User Experience for AI

Stop Building Chatbots

The default AI interface is a chat window. It's almost always wrong. Chat puts the burden on the user to know what to ask, how to phrase it, and what's possible. That's a terrible experience for anyone who isn't already an AI power user.

Instead, build structured interfaces. Buttons, dropdowns, templates, and pre-filled forms. Guide the user toward the input your model works best with. A form that says "paste your job description here" converts ten times better than a blank chat box that says "how can I help?"

Edit, Don't Create

The most successful AI interaction pattern is: the AI generates a first draft, and the user edits it. This works for emails, code, designs, summaries, and recommendations. The user feels in control. The AI handles the blank-page problem.

Design your UI around editing. Inline editing, track changes, version comparison, "regenerate this section" buttons. The AI is the first drafter. The human is the editor-in-chief. This relationship feels natural because it mirrors how humans already collaborate.

Loading States and Streaming

AI responses take seconds, not milliseconds. In traditional software, a 3-second wait feels broken. In AI, you need to reframe waiting as processing. Show what the AI is doing: "Analyzing your document..." "Generating recommendations..." "Checking against best practices..."

Streaming responses — showing text as it's generated — is the single biggest UX improvement you can make. It reduces perceived wait time by 60-70%. Users start reading immediately instead of staring at a spinner. Every major AI product streams for a reason.

Design for the Wrong Answer

Every AI output should be treated as a suggestion, not a declaration. Your UI must make it trivially easy to reject, edit, or regenerate AI outputs. If a user has to accept an AI result because there's no alternative, your UX has failed.

Add "thumbs up/down" on every AI output. Not just for feedback — it teaches users that evaluation is part of the workflow. Add "try again" buttons. Add "edit this" links. The user should always feel like the pilot, never the passenger.

AI for Everyone

AI interfaces often exclude people with disabilities. Streaming text can be unreadable for screen readers. Loading animations can trigger seizures. Auto-playing AI responses can overwhelm users with cognitive disabilities. Build with accessibility from day one — it's not a feature, it's a responsibility.

Progressive Disclosure for AI Complexity

AI products often have a complexity problem: power users want fine-grained control (model selection, temperature, output length), while casual users just want the magic trick. Progressive disclosure solves this by hiding complexity behind layers.

Layer 1 — One-click: The default experience. User provides the minimum input, AI uses smart defaults for everything else. This should handle 70% of use cases perfectly. No settings. No options. Just input and output.

Layer 2 — Guided options: An "Advanced" toggle reveals 3-5 meaningful controls. Output length (short/medium/long). Tone (formal/casual/technical). Format (bullet points/paragraphs/table). These controls use human language, not technical parameters.

Layer 3 — Expert mode: For the 5% of users who want full control. Model selection, temperature sliders, custom instructions, raw prompt editing. Hidden behind a developer tools panel. Never shown to casual users.

The key insight: each layer should be fully functional without the layers above it. A user who never opens advanced settings should have an excellent experience. Progressive disclosure means every user gets exactly the complexity they want — no more, no less.