UX Interaction Paradigms in the AI Era (7 Types: AI Embedded in Traditional UI / Agentive / Conversational / Human-in-the-Loop / Hyper-Personalization / Multi-Modal / Trends)

This article answers the first question every PM should think through clearly:In my product, where exactly does AI "live", and in what posture does it interact with people? It does not talk about how to draw a specific button (that is the job of the Shape of AI AI Product UX Pattern Library (7 major categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines) pattern library), nor does it dive deep into empirical research on conversational interfaces (that is the job of NNGroup AI Product and Conversational Interface Design (20 articles · 6 types of conversation / 4 degrees of anthropomorphism / sycophancy / intelligence over human-likeness / explainable AI citations and disclaimers / cut the pyramid to get straight to the answer with no nonsense / 3 types of prompt suggestions · CARE framework · 4 uses of prompt controls / GenUI vs Vibe Coding · return of buttons and checkboxes · Promptframe · designing interfaces with prompts / 4 roles of AI assistants · creative partner P&G 776-person experiment · initial review of ChatGPT Agent · Google AI Mode · why users switch between GenAI and search) ), but first lays out the seven macro postures in which AI intervenes in products, so you can pick the right "form" before you start building.

One-sentence map of the seven paradigms (sorted from shallow to deep "AI intervention feel"): Human leads, AI waits to be summoned → Human leads, AI helps in the background → Human and AI work side by side → AI takes initiative → Human feeds information to AI in the most natural way → AI becomes the system foundation → Pure chat.

Three overarching core principles that run through the whole article, remember them first:

AI = helps you do things, not does things for you. Almost every good case leaves users an opening to "undo/ignore/take over". Keeping control in the user's hands is the safety baseline for AI experiences.

Don't leave users staring blankly at an empty input box. Preset buttons, natural language, photo taking and handwriting, highlight selection — all paradigms essentially solve the problem that "ordinary people don't know how to start directing AI".

The deeper AI intervenes, the heavier the design responsibility. Silent background optimization, proactive prediction, standalone hardware — every layer deeper in intervention feel raises the cost of errors, privacy risks, and trust collapse. The depth must match the responsibility you can bear.

Quick Reference Table for the Seven Paradigms

Paradigm (EN / CN)	One sentence	AI intervention feel	Typical products	Core scenarios
Conversational	Pure chat where you ask and AI answers	Shallow (passive response)	ChatGPT / Gemini / Dot	General Q&A, companionship
AI in old UI: AI embedded in traditional UI	Attaching AI to familiar old software	Shallow (embedded)	Notion / Figma / Linear / Grammarly	Adding AI to existing products
Agentive UX	AI works silently in the background	Medium (proactive but not pushy)	Gmail Smart Compose / Waze / Netflix	Efficiency improvement, noise reduction
Human-in-the-loop	Human and AI co-create side by side	Medium (collaborative)	Photoshop / v0 / Krea	Creation, writing, design
Hyper-personalization	AI as system foundation / standalone device	Deep (full perception)	Rabbit R1 / aiOS concept	Next-generation OS, hardware
Multi-modal	Shooting, writing, seeing, listening, drawing, and moving can all direct AI	Medium-Deep (multi-channel)	Photomath / Apple Pencil / Arc	Natural input
Trending	Interface evolution history + macro directions	— (meta perspective)	——	Strategic judgment

How to use this table: Before building new AI features, first ask "which column do I fall into". Adding AI to an old product → look at the second column (AI in old UI); wanting AI to proactively save users trouble → the third column (Agentive); building creation tools → the fourth column (Human in the loop).Most mature products are a combination of several paradigms (for example, Notion is simultaneously "Conversational + AI in old UI + referencing for trust").

1. Conversational: The Starting Point of the AI Revolution

What it is: The simplest posture — users ask AI in a linear back-and-forth way, just like chatting with a person. No special commands, no learning threshold, ready to use right after opening. ChatGPT, Gemini, and Bing AI all follow this model.

Why (it is the starting point, not the end point): The core reason conversational UI ignited the AI wave is not "how strong the technology is", but that everyone can use it, everyone can understand it — it packages powerful AI computing power into "chatting with an all-knowing friend". This is the first time AI has been demystified for ordinary people. But be clear: pure linear conversation is only an entry-level posture, and the following six paradigms all address its shortcomings (not knowing what to ask facing an empty box, being unable to scroll back in long conversations, low efficiency of pure text).

Design key points (5 points extracted from real products):

Stay authentic, don't pretend to be human: AI directly says "I am an AI" and does not pretend to be human (echoing the NNGroup AI Product and Conversational Interface Design (20 articles · 6 types of conversation / 4 degrees of anthropomorphism / sycophancy / intelligence over human-likeness / explainable AI citations and disclaimers / cut the pyramid to get straight to the answer with no nonsense / 3 types of prompt suggestions · CARE framework · 4 uses of prompt controls / GenUI vs Vibe Coding · return of buttons and checkboxes · Promptframe · designing interfaces with prompts / 4 roles of AI assistants · creative partner P&G 776-person experiment · initial review of ChatGPT Agent · Google AI Mode · why users switch between GenAI and search) principle of "intelligence over human-likeness").
Ask clearly when unsure: When AI does not understand, it asks follow-up questions to confirm first, instead of making up answers.
State capability boundaries at the start: Tell users "what I can do and what I cannot do" to manage expectations.
Allow editing questions and stopping at any time: Users can edit historical inputs and clear conversations at any time — control remains in human hands.
Preset templates to lower the threshold: Provide quick entries such as "write an email / explain a concept / generate a recipe" on the first screen to cure "empty box fear".

Cases and variants:

ChatGPT: As the basic paradigm, it uses a typing cursor animation during loading to indicate "AI is working", a feedback button at the bottom to collect data, and supports voice read-aloud (a multi-modal extension).
Dot(Emotional companionship / diary reflection type): Uses chat bubbles + human-like character-by-character typing to enhance intimacy;Retells user experiences in the third person ("Mei enjoys cooking…") to create the resonance of "being understood and recorded"; pinching the screen brings up a calendar to review past conversations. This is a representative of the "companionship" end of conversational UI, directly relevant to PMs building AI chat / companion products.

When to use / Counterexamples: First choice for general assistants and emotional companionship products.Counterexample: If what the user wants is "change a formula in the spreadsheet" or "make this paragraph more concise", forcing a standalone chat box is a detour — you should use the "AI in old UI" below, letting AI live right at the task site.

2. AI in old UI: Attaching AI to Familiar Products

What it is: Instead of building from scratch, AI is embedded into old software users are already familiar with (Word, Notion, Photoshop, Figma, Linear) to do completion, search, generation, and layout.

Why (this is the optimal solution for adding AI to old products): Users have zero learning cost — AI naturally integrates into existing workflows without interrupting work. The practical significance for PMs: most products in your hands are not AI apps built from scratch, but "adding AI capabilities to existing products"; this paradigm has the highest ROI, lowest risk, and is easiest for existing users to accept. The core principle is "add-on enhancement, no mandatory interface restructuring".

There are 6 sub-patterns under this paradigm, from shallow to deep:

2.1 Natural language search

Turns "saying a sentence" into a universal remote control. No need to click options, pull down dropdowns, or flip through menus; a single natural language sentence ("best-selling product last year" "open bugs with SLAs") can search, filter, jump pages, or even trigger actions.

Linear: Adds an "AI Filter" on top of the original "filter + list" architecture, which accepts natural language and automatically parses it into standard filter conditions —zero changes to the old interface, rapid deployment of new capabilities.
Stripe developer documentation: Adds "Ask AI Assistant" to the search bar; natural language questions → two-stage loading ("Gathering sources" → answering while writing) → clickable sources listed below the answer → "Was this useful?" feedback.
Design key points: Retain traditional UI, output structured results compatible with original system logic, provide preset prompts ("assigned to me") to incidentally teach users how to phrase requests.

2.2 Referencing for trust

Make AI answers "have sources" so users dare to trust them. Because everyone knows AI can "hallucinate" (make things up), clickable, verifiable reference sources must be listed below answers.

Notion: Marks answers with "8 pages found", lists cards of actually cited pages that can be clicked to jump — "not only results, but also why these results".
Arc Search / Adobe Acrobat: Citation is not just links, but"semantic citation" based on original text position (hovering highlights the actual paragraph in the PDF), paired with a clear disclaimer ("AI answers may be inaccurate, please verify").

Core: Trust comes from "being verifiable, traceable, and controllable"; it is a rigid requirement especially in high-authority scenarios such as technology, compliance, healthcare, law, and scientific research. This point is aligned with NNGroup AI Products and Conversational Interface Design (20 articles · 6 conversation types / 4 degrees of anthropomorphism / sycophancy / intelligence prioritized over human-likeness / explainable AI citations and disclaimers / cut-the-pyramid to get straight to answers with no fluff / 3 categories of prompt suggestions · CARE framework · 4 uses of prompt controls / GenUI vs Vibe Coding · return of buttons and checkboxes · Promptframe · designing interfaces with prompts / 4 roles of AI assistants · Procter & Gamble 776-person creative partner experiment · initial review of ChatGPT Agent · Google AI Mode · why users switch between GenAI and search) deeply echoes the "explainable AI / source citation" section in that piece — that piece also adds a key counterexample:A "hallucinated source" that gives a link but leads to a 404 when clicked is instead a dangerous trust trap(users trust it more simply because it "has the form of a source," without actually verifying it). Therefore citations must be "actually clickable and authentic."

2.3 Context bundling

Embed preset prompts inside buttons, solving the problem of "not knowing where to start with a blank input box." When a user clicks a button ("Write a resignation letter," "Summarize this report," "Explain this image for me"), the AI automatically brings in the page context, so the user does not need to write the prompt themselves.

Grammarly: Selecting text automatically pops up "Rewrite with Grammarly," which turns "style selection" into buttons (assertive/confident/diplomatic), and also provides Insert/Retry/Shorten after output —output is not the end point, but the starting point for further fine-tuning.
Apple(Mail/Notes): Non-intrusive embedding: select text and click "Rewrite/Proofread," the AI has almost no "presence," keeping the system extremely minimal.
Figma (spatial AI): Click the "magic wand" to bring up a 2D tone matrix(Casual↔Professional × Basic↔Technical),Dragging the orange dot = adjusting prompt parameters, but far more intuitive than typing — the AI is positioned as a "tone adjuster" rather than a "content generator."
Monday.com: AI is embedded into the formula bar / paragraph "⋯" menu, precisely optimized for "writing formulas" and "summarizing paragraphs" rather than general-purpose chat.

Core principle: Reduce cognitive load + context awareness (the system already knows what you want to write, no need for you to say it) + in-flow assistance (no jumping out of the workflow). This is exactly the productized implementation of "prompt suggestions / prompt controls" in NNGroup AI Products and Conversational Interface Design (20 articles · 6 conversation types / 4 degrees of anthropomorphism / sycophancy / intelligence prioritized over human-likeness / explainable AI citations and disclaimers / cut-the-pyramid to get straight to answers with no fluff / 3 categories of prompt suggestions · CARE framework · 4 uses of prompt controls / GenUI vs Vibe Coding · return of buttons and checkboxes · Promptframe · designing interfaces with prompts / 4 roles of AI assistants · Procter & Gamble 776-person creative partner experiment · initial review of ChatGPT Agent · Google AI Mode · why users switch between GenAI and search); this is exactly the productized implementation of "prompt suggestions / prompt controls" in that piece.

2.4 Living documents

Leverage AI's "parallel processing" superpower— humans can only handle a few things at a time, while AI can process hundreds of files simultaneously, batch extract times, locations, people, and emotions, and conduct large-scale horizontal comparisons.

Elicit: Embed academic paper processing into structured tables, automatically compare multiple documents (fields such as Hypothesis/Findings), and include confidence prompts like "Low confidence / answer may be inaccurate."
V7 Go: Create a new AI attribute column in the table (e.g., "Extract emotion keywords"), select the GPT tool + image source, and generate structured fields (Joy/Anger/Admiration colored tags) for dozens or hundreds of images at once.

Why: This is a typical application of AI's "capacity expansion" ability, suitable for data-intensive research. In terms of design, it still "hangs on the old logic of tables," with low migration cost.

2.5 Infinite Canvas

Integrate AI into the spatial workspace of an "infinite whiteboard"; saying "Help me summarize" or "Organize icons" can mobilize the AI, suitable for design, planning, creativity, and group discussion.

Figma: Frame an area and click "Replace content," the AI reads the spatial structure to complete the content (e.g., automatically filling in a recipe table), uses yellow highlighting for processed areas and transparent transitions to emphasize that "AI is a collaborator, not a takeover," and allows fine-tuning and rollback after output.
tldraw:Canvas = input device, input "Arrange rectangles along the curve," and the AI directly performs the "geometric action" — language becomes spatial operation, no pop-ups, no confirmations, instant execution.

2.6 Non-linear conversation

A combination of "mind map + chat + editor." Instead of chatting from start to finish, you can click any content block to have an in-place conversation; answers are displayed at the corresponding position, you can switch topic nodes at any time, run multiple branches in parallel, and ultimately form an AI-integrated knowledge graph.

Why: Replace linear timelines with spatial layouts, solving the biggest pain point of linear conversations — being unable to scroll back to earlier content in long conversations (NNGroup calls this "page parking / excessive scrolling"). Suitable for literature research, product deconstruction, knowledge maps, and other non-linear thinking. The AI "remembers" the context for each branch, and asynchronous generation does not block reading.

When to use AI in old UI overall / counterexamples: When adding AI to existing products where users already have firmly established usage habits, this is the most stable path.Counterexample: If you are building a brand-new AI-native product (with no "old UI" to embed into), forcing this framework will instead tie your hands — in that case you should consider conversational UI or the hyper-personalization paradigm below.

III. Agentive UX: AI works quietly in the background

What it is: A group of "invisible AI butlers," on call but unobtrusive. They auto-complete and optimize settings, but always stay behind the scenes, and you can take back control at any time.

Why (why "not seizing control" is its soul): The value of agentive UX lies in "efficiency improvement + noise reduction," but its easiest pitfall is "overstepping its bounds." Therefore every good case strictly adheres to one rule:AI proactively helps, but users can adopt/ignore/take over at any time. This is the principle of "assist but do not interfere." If the sense of intervention is not well calibrated, thoughtfulness becomes offensive (for example, the controversy caused by Samsung's "AI moon painting" over-enhancement).

Four subtypes, progressing by "AI proactivity level":

3.1 You complete me! (Intelligent completion)

AI predicts your next step and completes it; you can adopt it with one click, or ignore it.Gmail Smart Reply Gmail Smart Compose (after typing "Dear Sir," predicts the next sentence, press Tab to adopt),Google Sheets Smart Fill(identifies data patterns to auto-complete cells). Core principle: a "half-step ahead" rhythm, minimizing input burden without interrupting the workflow.

3.2 You surprise me! (Proactive prediction)

Provides thoughtful services beyond your expectations, and is always revocable. Siri Suggestions(automatically recommends the location on the map when receiving a message with an address),Arc auto-renames downloaded files(reads email titles to rename "flight confirmations" into readable names),Pierre code review(automatically summarizes code changes into a one-line description).

3.3 You exist? (Invisible learning)

AI is like an invisible person,observing habits in the background and optimizing silently, and you only realize it's there when "the search bar auto-complete is really accurate" or "homepage recommendations become thoughtful."

Netflix three-layer perception: ① Explicit feedback (likes and favorites) → ② Implicit context (device/time slot/region/whether binge-watching) → ③ Deep personalization (for the same show, serve couple posters to romance lovers, thriller posters to suspense lovers). Intelligent fine-tuning of "same content, different packaging."
mymind: Interface-free intelligence: when saving webpages/images, automatically extracts keywords, categorizes and tags them (color orange / item table lamp / designer / atmosphere warm); users "often forget AI exists, but every time they retrieve content they feel it has already done the work for them."
Easter eggs: Not for efficiency, but to create "Wow!" positive emotional feedback — Apple video call gestures trigger fireworks and balloons, YouTube subtitles highlight the button when they recognize "like and subscribe," Samsung moon photography enhancement.

3.4 You assist me! (Returnable control)

AI handles repetitive tasks in the background; the core is seamless take-back / handoff of control.

Ford assisted driving: Lane keeping automatically fine-tunes steering (handoff to the system), but pressing the brake/turning the steering wheel immediately switches back to human control (take-back); haptic vibration + voice reminders provide feedback. This is the most straightforward physical metaphor for "assist but do not dominate."
Waze: Proactively recommends better routes but retains your commonly used routes, marks estimated time saved, and you are free to choose "adopt" or "ignore."

When to use / counterexamples: Suitable for scenarios where "users do things repeatedly and AI can predict" (writing emails, navigation, organizing, recommendations).Counterexample: For irreversible/high-risk operations (transferring money, deleting data, submitting sensitive information), agentive UX must not make decisions on its own — it must escalate to the "Human in the loop" paradigm below, pausing for human confirmation (NNGroup calls this a decision gate / human-in-the-loop, which is the safety baseline for AI Agents).

IV. Human in the loop: co-creation side by side

What it is: AI is an assistant, not a boss;it does not fully replace you, but gives suggestions when you need them. It suggests text revisions, color schemes, and detail supplements, like a partner who understands your rhythm, adding the finishing touch at critical moments.

Why ("keeping humans in the loop" is necessary in creative scenarios): Value judgments in creation and review must be made by humans — AI does not know "which is better" or "which sentence to keep." Therefore the two iron rules of this paradigm are Human in the loop(humans always control the process, AI only acts within specified scope) + User curation(AI gives suggestions, humans decide whether to adopt them). It is an "editor who polishes but does not steal the manuscript," rather than an "overbearing AI that rewrites without permission."

Four subtypes:

4.1 Highlight text content

Highlight = input, question = command, return = revised draft. Select text and ask questions directly to it, no need to restate context (highlighting is the context anchor), greatly reducing cognitive load.

Lex.page: Highlight + input @lex, request polishing/style conversion in place ("make this more casual"), no prompt experience required.
ChatGPT: Select a sentence and directly say "suggest more quirky and funny ones," AI continues writing from that sentence without opening a new conversation thread.

4.2 Highlighting visual content

Bring "point to where to change" onto the screen — lasso/click visual elements to initiate local modifications; AI executes but humans retain dominance.

Vercel v0: Click any component (title/button) → enter "Make the heading larger and darker" in the floating layer → AI makes changes + version history (v1/v2) allows rollback. It also automatically adds social proof ("Learn More (10,000+ users)").
Photoshop Generative Fill: Lasso an area → enter "Cloud" → AI generates 3+ variants for selection, blended with the original image's light and shadow, and operation history allows rollback.

4.3 Work with me

AI does not take over work or ramble, only gives light prompts when needed ("Need help?" "Got a better version"), and focuses on enhancing rather than overwriting the user's existing input.

Granola: Instead of making a "unified summary," it provides personalized enhancement around your own quick notes — "you write what you want, I fill in what you need."
Dovetail: In research tools, AI provides "suggested highlighting" (Suggest), but Suggestions must be manually Accepted/Dismissed to take effect, the system does not judge value on its own.
Captions: Video editing provides three style templates (Impact/Cinematic/Paper), marking at the frame level which segments have been processed by AI (to prevent "black boxes" and enable rollback), while subtitle copy/paragraph order can still be modified.

4.4 Real-time dreaming: Think while doing (real-time generation)

"As I think, AI moves along with me." Instead of writing a prompt, clicking generate, and waiting for a long time, you get feedback as soon as you drag something, and a new image appears as soon as you change a word —The core selling point is "no loading state, this is real-time".

Meta AI: As you type, you see the image grow automatically, hiding all waiting/confirmation actions; after generation, you can still Edit/Animate for secondary creation.
Krea AI:Image = prompt + dynamic operations, dragging geometric shapes on the left changes the image form on the right in real time, supporting dual drive (multimodal) of structural sketches + text. This makes users feel "I am creating" rather than "I am waiting for AI to give an answer".
Decohere: Supports prompt append evolution (version tree), selecting a variant as the starting point for the next step, "frame-by-frame growth"; lightweight loading ("Slight loading, good feedback") — a conscious trade-off different from the "zero loading" of the previous two.

Design insights (for PMs building creative tools): Real-time feedback shifts user perception from "making a request" to "being in control", which is the moat of creative AI experiences. But "zero loading" is not absolute — Decohere intentionally leaves a little light loading to give users a sense of certainty that "the system is responding".Trade-off point: Immersion vs. progress certainty, depending on whether your users are more afraid of "stuck anxiety" or "being interrupted".

When to use / Counterexamples: Scenarios requiring human judgment such as creation, writing, design, and research.Counterexample: For pure factual queries (booking a restaurant, checking the weather), forcing people to confirm each step step by step is instead torture — that should be handed over to agentive or conversational paradigms for quick results.

V. Hyper personalization: AI becomes the system foundation

What it is: AI upgrades from "a feature in an App" to running through the entire system or on independent hardware. When you say "I want to relax", it mobilizes entertainment, food, and environment, like a friend who knows you best.

Why (this is the most radical paradigm with the heaviest responsibility): It pushes AI from a "tool" to a "proactive service provider". The benefit is extreme convenience (completing tasks across Apps with one sentence), but the cost is omnidirectional perception = omnidirectional privacy risks. It has the deepest sense of intervention, and the consequences of errors and abuse are also the greatest. This is a "future-style" paradigm, most of which are still in the concept/early hardware stage.

Two subtypes:

5.1 aiOS AI operating system (conversational operating system)

No need to click buttons to find menus,just tell the computer what you want to do. Core features: proactive information service (the system delivers information to you), understanding your context (remembering habits/location), omnidirectional perception (listening + viewing content + behavioral habits are all inputs), and software cooperating with each other (you don't care which App to use, the system automatically selects tools).

Analogy: Just like the "flashlight" changed from a standalone app on the App Store to a system feature in iOS Control Center — commonly used capabilities will gradually settle into the system foundation, and aiOS takes this path with AI. Designer Alexander Obenauer's vision: processing to-dos and emails in the same interface, shifting the focus of "software" to "users' actual needs".

5.2 AI as a product: Independently existing products (AI hardware)

AI is no longer a feature in a phone, but independent hardware that can converse, perceive the environment (location/sound/vision), and automatically generate context.

Rabbit R1: Multimodal physical device (button-activated voice + camera recognition + touchscreen), driving tasks through language and behavior (checking menus/placing orders), outputting key information in cards, using a bouncing rabbit animation + voice for anthropomorphic loading feedback, and even attaching a transparent hard box to strengthen the "product feel".
Humane AI Pin: Screenless + wearable + multimodal. Magnetically attached to the chest,projecting information onto the palm as a screen, interacting via voice + gestures (wave/slide/pinch). But it exposes a key design challenge:There is almost no visual feedback when AI is processing (even "awkward silence"), and users cannot tell whether it is working— this is the perceptibility problem of "interface-less interaction".

Feature summary: Extremely simple operation (just speak), usable for all ages, Everything in = Everything out (everything you say, do, and everything around you is a clue), and Apps are interconnected (cross-App with one sentence).

When to use / Counterexamples: This is a strategic reference for "next-generation OS / AI hardware" teams; ordinary functional PMs will not need it in the short term.Counterexample / Warning: The commercial failure of Humane Pin reminds us — even with a full sense of intervention,if perceptibility (whether users know what AI is doing) and clear value are not in place, no matter how flashy the hardware is, it cannot stand. Privacy + perceptibility are the lifelines of this paradigm.

What it is: In the past, you had to write precise prompts to make AI understand; now you interact in the most natural ways such as drawing, taking photos, seeing, writing, listening, and moving. Take a photo and say "copy this style", write a draft and say "polish this", speak "summarize the meeting", circle and select "simplify this section".

Why (this is the ultimate direction of AI experience): One sentence sets the tone — "In the future, AI will understand you, not you adapt to it."Multimodality reduces the "threshold for commanding AI" to near zero: you express in ways you have known since childhood (speaking, taking photos, handwriting, drawing), and AI is responsible for understanding on its own. This is a fundamental dissolution of the "prompt engineering" threshold.

Six input channels:

Subtypes (EN / Chinese)	Input method	Cases	What it solves
Can you just snap? Photography	Take a photo/screenshot, and AI understands	Photomath(automatically solving math problems + steps after taking a photo), iPad popping up a calculator when copying numbers	No need to manually input complex content (formulas/problems)
Can you just write? Handwriting	Natural handwriting, AI responds in the same style	Goodnotes(handwriting recognition + spell check + completion),Calculator for iPad(handwritten arithmetic problems, writing answers in your own handwriting)	Retaining handwriting flexibility + word processing precision
Can you just look? Sight	AI reads you through expressions/gaze	JAWS(green dot guiding visually impaired users to align with the camera),Hunger Station(eye-tracking scoring to predict "what you most want to eat"),New Computer(looking at screen → keyboard mode, looking away → voice mode)	Accessibility, emotion recognition, attention analysis
Can you just listen? Listening	Operate just by speaking	Alexa / Siri(wake word + continuous natural conversation),Arc(holding up the phone like making a call triggers voice search)	Faster than typing, zero threshold
Can you just draw? Drawing	Draw a sketch, AI fills in the details	Apple Intelligence(hand-drawn building outline → generating realistic images, displaying recognition keyword concept maps),Adobe Firefly(structure reference, sketch + intensity slider to control reference weight)	"You draw the framework, AI fills in the details", for creative drafts
Be my extension? Become an extension	Understands with a single movement (gestures/actions)	Arc(pinching a webpage triggers a summary, bringing it close to the ear triggers voice),Google Lens(circling/doodling an image area to initiate visual search)	Without speaking or clicking buttons, AI knows the next step with just "a movement"

Design commonalities: ① Inputs are natural actions (snap/write/look/listen/draw/move), not prompts; ② Loading states either fill the environment ("Summarizing for you…" + light music to ease waiting anxiety) or are directly hidden; ③ Most retain User feedback (thumbs up/down) for closed-loop optimization. This section and Shape of AI AI Product UX Pattern Library (Seven categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines)can be cross-referenced for specific input components.

When to use / Counterexamples: When input content is hard to describe in text (math formulas, visual styles, image regions), multimodality crushes the keyboard.Counterexample: For scenarios requiring extremely high precision and auditable records (legal documents, financial number entry), recognition errors of pure voice/handwriting are risks — either equip with secondary confirmation, or still use structured input.

VII. Trending UX trends in the AI era: Where we are now

What it is: This is a "meta-perspective" — it does not talk about individual patterns, but lays out the history of interface evolution and macro directions to help PMs make strategic judgments. One original quote sets the tone: "History rhymes, everything is repeating, but where exactly are we now?"

7.1 Five stages of interface evolution (where we are)

Stage (EN / Chinese)	Features	Markers
CLI Command Line Interface	Type commands to interact directly with the computer, "the code era"	Early technology stage
GUI Graphical User Interface	Package commands into clickable interfaces, usable by ordinary people	1984 Mac
GUI 2.0	Add cloud connectivity	Modern desktop/web applications
Conversational UI	LLM-driven, the first time "natural language" can be used to access AI computing power	ChatGPT
Intent based(under exploration)	Users no longer perform explicit operations,the system proactively understands intent and responds— AI changes from a "tool" to a "proactive service provider"	Future direction

Strategic implications for PMs: We are standing at a critical turning point from GUI → AI-driven interaction. "Intent-driven" is the endgame direction (corresponding to aiOS in Section V on hyper-personalization in this article), but it is still in the exploration stage —don't treat concepts as ready-made solutions, but know which way the wind is blowing.

7.2 Six macro trends (panoramic index of this article)

These six trends are exactly the first six paradigms in this article plus conversational UI, forming the top-level skeleton of the design resource section:

Conversational UI— the main way to access AI with natural language (§1 of this article).
Hyper personalization— the system proactively provides personalized services based on behavioral preferences (§V).
AI in old UI— AI integrates into existing interfaces, upgrading old systems (§II).
Human in the loop— humans intervene and supervise the AI process to ensure transparency and safety (§IV).
Multi-modal AI–UX— integrating text/voice/image/video (§VI).
Agentive UX— AI proactively executes tasks instead of waiting for your commands (§III).

When to use: When making product strategy and explaining to the team/boss "where we should go in the AI era", use this evolution axis + six trends as a framework. Its value is helping you locate "which segment of the evolution curve your product is on, and which paradigm to migrate to next".

Implementation checklist for AI product PMs

Condense the seven paradigms into executable judgments:

Define the form before drawing the interface: Before starting, use the "seven paradigms quick reference table" to ask yourself which column you fall into. Adding AI to an old product → AI in old UI; wanting AI to proactively save trouble → Agentive; doing creation → Human in the loop. Most products are in a superposition state.
Always leave an opening for undo/ignore/takeover: No matter which paradigm, "AI = helps you do things, not does things for you". Keeping control in the user's hands is the safety bottom line.
The deeper the intervention, the heavier the responsibility: Silent background < proactive prediction < independent hardware; each level deeper brings higher risks of privacy, perceptibility, and trust collapse (the lesson of Humane Pin).
Irreversible/high-risk operations = mandatory human in the loop: For transfers, deleting data, submitting sensitive information, agentive paradigms must be upgraded to a decision gate, pausing for human confirmation.
Curing "blank box fear" is a common motif: Preset buttons, natural language, photo taking and handwriting, highlight circling — all solve the problem of "ordinary people not knowing how to command AI". This is the core battlefield for lowering the threshold.
Citations must be truly verifiable: When doing "citations to build trust", links must be real and clickable; fake "hallucinated sources" are more dangerous than none.
Loading states are the experience watershed: Real-time generation pursues "zero-loading" immersion, but you must weigh "users' fear of getting stuck vs. fear of being interrupted" — sometimes deliberately leaving a slight light loading state actually provides a sense of certainty.

Derived from will-design-notes / AI-UX interactions (30 articles, 7 major paradigms). This article is at the "paradigm classification skeleton" level: for more fine-grained reusable AI design components, see Shape of AI AI Product UX Pattern Library (seven major categories: Wayfinders/Inputs/Tuners/Identifiers/Governors/Trust/Dark patterns + design guidelines) pattern library; for in-depth empirical research on conversational interfaces (6 types of conversation / no fluff / explainable AI / sycophancy / anthropomorphism / prompt UX / GenUI), see NNGroup AI Product and Conversational Interface Design (20 articles · 6 conversation types / 4 degrees of anthropomorphism / sycophancy / intelligence prioritized over human-likeness / explainable AI citations and disclaimers / no-fluff straight-to-answer truncation pyramid / 3 types of prompt suggestions · CARE framework · 4 uses of prompt controls / GenUI vs Vibe Coding · return of buttons and checkboxes · Promptframe · designing interfaces with prompts / 4 roles of AI assistants · creative partner Procter & Gamble 776-person experiment · initial review of ChatGPT Agent · Google AI Mode · why users switch between GenAI and search). The three pieces complement each other — this article helps you "choose the form", the pattern library helps you "pick components", and NNGroup helps you "dive into the deep end of a specific interaction type". Original screenshots are retained in sources.