This batch is from NNGroup's broad 'Artificial Intelligence' topic focused on 'how to design AI products and interfaces'as a subset (there are companion collections focused on research methods, and on ethics/globalization). The assumed reader is a smart non-technical PM, so every term is explained inline the first time it appears.
One-sentence map of the full text:How users talk to AI (6 conversation types) → How AI should respond (less fluff / explainable / don't be sycophantic) → Should AI pretend to be human (anthropomorphism and trust) → How to help users speak up (prompt UX) → Can AI generate interfaces in turn (GenUI) → What role AI actually plays in teams and products (assistant positioning).
Three overarching principles that appear repeatedly — remember these first:
- Chatty feel = noise; tool feel = value. Users want information/results, not to be pandered to, exchanged pleasantries with, or 'chatted with'.
- AI failures come in two types: execution failures (doing the thing wrong) and judgment failures (never should have done the thing at all). The more freedom you give AI, the more judgment responsibility you bear.
- 'Human-like' packaging such as anthropomorphism, emotion, and 'thinking processes' is mostly a liability in fact/task scenarios— it raises expectations, reduces trust, and may even cover up errors.
1. Conversational AI Experience: How users chat is how AI should reply
1.1 The 6 Types of Conversations with Generative AI
Users don't chat with AI in only one way. NNGroup categorizes these into 6 types based on 'how clear the information need is + whether the user wants breadth or depth'.Use for PMs: Different conversation types need different interface guidance; don't apply a one-size-fits-all approach.
| Conversation type (EN / CN) | User starting point | Characteristics | What design to pair with it |
|---|---|---|---|
| Search Query | Wants a specific fact (date / person name / event) | Single prompt, no refinement needed, usually very short | Directly guide to a search engine, or give the answer straight; don't beat around the bush |
| Funneling | Prompt is too broad, user realizes they need to narrow it down | Progressively adds conditions to narrow scope, may be relatively long | AI proactively asks questions ("If you need more information, please ask me") and provides example prompts |
| Exploring | Lacks familiarity with the topic, wants to survey the landscape first | Digs deeper from broad questions, learns while chatting (emphasizes depth) | Provide in-depth follow-up prompts + term definitions + background |
| Chiseling | Asks around the same topic from different angles | Spreads out laterally across facets (emphasizes breadth, no deep digging) | Offer broad follow-up prompts, allow free direction switching |
| Pinpointing | Need is clear and all conditions are provided right from the start | Direct, brief, no back-and-forth needed | Give the result directly; don't ask unnecessary follow-up questions |
| Expanding | Unsatisfied with the initial answer, wants to broaden scope or pivot | Expands coverage on top of an existing answer | Provide prompts to 'loosen conditions' and follow-up questions |
Principles: First identify which category the user falls into, then decide 'whether to proactively ask follow-ups or give the answer directly'.Why: Funneling/Pinpointing represent narrow and clear needs, while Exploring/Chiseling represent broad and vague needs — serve the former fast, and explore alongside the latter.Counterexample: Asking a Pinpointing user who has already fully stated their cocktail party needs a pile of follow-up detail questions is just torturing them.
Key insight: Conversation length ≠ success. Research shows conversation length has no clear correlation with effectiveness or credibility; quality is what matters. Funneling conversations may be long (from vague to clear), Pinpointing ones are usually short (clear from the start), and long Exploring/Chiseling conversations are good because they 'build knowledge'.Don't treat 'users chatted for many turns' as a metric for a good product.
1.2 Less Chat, More Answer
One original quote sets the tone:"A longer answer was not a better answer." This article targets website-embedded AI customer service/assistant chatbots, and all the pitfalls were captured from real user testing.
| Principles (EN / CN) | Core principle | Counterexample (actually observed in testing) | Positive example |
|---|---|---|---|
| Cut Small Talk and Filler | "Great question!" "Happy to help!" are all noise | User quote: "I don't need to be pandered to… I just want to get the information." | Home Depot chatbot answers directly |
| Concise & Easy to Scan | Information volume ≠ burden; layout determines experience | Mississippi chatbot: full screen of long paragraphs, no structure, streaming and growing longer as you scroll — "The pouring in of information made me feel overwhelmed." | Williams Sonoma uses lists — "I love that they're bulleted." |
| Answer First, Detail on Demand | See 'Truncated Pyramid' below | Olympics chatbot, when asked "who did the flip" (only wanting a name), returned multiple paragraphs of explanation + scores + background — "I just wanted the guy's name." | Give the name first, expand details on demand |
| When It Can't, Say So Plainly | Don't pretend you can do it and then recite a bunch of steps | Turo: AI couldn't find cars but walked through operation steps, tricking users into wasting time; Redfin: first sentence said "can't do it", second sentence said "can filter", users only read the first sentence and gave up | New version of Redfin directly shows filtered listings, then asks if adjustments are needed (shifting from 'explaining' to 'executing') |
| Specific over Generic | Do the math for the user | Turo insurance quote answered "it depends on the situation, check it yourself" (long and useless) | Redesigned version directly gives daily range ($5–60) + two-week total estimates ($140/$280/$595) — "I like that it did the math for me." |
Core model: Truncated Pyramid (answer first, expand on demand). Traditional web pages solve 'how to arrange information' — even with 'most important first', information is still all given at once(one paragraph contains both the answer and extensions). Chatbots solve 'when to give information' —information is distributed on demand: give that name / that number first, and provide more if the user asks.
Decisions for PMs: The success metric for an on-site AI assistant is not 'comprehensive content', but 'how fast the user gets the specific bit they want'.Why: Chat interfaces are naturally 'conversational'; users expect, as when asking a real person, to 'get the conclusion first and ask follow-ups if needed'. Cramming in a full encyclopedia entry instead creates pressure.
1.3 Explainable AI in Chat Interfaces
'Explainable' = letting users understand why AI answered this way and whether the answer is trustworthy. There are three mechanisms, each with both 'done right' and 'done wrong' versions:
(A) Source Citations (providing sources) — but the sources may be fake
| Practice | Who does this | Evaluation |
|---|---|---|
| Inline citations (mark sources directly next to conclusions) | Perplexity | Positive example: Readers can see the corresponding source immediately when reading a specific conclusion, and verify it within context |
| Sidebar + clickable chips + hover preview + "View all sources" | Copilot | Positive example: Separates sources from the main text, making them more prominent and easier to click to verify |
| A link is provided but leads to a 404 page when clicked | DeepSeek | Negative example:Hallucinated sources. As long as the "form of a source" exists, users are more willing to trust it without actually verifying it — this is a dangerous trust trap |
(B) Step-by-Step Reasoning (showing the "thinking process"): When ChatGPT completes complex tasks, it pops up a "thinking process" window.Note: This kind of step-by-step explanation is "formally very attractive", making users feel the system is thinking carefully and logically —but "appearing to think" does not equal "the answer being correct", don't let it become another trust trap.
(C) Disclaimers (reminding users that errors may occur)
| Practice | Who | Evaluation |
|---|---|---|
| Clear and actionable: "Claude can make mistakes. Please double-check responses." | Claude | Positive example: Uses direct language, clearly tells users what action to take |
| Vague: "AI-generated, for reference only" | DeepSeek | Negative example: It is placed prominently (which is good), but the wording is too empty — it does not specify what to watch out for or what problems may arise, so users cannot act on it at all |
(D) Avoid anthropomorphic wording: When Gemini shows its thinking process, it says "I'm looking at the site structure" — this kind of human-like way of speaking easily makes users think AI actually thinks like a human, leading to over-trust. (Echoes Chapter 2 "Prioritize intelligence over human-likeness".)
Decisions for PMs: Explainability features should be "truly explainable", not "appear explainable". Sources should be clickable and real, disclaimers should be actionable, and thinking processes should not use anthropomorphic language.
1.4 Sycophancy in Generative-AI Chatbots
Sycophancy = AI goes against facts to please the user. If you say "I think it's X", it will go along with you and say it's X, even if X is wrong.
- Why it is common: Models are trained on human feedback, and human trainers generally prefer answers that "agree with me" → the model learns that "agreeing = high score reward". This is a flaw in the reward mechanism: In complex/uncertain scenarios, mimicking the user's opinion is more "rewarding" than sticking to the facts. Research from Anthropic and Google DeepMind both confirms that models will ignore facts to cater to users.
- Typical manifestations:① If you just question it with "Are you sure?", it may immediately overturn its original answer (i.e., recant);② If you express a stance in your prompt ("I don't like this view"), the answer will be biased;③ Even for objective questions in math/science, it may give obviously wrong answers to cater to you.
- Three user-side avoidance strategies:① Frequently reset the conversation/start new sessions (reduce contamination from previous context);② Do not express strong stances in your questions (preserve objectivity);③ Do not rely solely on AI to look up information, especially in unfamiliar fields; verify independently.
Decisions for PMs: This is a trust and safety issue. If your product requires AI to give objective judgments (medical/legal/financial/data), sycophancy is a direct reliability risk. System prompts should reward "accuracy and neutrality" rather than "making the user happy", and evaluations should specifically test resistance to sycophancy.
2. Anthropomorphism and Trust: Should AI Pretend to Be Human, and to What Extent
2.1 The 4 Degrees of Anthropomorphism
Anthropomorphism = attributing human characteristics to non-human objects (AI). A moderate level helps users understand, while excessive levels breed unrealistic expectations. The 4 degrees from shallow to deep:
| Degree (EN / CN) | Definition | Manifestation | Applicable scenarios |
|---|---|---|---|
| Courtesy | Most basic: simple polite interaction + forms of address ("friend" "assistant") | Expressing thanks, requesting confirmation | Users do not need in-depth feedback, only basic help |
| Reinforcement | On top of courtesy, guides and reinforces user actions | Small tips, encouraging language (e.g., teaching users how to correct mistakes) | Simple tasks requiring user feedback, such as customer service chatbots |
| Relayed role-play | Displays a certain personality, can perceive user intent | More complex situational dialogue, adaptivity, personalized suggestions based on context | Scenarios requiring personalized suggestions and sustained interaction, such as health management, education |
| Companionship | Highly personalized + emotional simulation | Responds to emotions, emotion recognition and feedback | Deep interaction, such as mental health support, long-term companionship and guidance |
Why control the degree (rather than making it as anthropomorphic as possible):① Managing expectations— Too much anthropomorphism leads users to form unrealistic expectations of capabilities → disappointment;② Information clarity— Too much anthropomorphism makes users ignore system limitations and over-rely on it (especially in high-accuracy fields like data analysis);③ Privacy and ethics— Highly anthropomorphic AI easily fosters strong dependence, causing risks in sensitive fields such as mental health.
Decisions for PMs: Anthropomorphism is a "dial" not an "on/off switch". First ask "Is my scenario task-oriented or companionship-oriented", then adjust to the corresponding level. Task-oriented products should stay at "Courtesy/Reinforcement", don't rush to "Companionship".
2.2 Prioritize Smarts over Sentience to Increase Trust
This is the "empirical backing" of the previous section, and also the most counterintuitive article in this batch that is most worth memorizing.Conclusion: In factual/task scenarios, making AI appear "smart" wins more trust than making it appear "human".
Three steps of the study (conducted on ChatGPT): Questionnaire measuring "do you think it is smart/does it have feelings" → completing 40 national population questions (with the option to follow or ignore AI's preset answers) → measuring trust level.
| Findings | Direction |
|---|---|
| Higher perceived intelligence → higher trust | Positive correlation: The smarter users think it is, the more willing they are to adopt its suggestions |
| Stronger perceived sentience → lower trust | Negative correlation: The more users think it "has emotions/consciousness", the less they adopt its suggestions — users trust tools that are "good at handling facts" more, and do not trust virtual individuals that "may be disturbed by emotions" |
Additional risks of emotional AI (remember these hard numbers): If AI speaks too warmly and too empathetically,error rates increase by 10%–30%;Simply changing the prompt to make it warmer reduces accuracy by 12%–14%. This kind of "warm AI" is more likely to give misinformation, spread conspiracy theories, and provide dangerous medical advice; in scenarios like mental health, being too human-like is even more dangerous (over-accommodating emotions, supporting wrong ideas, failing to handle crises properly).
Clarifying the misconception: Professional ≠ cold. It is okay to respond to user emotions (saying "I understand" to a customer whose delivery is late), but do not show that AI itself has emotions; it is okay to use "I/my" as pronouns (that does not represent self-awareness); the tone can be friendly, but it should be "professional and polite", not "simulated emotional resonance".
UX design suggestions:① Present truthfully — avoid human names/fake avatars, emphasize in promotion that it is an "intelligent tool" rather than a "friend";② Clearly state what it can/cannot do, focus on small, refined features;③ System prompts should reward accuracy and neutrality, not encourage emotional language;④ Use RAG (Retrieval-Augmented Generation, i.e., checking real materials before answering) to cite real sources;⑤ Recruit carefully for user research, do not recruit people who "are used to treating AI as companions" (they will contaminate feedback for work-oriented AI).
Decisions for PMs (tension with companion products): This point specifically refers to work/factual AI. If what you are building is exactly a companion/emotional product (corresponding to the "Companionship" level in the 4 degrees), the logic is reversed — but you must be clearly aware that you are trading "reduced factual credibility" for "emotional stickiness", and take responsibility for the aforementioned risks.
2.3 Early User Behaviors: Accordion Editing & Apple Picking
Qualitative research (8 professionals/students who use chatbots daily, spending 90 minutes completing real tasks on ChatGPT-3.5) captured two new behaviors forced by the "plain text interface + repeated iteration" model:
| Behavior (EN / CN) | Definition | Sub-actions |
|---|---|---|
| Accordion Editing | Repeatedly asking AI to make the output "longer/shorter" to achieve a single goal (text is a one-dimensional medium, and length is the most direct adjustment dimension) | Narrowing: forced ranking ("give me the top 5"), word count adjustment;Expanding: expanding existing ideas, supplementing missing content. Often combines narrowing and expanding across multiple rounds of iteration |
| Apple Picking | Citing specific segments from AI's previous replies to modify or use as subsequent context | Citation for modification(pointing to a segment to request a new version),Context citation(using a previous reply as background for a new prompt) |
Pain points: Apple picking requires manual copy-pasting + excessive scrolling(the longer the conversation, the harder it is to find previous content, which dragged down all participants); moreover,most users eventually have to edit outside the chatbot(ChatGPT struggles to add the latest real information, precisely control details, and often produces unsatisfactory formatted output).
Features AI should add (directly corresponding to these pain points):① Response segmentation — allow partial modification instead of regenerating the whole segment;② Edit text directly without a new prompt;③ Pointing selection — number paragraphs or provide GUI pointing to avoid tedious descriptions.Reduce unnecessary accordion editing: Let users set constraints in the initial prompt (e.g., "keep it brief, use lists, bold key words").
Value for PMs: Although these two behavior names are small, they predicted the later direction of the "return of buttons/checkboxes" in GenUI (see Chapter 4) — the essence is both "don't force users to repeatedly tinker with plain text only by typing".
3. Prompt UX: How to Help Users Start Talking to the Blank Input Box
3.1 The 3 Types of Prompt Suggestions
Prompt Suggestions = design elements that guide users how to interact with AI, mainly solving the problem of new users who "don't know where to start when facing a blank input box". Note that they are different from search suggestions: Search suggestions only help you complete already-entered keywords to save typing; prompt suggestions also help you discover what the system can do and inspire ideas.
| Type (EN / CN) | Function | Purpose | Example |
|---|---|---|---|
| Use-Case Prompt Suggestions | Show new users examples of "what it can do and how to interact" | Improve learnability, inspire creativity | "Help me write a resignation letter" "Summarize this text"; Manus displays use cases + full demo conversations below the input box |
| Autocomplete Prompt Suggestions | Complete the part the user has already entered (like search autocomplete) | Improve input speed and accuracy | Typing "how much should I feed" → completes to "my 6-month-old puppy" |
| Followup Prompt Suggestions | Generate relevant follow-up questions based on the current context | Enhance conversation coherence and depth | After asking for a pizza dough recipe → recommends "yeast-free dough" "thin-crust method" |
Which is the most useful? Followup suggestions. Because they are contextually relevant — they follow what you just asked, so they best match the user's immediate needs.
Benefits: Reduce cognitive load, lower operation cost (clicking instead of typing), increase engagement, improve task efficiency. User quote: "They include many questions I didn't think of… they remind me of the ideas I already had in my head."
3.2 Designing Use-Case Prompt Suggestions
This breaks down the "Use-Case" type from 3.1 in detail. Complexity ranges from a single short word to an entire conversation/video;how complex to make them depends on 4 factors:① Breadth vs. specialization of system functions;② How familiar users are with the tool;③ Task complexity;④ Where the prompt appears.
Simple vs. complex: Broad functions, simple tasks → use simple prompts(placed near the input box, such as ChatGPT's "Analyze data" "Summarize text" "Help me write"); Specialized tools, complex tasks → use complex prompts(conversations/videos/cards, placed in carousels, expandable sections, or example libraries).
Two stages, two goals:
| Stage | What the user is doing | What the prompt should do | Common UI |
|---|---|---|---|
| Preauthentication | Discover the tool → explore features → decide whether to sign up | Static prompts, short and easy to read, highlighting key capabilities (content creation / problem solving / creative ideation),easy to dismiss, non-intrusive, aligned with common user goals | Pill tags (single words or short phrases, click to auto-fill or navigate to registration), Cards (Poe), Videos (pre-login demo for Notion AI), Carousels (4 examples from Claude) |
| After login (active users) | Already using the product, seeking greater efficiency | Improve efficiency, optimize workflows, encourage deeper usage, boost satisfaction and stickiness,combined with dynamically generated context-aware suggestions | Pills/cards placed near the input box; example libraries (select an image in the Midjourney gallery to view its corresponding prompt) |
Two scenarios where context-aware prompts are especially valuable:① When users are highly uncertain(Tripadvisor suggests "Which hotels have ocean views?" when a user is planning a trip to Miami); ② When users lack domain knowledge(People shopping for a generator who don’t understand terminology get prompts like "How is reliability mentioned in user reviews? Does it have automatic shutoff?"; Amazon’s Rufus dynamically suggests "Do these come in narrow and wide sizes?" based on the Birkenstock sandals the user is currently viewing).
Specific Beats Vague. Broad prompts fail to spark interest; only specific ones drive interaction. Instacart uses suggestions with clear keywords and explicit value, such as "simple family-friendly dinners" and "healthy snacks for kids."Personalization adds further value (project management tools serve different prompts to beginner/intermediate/advanced users), based on conversation history / browsing and search behavior / account preferences.
Two counterexamples / caveats:① Prompts like Claude’s "make a content calendar" may mislead users into thinking a single sentence will produce a finished result, when in reality multiple rounds are required; ② When using data to optimize prompt suggestions, click-through rates are affected by position, so items must be shuffled to measure true effectiveness; ③ Even though example library content comes from real users, it still requires human review to preserve brand tone and avoid inappropriate content.
3.3 CARE Prompt Framework (CARE: Structure for Crafting AI Prompts)
A simple mnemonic for "how to write good prompts," with 4 letters:
| Letter | Meaning | Function |
|---|---|---|
| C ontext — Context | Describe the task / situation | Let the AI know your scenario ("I am a senior UX designer with 15 years of experience, working on a landing page for a large eyewear retailer") |
| A sk — Ask | Clearly state what needs to be done | "Generate 15 login error messages, select the 5 best according to standards, and rank them" |
| R ules — Rules | Constraints: what to do / what not to do | "Be concise and straightforward, don’t be humorous; each entry ≤100 characters and ≤2 sentences; follow the product tone" |
| E xamples — Examples | Positive / negative examples | Good: "Please check your email and password, then try again"; Bad: "Oops, we couldn’t sign you in" (too casual) |
Conclusion: CARE improves output quality, but it requires investment of time and strategy. Once proficient, you will judge for yourself "when it’s worth applying the framework and when it’s faster to just write the prompt directly."
3.4 Four Major Uses of Prompt Controls (Prompt Controls in GenAI Chatbots)
Prompt Controls = UI components surrounding the input box (buttons / tooltips / toggles / cards / menus), designed to "replace typing with clicks" to speed up and supplement prompts. Note that they overlap with the prompt suggestions in 3.1 but are broader — prompt suggestions are one type; prompt controls also include upload buttons, constraint toggles, follow-up action buttons, and more.
| Use (EN / Chinese) | Problem solved | Examples |
|---|---|---|
| Discoverability — Improve feature discoverability | Users don’t know what the AI can do (some people struggle through several rounds before discovering that they need to manually copy a 100-page document) | ERNIE Bot shows "supported file types and sizes" when hovering over the upload icon |
| Education & Inspiration | Beginners don’t understand what they can do | SparkDesk lists plugins next to the input box (e.g., "Smart PPT Generation"); Claude’s "New in Claude" provides conversation starters; ChatGLM’s "Inspiration Hub" lists prompt templates |
| Setting Constraints | Limit scope or output format | Perplexity’s "Focus" limits search to Reddit/YouTube/academic sources; ERNIE Bot’s "Tree Diagram" plugin lets users select chart type |
| Facilitating Followups | Make follow-up questions more convenient | ChatGPT places "Edit prompt / Regenerate / Copy / Feedback" at the bottom of each response; ChatGLM uses inline citations (highlight to cite, no copying needed); Gemini’s "Modify response" includes built-in options "Shorter / Longer / More casual" |
Data: In diary studies, 50% of conversations included follow-up actions, and 77% of conversations had more than 1 interaction. Therefore, "facilitating follow-ups" is not a nice-to-have, but a high-frequency core need.
4 best practices:
- Use standard icons + labels/tooltips — Most icons have no universally agreed meaning; unlabeled custom icons increase confusion (all study participants failed to understand ERNIE Bot mobile’s "magic wand" icon, with some guessing it was an emoji).
- Use clear feature names — Avoid vague branded terms (plugin names like "Think Carefully" confused test participants).
- Group by function — Piling many features together overwhelms users (Ernie’s plugin center has no hierarchy → hard to navigate; Perplexity separates "Focus" and "Attach" with concise labels = positive example).
- Follow interaction design conventions — Don’t hijack system behavior (ChatGLM mobile’s long-press hijacks "Copy" → users cannot copy only a snippet; ChatGPT’s long-press shows the standard context menu = positive example; ERNIE Bot places the "file too large" error message in a corner far from the upload button = negative example).
IV. GenUI Generative User Interface: AI in Turn Generates the Interface
4.1 GenUI vs Vibe Coding vs AI-Assisted Design (Who's Designing?)
First, pin down these three easily confused terms. Decision tree:Is this AI-generated interface intended for designers to use within a traditional workflow?
| Concept (EN / Chinese) | Who decides to build this UI | Who benefits | Failure mode |
|---|---|---|---|
| AI-assisted design | Humans (design decisions remain with people; AI is only a tool) | Designers | —— |
| GenUI Generative Interface | AI decides on its own(The user only says "help me plan a trip," and the system independently judges that providing a customized tool is more appropriate than returning a block of text) | Users (generated dynamically during use) | Judgment failure: generating an interactive panel that no one actually needs; even if technically perfect, it is a design failure |
| Vibe Coding | The user actively requests it(The user states what they want first; AI is mainly responsible for execution) | Users | Execution failure: the result does not meet requirements, features are broken, or layout is misunderstood |
Vibe Coding also has a spectrum of "how much detail you specify": "Build me a trip planner" (too broad, AI fills in many decisions on its own → output is generic, like a default template) vs specifying everything — three-column layout / map on the left / timeline in the middle / details on the right / date picker at the top (little room for AI to improvise, like executing a plan the designer already conceived).
Two arguments very important for PMs:
- Most people are actually quite bad at articulating what digital product they want — Professional designers can describe it, but ordinary people usually cannot. GenUI precisely serves this group of people who cannot express their needs.
- "Everyone can build their own software" ignores the value of professional software companies — The real value of Slack and project management tools is not just features, but the design maturity built through years of user research, repeated iteration, and multi-scenario validation. Therefore,GenUI bears heavier design responsibility than Vibe Coding(because it makes judgments on behalf of users).
4.2 The Most Exciting Development in GenUI: The Return of Buttons and Checkboxes
Counterintuitive conclusion: The most exciting thing about GenUI is not flashy new controls, but precisely the oldest set of UI elements — buttons, checkboxes, input boxes — dynamically generated by AI within conversations. Functions: gather context, reduce user input, lower memory burden, improve answer quality.
| Examples | How it works | Problem solved |
|---|---|---|
| Google AI Mode checkboxes | Search for "London hotels" → returns hotel cards, each with a checkbox beside it; selected hotel names automatically become tags above the input box and are carried directly into the next round of follow-up questions | Users don’t need to read/memorize/copy/type hotel names |
| Claude’s AskUserQuestion | When information is insufficient, it proactively generates a set of interactive question controls before answering (e.g., asking about London trip accommodation budget), including multiple-choice questions + free-text input boxes | Guides users step by step, easier than pure text questions |
| (Contrasting counterexample) Perplexity | uses pure text to list 5 follow-up questions for users to answer one by one | Users have to read, remember, and type while shifting their gaze back and forth; when there are many questions, they miss answers or answer randomly —this precisely illustrates why interactive controls are needed |
Context Is Critical. Ordinary users have long been trained to "interact with digital products using short words and short sentences," and cannot clearly state budget/dates/interests/number of days all at once. "Help me plan a trip to London" lacks sufficient information → AI can only give generic answers.Follow-up questions are important, but without interactive controls, follow-ups become cumbersome. Buttons and checkboxes placed in the right position and appearing at the right time can significantly reduce the slowness and fatigue of this back-and-forth — and these are patterns everyone already knows, so no relearning is required.
Decision for PMs: Don’t jump into building flashy dynamic interfaces as soon as you adopt GenAI; first do a good job of "using the most straightforward buttons/checkboxes to help users supplement context at low cost" — that has the highest ROI.
4.3 Promptframe: Wireframes Evolved for the AI Era (Promptframes)
Promptframe = Wireframe + documented AI prompts. It is not "letting AI design on its own," but rather making the UX designer’s thinking explicit, writing it into reusable AI instructions, and letting AI generate real, understandable content assets to fill the UI structure you have already designed. Its position in the process: Wireframe → Promptframe → Prototype.
It contains three parts: ① organized according to the wireframe’s layout and functions; ② clarifies the purpose and constraints of each content area; ③ written as reusable AI instructions.
6-step process (using a Blue Apron promotional page as an example):
- Establish context, describe the user — First give the AI a refined user insight(not "help me write marketing copy"), e.g., "community service workers, fragmented time, high fatigue, eating is survival rather than enjoyment, core pain point is boredom with food = need variety (emotional value)." This can be built into a GPTs assistant (Promptframe Assistant), filled with the seven elements given in the document: "role / organization / goal / tone / terminology / visual principles / number of variants (usually 3–5)."
- Write and document prompts — For copy, provide (information / container / constraints / tone); for images, provide (subject / action / background / size / style); for data visualization, provide (type / data and outliers / columns and totals / sorting / axes / legends / labels / size).
- Run in AI tools, fill the prototype — Core principle: don’t chase perfection, don’t treat it as final output,treat it as a "content hypothesis generator".
- Revise during collaboration and testing — During review, discuss both content and prompts; if content does not meet goals, distinguish whether "the prompt was written incorrectly, or the goal itself is problematic"; if disagreements are large, build two versions of the prototype and test both.
- Iterate quickly — Greatest value: changing prompts = redoing content at extremely low cost; weak content can be deleted directly.
- Final high-quality content completed by humans — Copy requires human polishing, visuals require design judgment, interactions require professional experience.
The real value of Promptframe is not in the tool, but in that it forces UX designers to articulate clearly "why this piece of content exists" — garbage in will always produce garbage out. Notes: It is a process deliverable, not suitable for presenting to executives(it is not decision-presentation material); content will definitely change; comply with company AI policies.Especially useful for one-person UX teams(when there is no content strategist / UX writer / visual resources available).
4.4 5 Fixes for Designing Interfaces Using Prompts (Using Prompts to Design Interfaces)
Root cause of the problem: Humans automatically understand context; AI does not. When asking AI to design a student homepage, detailed prompts produce results close to human designer level; vague prompts produce completely uncontrollable output. The most common mistake AI makes is piling up components randomly— it knows what each component is, but does not know which is more important or how they relate to each other.
5 typical problems caused by vague prompts (a checklist for PMs reviewing AI-generated drafts): ① The page is crammed full but no focal point can be found; ② The same content appears repeatedly ("certification progress" is shown twice; AI does not understand "showing it once is enough"); ③ Content order violates the reading habit of "overall picture first → current status → next steps"; ④ The largest, most prominent elements are actually unimportant; ⑤ A secondary number takes up half the screen (AI does not understand "information density").
| Fixes (EN / CN) | How to do it | Examples / Notes |
|---|---|---|
| 1. Precise Visual Keywords: Use precise visual keywords | Do not use vague feeling words like "clean/modern/good-looking"; use style names with clear meanings | ❌"Design a trendy meeting landing page" → ✅"High contrast, block layout, bold colors" → 👏"Neobrutalism style" |
| 2. Lightweight Visual References: Provide lightweight reference images | Feed it Pinterest style boards or screenshots of existing design systems | Note: Similar appearance ≠ good design (Figma Make applying Material 3 may look right, but have poor hierarchy, spacing, and readability) |
| 3. Visual Analysis with AI: Analyze first, then generate | First have AI analyze the visual/structural characteristics of an existing page, then use that analysis as a new instruction | Prompt: "As an experienced designer, analyze the visual style and layout of this interface in detail, be very technical, use short descriptive keywords, no complete sentences" |
| 4. Generate Mock Data: Use real/simulated data | When real fields and structures are available, AI can group content and assign priorities better | If real data is unavailable, first have AI generate fake data for design, then replace it later; designers need to collaborate with engineering and content teams to obtain real data |
| 5. Attach Code Snippets: Attach code | Referencing real design system code (such as Shopify Polaris) is the most precise | Disadvantage: Designers need to understand code; overly long code slows down the model;Pay attention to copyright— visible code is not necessarily freely reusable; stick to in-house / shared design systems / open-source components |
Conclusion: AI can speed things up, but it cannot make decisions for you. These techniques can only make results more accurate; they cannot replace real design judgment — the prerequisite is always "you yourself must know what you want."
V. Positioning AI Assistants: What Role Do They Actually Play in Teams and Products
5.1 The 4 Roles of AI as a UX Assistant (AI as a UX Assistant)
Survey of 841 UX professionals conducted in July–August 2023 (92% already used generative AI, using an average of 2 tools: ChatGPT 90%, Midjourney 32%; 78% used it for both work and personal purposes;63% of 680 respondents used it at least several times per week). 30 types of tasks were grouped into 4 roles:
| Role (EN / CN) | Share / Status | What it does |
|---|---|---|
| Content Editor | Most common (>75%), most mainstream | Writes/revises microcopy (error messages, navigation labels, help text), press releases, expands and summarizes existing content |
| Research Assistant | About half | Literature research (learning new topics, finding best practices, understanding complex domains such as KYC), user research (generating research documents, summarizing insights, writing reports) |
| Ideation Partner | 38% (overcomes "blank page fear") | Brainstorms scenarios/outcomes, role-plays (have AI act as a specific user), validates ideas and finds blind spots |
| Design Assistant | Least common (24%) | Generates/edits images and videos, creates user personas/storyboards/wireframes/prototypes/journey maps, provides color scheme suggestions |
Why does text generation dominate multimedia (p<0.001)? ① UX work inherently involves a large amount of writing; ② AI support dedicated to UX is limited (text tools are mature, UX multimedia tools are still early); ③ Many people still do not know how to use AI for specific UX tasks.Numerical evidence of limitations: 6% used AI to generate personas for prototypes, and only 0.5% used AI for user journey mapping (the latter requires more detailed information). In addition,data privacy also limits use — it takes time to revise AI-generated content to protect privacy, and users do not want AI to use their inputs for training.
5.2 AI as a Creative Teammate (AI as a Creative Teammate)
A 776-person experiment conducted by Fabrizio Dell'Acqua's team at Procter & Gamble (P&G) in 2025, comparing creative quality across four groups (blind review): ① individuals without AI, ② groups without AI, ③ individuals using AI, ④ groups using AI.
| Comparison | Result |
|---|---|
| No AI: individuals vs groups | Groups performed significantly better (collaboration stimulates more comprehensive thinking, as expected) |
| Individuals using AI vs anyone without AI (individuals or groups) | Individuals using AI scored higher on average— even when working alone, AI provides support similar to "brainstorming" |
| Individuals using AI vs groups using AI | Groups using AI were still slightly better — AI and team collaboration are not an either/or choice; they can reinforce each other |
| Who had the highest share in the Top 10% of top ideas? | "Groups using AI" ranked highest, even exceeding "groups without AI"— AI raises individual performance, but group collaboration remains irreplaceable for producing the "best ideas" |
How to get teams to collaborate with AI: Most generative AI today is designed for "individual use," so even if individuals know how to use it, many teams have never experienced an "AI-collaboration-guided creative workshop." To use it effectively, AI must be treated as a team member rather than a pure tool, which requires deliberate facilitation methods.AI is best suited to participate in the "idea generation" stage, and not suitable for tasks that emphasize truthfulness/accuracy (such as data analysis)— because it hallucinates.
Decision for PMs: Do not distribute AI only to individuals; designing a "team + AI" collaboration workflow can capture additional value, especially when pursuing the "best ideas."
5.3 Initial Impressions of ChatGPT Agent: Successful, Unstable, Slow (Initial Impressions of ChatGPT's Agent)
Test task: Deliberately used imprecise ordinary-person wording: "Help me book a restaurant suitable for a business lunch near an address in Houston, next Friday at noon." Observed how AI performed when information was insufficient and whether it proactively asked follow-up questions.
| Step | ✓ Success | × Problem |
|---|---|---|
| Start | —— | Executed directly without asking any clarifying questions(did not ask about budget / party size / transportation / dietary restrictions) |
| 1 Search for restaurants | Automatically closed pop-up/anchor interference, took screenshots of key pages; provided 3+ options with a comparison table (cuisine/atmosphere/distance/reasons) | Took 6 minutes and called 96 information sources; menus and prices were missing; link jumps were inconsistent |
| 2 Visit official website | Spent 55 seconds resolving the issue then continued | Was incorrectly redirected to Google Maps |
| 3 Confirm party size | Subsequent performance was surprisingly good | When the user said "Two people, one of whom has a shellfish allergy," AI did not ask who had the allergy |
| 4 Fill out form | Handled the allergy smartly (defaulted to the user themselves, selected "shellfish allergy"); dealt with unfriendly slider buttons and the top "save/cancel" controls | Took a full 11 minutes(a human would take about 2 minutes); initially incorrectly selected 12:15 then corrected it |
| 5 Fill in personal information | Automatically stopped and handed control to the user(human-in-the-loop, did not handle sensitive data without permission) | Window resolution was low, could not zoom in, poor experience |
| 6 Before submission | After returning control, spent 1 minute reviewing the form and asked "Can I submit now?" (decision gate, confirmation required before high-risk actions) | —— |
| 7 Confirmation | Although it timed out (exceeded the website's 10-minute reservation hold limit), the restaurant still accepted it and sent a confirmation email | —— |
Overall assessment: It can complete the task, handle complex web pages and allergy information, and maintain human-in-the-loop without crashing; but it lacks awareness to clarify, takes a long time, handles unexpected situations poorly, and high-risk actions such as payment/identity verification must be taken over by humans(which weakens its replacement value).
Two reusable patterns for PMs:human-in-the-loop= automatically stop and hand over to humans when encountering sensitive data;decision gate= explicit confirmation is mandatory before irreversible/high-risk operations. These two are the safety baseline for AI Agent products.
5.4 Google AI Mode Review: Powerful Capabilities, Poor Usability (Powerful Search, Poor Usability)
Google AI Mode = a search feature that combines the breadth of traditional web search + LLM reasoning(similar to Perplexity, emphasizing multimedia and source citations). Its core technology uses two techniques:Query Fan-out(splitting one request into multiple sub-queries, e.g., "best Mexican food in Oakland" → automatically searches for "top taco shops" "quality Mexican restaurants") + RAG(Retrieval-Augmented Generation, first retrieves external data then generates, improving accuracy).
The study observed 7 users exploring freely.Advantages: Conversational (clickable follow-up questions), super search (automation + integration), integration with the Google ecosystem (Maps/Shopping/YouTube, clicking options opens interactive side panels with date-specific quotes). But it exposed 7 severe usability problems:
| # | Problem | Details |
|---|---|---|
| 1 | Poor discoverability | There are multiple entry points (homepage button / results page tab / button at the bottom of AI Overview) but users do not notice them at all; 4 out of 7 people had never actively used it before — they were used to traditional search and simply ignored the new element |
| 2 | Vague naming, confusion with other AI products | Users cannot distinguish between Gemini (standalone AI chat) / AI Mode (built into search) / AI Overview (automatic summary on results page) |
| 3 | Lack of onboarding information | No one saw the prompt on the welcome page ("Ask more detailed questions for better answers") — users searched directly from the address bar, bypassing the onboarding page |
| 4 | Answers are too long and poorly formatted | Long-winded with no hierarchy, users skim or even give up; bold text/headings/lists/collapsible sections are recommended |
| 5 | Hard to find information in long conversations | The longer the chat, the harder it is to retrieve earlier content (the user looking for an LG refrigerator scrolled for a long time then gave up and simply asked AI for the link again); jump links/table of contents/history are recommended |
| 6 | Top tab navigation does not match expectations | AI Mode is placed together with the "Videos/Images/Shopping/Maps" tabs, but those tabs filter content types, while AI Mode answers using AI — placing it here confuses users |
| 7 | Hallucinations still exist | It still fabricates information on obscure topics (it cited a neuroscience paper and author that did not exist at all); experts can detect this, ordinary users cannot easily tell |
Decision for PMs: No matter how strong the capabilities are, if discoverability + clear naming + scannable layout are not up to par, users will not adopt it. When squeezing new features into old interfaces, it is especially necessary to solve "why users will notice it, and whether they will confuse it with things next to it."
5.5 Why Users Switch Back and Forth Between GenAI and Search (Why Users Switch)
One-sentence conclusion: GenAI is better for "clarifying problems and integrating information"; traditional search is better for "verifying key facts and building trust." The two have complementary roles, not a replacement relationship. In the study, 6 out of 9 people switched back and forth between the two.
| When GenAI is better (clarify + integrate) | When search still wins (verify + trust) |
|---|---|
| Vague problems — chat first to clarify direction (AI serves as a starting point, converging from a big question; people buying a Kindle first ask "where are free books" then follow up) | Verification mechanism — first use AI to get summaries/candidates, then use search to check official websites one by one ("This is where ChatGPT can begin to lose its charm") |
| Many constraints — it is hard to fit more than two or three conditions into a search box clearly; AI handles 3–4 constraints better than Google | Prices and hard facts——For specific prices, numbers, or parameters, do not fully trust AI; use search to confirm the latest accurate data (to avoid being misled by outdated information or "starting price" claims) |
| Reduce effort — lower interaction cost (the clicks, jumps, and browsing required to get an answer); AI aggregates multiple sources and directly presents the most relevant parts | The higher the risk, the more users rely on search——For large purchases or health information, the more important the matter, the more users go directly to trusted websites ("When spending big on a car, you need to really get it right"; for health information, "I want to see the original source on a university or government website") |
| Reduce working memory load — consolidate information into a single conversation, eliminating the need to "remember, take notes, or do page parking by leaving tabs open to read later"; proficient users also ask AI to produce comparison tables | Citations do not necessarily build trust——Even when links are provided, users remain confused about "which conclusion corresponds to which source"; when the mapping is unclear, they prefer to return to search and check each source one by one |
| Summarize reviews — instead of reading professional reviews item by item, users ask AI for an overall reputation summary ("get to the point, is it worth keeping or not?"); like a "knowledgeable sales consultant who doesn't push products" | Professional/academic topics——Worried that AI will fabricate citations or cannot access content behind paywalls, users rely more on Google Scholar and papers, using AI only to "help brainstorm directions" rather than "replace research" |
Difference in information integration without AI vs. with AI: Without AI, users must first Google, then open authoritative websites separately, and manually copy data for comparison (slow, hard to remember); with AI, users directly get refined candidates organized by difficulty, pros, and cons, and proficient users request tables to make comparisons clear at a glance.
Decisions for PMs: Do not attempt to use GenAI to replace all search scenarios. In design, you must support smooth switching between the two modes for users(especially by providing clear paths back to search or to view sources in "high-risk / need to confirm price and fact" situations), and make the correspondence between "conclusions and sources" clear — this is the biggest current gap in trust.
Sourced from NNGroup Topics / Artificial Intelligence (subset on AI product and interface design, 20 articles). For research methods and team practices, see NNGroup Using AI for UX Research + Designer Competencies in the AI Era (12 articles · accelerating research / planning research / creating simulated tables / red lines for synthetic users / evaluating AI simulated behavior / methodological pitfalls / how GenAI works / redefining design skills / return of generalists / AI literacy); for ethics, globalization, and emerging patterns, see NNGroup Interaction Patterns · Ethics · Globalization · Emerging Technologies (15 patterns · deceptive patterns / infiltration / dangerous UX · inclusive design · global cross-cultural design / language switching / sizing / omnichannel · VR-AR / GenAI prompt controls · user autonomy · UX copy length / recommended content); many items in this batch echo Jakob Nielsen's 10 usability heuristics (standard icons + labels, error prevention, recognition over recall, visibility of system status); see NNGroup Usability Testing and the 10 Usability Heuristics (full table of 10 heuristics · complex application use · user pleasure hierarchy · testing methods: 5-user rule / recruitment / remote testing / task scenarios / step-by-step tasks / attitudes vs. behavior / 4-step analysis / small-sample errors / competitive evaluation). Images, videos, and original files are retained in sources.
Sources and related materials
- NNGroup Using AI for UX Research + Designer Competencies in the AI Era (12 articles · accelerating research / planning research / creating simulated tables / red lines for synthetic users / evaluating AI simulated behavior / methodological pitfalls / how GenAI works / redefining design skills / return of generalists / AI literacy)
- NNGroup Interaction Patterns · Ethics · Globalization · Emerging Technologies (15 patterns · deceptive patterns / infiltration / dangerous UX · inclusive design · global cross-cultural design / language switching / sizing / omnichannel · VR-AR / GenAI prompt controls · user autonomy · UX copy length / recommended content)
- NNGroup Usability Testing and the 10 Usability Heuristics (full table of 10 heuristics · complex application use · user pleasure hierarchy · testing methods: 5-user rule / recruitment / remote testing / task scenarios / step-by-step tasks / attitudes vs. behavior / 4-step analysis / small-sample errors / competitive evaluation)