfers all articulation work to the user. Provide prompt affordances — suggestion chips, example prompts tied to real capabilities, templates, post-response follow-up controls (tone/length/format) — which expedite input and teach capability boundaries (NN/g "Prompt Controls in GenAI Chatbots"). Rotate by context; every chip must actually work well.
• Streaming and latency. Token streaming is the standard convention: multi-second generation feels responsive and users can abandon early (a Stop button is mandatory). Cost: streamed text reflows and users lose their place (NN/g) — stream coherent blocks, stabilize layout, auto-scroll only while the user is at the bottom. For non-chat generation show staged progress; the labor-illusion effect applies.
• Citations and grounding display. Link sources inline or per-paragraph, not a link pile at the end; make excerpts inspectable so verification is cheap. Grounding (RAG) plus visible citations is the strongest UI-level hallucination mitigation — but citation presence isn't citation accuracy; models can cite sources that don't support the claim. NN/g finds chat interfaces actively discourage error checking; design against that default.
• Human-in-the-loop for consequential actions. Anything that spends money, deletes data, messages people, or commits externally gets explicit review: preview/diff of what will happen, approval, friction scaled to stakes. Draft-first ("here's the email — send?") is the workhorse pattern.
• Undo and versioning for generative edits. Generation that overwrites user content must be non-destructive: keep the prior version, offer regenerate-as-alternative, make revert one action. Users iterate by alternating expansion with selective harvesting of AI output (NN/g's "accordion editing" / "apple picking") — support partial acceptance, not all-or-nothing.
• Feedback loops. Thumbs up/down with optional reason is the standard convention (response rates are low; don't depend on it). Richer signal comes from behavior: edits to AI output, regenerations, copy events, abandoned drafts. If feedback trains the model, say so (see privacy below). Close the loop visibly (HAX G16–G18).
• Agent progress disclosure. Long-running agentic tasks need a live activity log at "step" grain (searching X, editing file Y, ran tests), not a spinner: it enables trust calibration, early interruption, and post-hoc audit. Provide pause/stop/redirect mid-task; end with a summary of what was done and what changed. Emerging practice (2024–2026 coding and research agents), not settled research.
• Handoff to human. Make escalation discoverable at all times, transfer conversation context so users never repeat themselves, and hand off proactively after repeated failure (mirrors voice's reprompt-escalation ladder). Hiding the human exit to deflect tickets is a dark pattern and destroys trust in the bot.

Failure & safety UX

• Hallucination mitigation is a UI problem too: ground responses in retrieved/verifiable data, show sources, scope the assistant to domains it can answer, frame uncertainty honestly. No UI eliminates hallucination; the goal is errors that are detectable and cheap to check.
• Harmful-content refusals deserve error-message quality: say what can't be done and why (category-level), don't moralize, offer an adjacent path when a benign intent was likely misread. Over-refusal of innocuous requests is a real usability failure mode.
• Data-privacy disclosure: state plainly whether conversations train models, retention period, and who can see them; provide opt-out or ephemeral modes. Enterprise buyers treat this as gating.
• Disclosure of AI identity is becoming law, not just courtesy. The EU AI Act's Article 50 transparency obligations apply from 2 August 2026: users must be informed they're interacting with AI (unless obvious), AI-generated content must be machine-readably marked, and deployers must label deepfakes and AI text published on matters of public interest. California's SB 1001 (2019) and Utah's AI Policy Act (2024) impose narrower disclosure duties. Design default: label AI interlocutors and AI-generated content regardless of jurisdiction.

Anti-patterns

• Sparkle-icon everything: ✨ on every surface dilutes meaning and reads as marketing, not affordance. Label features by what they do ("Summarize"), not that they're AI.
• Forced AI insertion: interposing an assistant into flows that worked fine (search replaced by chat, mandatory AI summaries with no off switch). Solve a user need or leave it out.
• Fake anthropomorphism: simulated typing delays, first-person emotional claims ("I'm excited!"), hiding bot status to seem human — inflates capability expectations, creates misplaced social trust, and collides with disclosure law.
• Unlabeled AI content: publishing generated text/images/summaries unmarked shifts error liability onto the reader and (from Aug 2026, in the EU) violates Article 50.
• Over-automation of agency: auto-applying edits, auto-sending, auto-committing without review; burying the manual path; making the AI outcome the unremovable default. Automation should be adjustable — assist, draft, or act — with the user choosing the level.

Sources

• Amershi, S. et al. (2019). "Guidelines for Human-AI Interaction." Proc. CHI '19. Microsoft Research; operationalized as the HAX Toolkit (microsoft.com/en-us/haxtoolkit).
• Google PAIR — People + AI Guidebook, incl. generative-AI update (pair.withgoogle.com).
• Apple — Human Interface Guidelines: "Machine Learning" and "Generative AI" (developer.apple.com/design/human-interface-guidelines).
• Parasuraman, R. & Riley, V. (1997). "Humans and Automation: Use, Misuse, Disuse, Abuse." Human Factors, 39(2).
• Lee, J. D. & See, K. A. (2004). "Trust in Automation: Designing for Appropriate Reliance." Human Factors, 46(1).
• Buçinca, Z., Malaya, P. & Gajos, K. (2021). "To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI." Proc. CSCW '21.
• Nielsen Norman Group — AI topic collection (nngroup.com/topic/ai): "AI Chat Is Not (Always) the Answer"; "Prompt Controls in GenAI Chatbots"; "Accordion Editing and Apple Picking"; "Sycophancy in Generative-AI Chatbots"; "AI Chatbots Discourage Error Checking"; "10 Guidelines for Designing Your Site's AI Chatbots."
• EU AI Act, Art. 50 — transparency obligations, applicable 2 Aug 2026 (artificialintelligenceact.eu/article/50/; EC AI Act Service Desk). California SB 1001 (2019); Utah AI Policy Act (2024).