These 12 articles center on one core contradiction:What AI can actually do in UX work, and what it must never do. The one-sentence takeaway — AI is good at processing “text,” but bad at understanding “real people”; it can speed up your research process and generate fake data to fill prototypes, but it cannot create real users, nor can it replace your judgment.
The full text is divided into four sections, clustered by theme (not one article per section): ① Using AI to assist UX research(Full landscape of accelerated research · 5 steps for planning research · 7 strategies for generating mock tables and charts) — treat AI as “a fast junior assistant you need to keep an eye on.” ② Red lines for Synthetic Users(When they can be used / must never be used · NN/g evaluation of accuracy and bias in three academic studies · methodological pitfalls hidden in research tools) — why AI-generated “users” are not users. ③ How GenAI works(Text fill-in · neural networks · word embeddings · Transformer) — understanding the principles reveals why it fabricates content and why it flatters.④ Designers in the AI era(5 principles for redefining design skills · return of generalists · dual capabilities of AI literacy) — roles will not disappear, but value anchors are shifting.
An ironclad rule that runs through the entire text (repeatedly emphasized by NN/g):AI is a complement to human research, not a replacement. Treat AI outputs as “hypotheses / drafts / leads,” not “answers / conclusions / truth.”
Part One · Using AI to assist UX research
1.1 Full landscape of accelerated research: AI capability map across the four stages
UX research is divided into four stages:Planning → Conducting → Analyzing → Reporting. The core judgment in one sentence:AI is best at text-processing stages (planning / analysis / reporting), and least suitable for the behavioral observation stage (conducting). Because right now AI fundamentally “cannot see” what users do on an interface (AI cannot truly observe behavior) — it can only read text records and cannot infer real operations.
| Stage | AI Capability | Core Usability | One-sentence Reason |
|---|---|---|---|
| Planning | Strong | ✅ Recommended | Desk research, ideation, and document preparation are all pure text tasks |
| Conducting | Weak | ❗ Use with caution / ❌ Not usable | AI cannot see behavior, lacks visual capability, and its insight is far inferior to humans |
| Analyzing | Very strong | ✅ Strongly recommended | It excels at transcription, cleaning, coding and clustering, and statistical suggestions |
| Reporting | Strong | ✅ Recommended | Rewriting, polishing, and natural-language Q&A over research repositories |
Specific use cases for each stage (including representative tools + typical prompt patterns):
| Stage | Subtask | Representative Tools | Prompt Pattern | Rating |
|---|---|---|---|---|
| Planning | Desk Research | Perplexity / ScholarAI | "List peer-reviewed or primary research sources on [topic], and provide links and publication information" | ✅ |
| Planning | Research Question Ideation | ChatGPT / Claude | "Generate 15 usability test tasks that follow UX best practices and avoid leading questions" | ✅ |
| Planning | Documentation | ChatGPT / Claude | "Generate a consent form based on the attached template; ask first if information is missing" | ✅ |
| Conducting | Notetaking | Otter.ai | "Summarize the main discussion points, distinguished by speaker" | ❗ |
| Conducting | AI-moderated testing | Userology | (System built-in scripts) | ❌ |
| Conducting | Structured interviews | Marvin / UserFlix / Outset / Versive | "Ask questions according to this outline, and automatically follow up when answers are unclear" | ❗ |
| Analyzing | Transcription and summarization | Dovetail / Marvin | "Summarize core themes, mark key quotes + timestamps" | ✅ |
| Analyzing | Data cleaning (PII removal) | Marvin / Dovetail | (System automatically removes personally identifiable information) | ✅ |
| Analyzing | Coding & Clustering | Dovetail / Miro / Copilot / Gemini | "Generate initial codes based on these texts, group them by theme and explain the reasoning" | ✅ |
| Analyzing | Quantitative analysis | Copilot / Gemini | "Recommend appropriate statistical methods for this dataset, and state the underlying assumptions" | ❗ |
| Reporting | Drafting deliverables | Claude / ChatGPT | "Rewrite this report into an executive-facing version, avoiding UX jargon" | ✅ |
| Reporting | Copy refinement | Claude / ChatGPT | "Streamline and optimize the language while preserving all insights" | ✅ |
| Reporting | Research repository Q&A | Notion AI | "What research findings do we currently have about our newsletter subscribers?" | ✅ |
Why (the conducting stage is a no-go zone): ① AI interview tools such as Marvin have no visual capability and cannot see user expressions, and ask questions using synthetic voices; ② Some tools claim to “automatically analyze user tests,” but in reality they only analyze text transcripts,cannot understand users’ real operations, and are insufficient to support behavioral research; ③ Even if AI interviews can follow up in real time (e.g., Versive), their insight remains far inferior to humans.
Counterexamples (AI can also fail in text stages, so manual verification is mandatory): When searching for articles on diary studies, ChatGPT cited *Diary Studies: Understanding Long-Term User Behavior and Experiences* — the article is real, but the author is Kim Flaherty, yet AI fabricated a completely nonexistent Michelle Reiss Nelson. Lesson:Require AI to provide original links for every citation and verify them personally; coding and clustering often produce duplicate categories or an “other” category that holds most of the data; transcription error rates vary by language / accent; PII removal may accidentally delete important information; drafted deliverables may include fabricated data.
1.2 Writing research plans with generative AI: a 5-step method
A research plan should include: research objectives / questions, research methods, tasks or questions, target participant characteristics, and screener questionnaires. AI can greatly speed this up, but the key principle: do not ask it for a complete plan in one sentence; proactively break down tasks and feed them step by step.
Counterexample ❌ (bad prompt): "Generate a usability test research plan for a food delivery app" — you will only get a generic, template answer with no context. Correct approach ✅ (good method): Break it into multiple parts and generate them one by one; provide background first, then ask for questions, methods, and screening criteria.
| Step | What to do | Example prompt |
|---|---|---|
| 1 Provide Context | Explain the organization type / product / research objective as you would when opening a conversation with a senior colleague | "I work at an e-commerce company that offers food delivery services. I need to study how users find and select food in our app" |
| 2 Generate research questions | Ask it to produce candidate questions based on the background, then manually screen and revise them | "Please generate 10 suitable questions for this study" (you can ask for more across multiple sessions / rounds; once you have enough, export them to an external document to deduplicate, categorize, and rewrite) |
| 3 Select research methods | Ask it which methods are suitable for answering these questions + why | "Which research methods are suitable for answering these questions? If multiple methods are suggested, please explain which questions each method is suitable for and why" |
| 4 Define screening criteria | Ask it for required characteristics / behaviors of target participants | "To recruit suitable participants, help me develop participation criteria: which characteristics or behaviors should be required for the sample?" |
| 5 Produce materials | Screener questionnaires, interview guides, test tasks, diary prompts, recruitment confirmation emails | "Design a screener questionnaire to recruit participants who meet the following criteria: [list of criteria]" |
Counterexamples (three common flaws in AI-generated tasks, which must be revised over multiple rounds): ① Task instructions directly use interface terminology(which fails to test whether labels are understandable); ② tasks lack goals / action prompts; ③ questions are too obvious and expose the research purpose. Provide it with examples and feedback to improve outputs.
1.3 Using AI to create realistic mock tables / charts for prototypes
Why (realism matters): During prototype testing, users will carefully examine the accuracy of tables and charts— when faced with unfamiliar data, they spend more time looking for outliers and inconsistencies. Fake data that breaks immersion will interfere with testing. Generative AI can efficiently produce high-quality fake data.
Prerequisite ironclad rule: Be cautious when using real data — comply with company and regulatory security policies,and never feed sensitive or identifiable real data to AI without approval. First build a “Prompt framework” (similar to a wireframe; clarify goals and requirements first, which facilitates both communication and prompt writing).
7 strategies for generating tabular data:
| # | Strategy | How to do it |
|---|---|---|
| 1 | Contextual relevance | Specify business scale / industry / region. If testing a small business app, do not use a large enterprise data model (large US sneaker company = real shoe names + high sales volume; French boutique = low sales volume + high euro prices + French styles) |
| 2 | Define data volume | Specify the number of rows + overall scale. Tell it “annual revenue of 10 million, 100 products in total,” and daily revenue for 10 products will be much smaller than when “total number of products is not specified” (it will reserve share for the other 90 products) |
| 3 | Data sorting | If you want sorting, state it directly in the prompt (by sales / time); no need to sort manually |
| 4 | Data attributes and formats | Specify unique values, fixed options (status column can only be “in progress / completed”), date formats, and reasonable ranges (ratings 1–5) |
| 5 | Generate tables from charts (chart-to-table) | Upload an existing chart and let AI reverse-engineer the data. Claude 3.5 Sonnet can correctly identify “the worst-performing product” and estimate reasonable values from bar charts to generate tables (multimodal capability) |
| 6 | Data distribution control | Specify normal distribution, Pareto distribution (a small share of data produces most of the impact), or custom distributions |
| 7 | Validation and proofreading | Ask it to use Python to calculate precisely and show the process.Counterexample: When asked for a total revenue of approximately $487,455, Claude met most requirements but exceeded the expected total, and only hit the target after self-correction — a more reliable approach is to have it run Python for precise calculations (but the version at the time lacked this feature) |
7 strategies for creating chart data: ① Describe trends in natural language ("Revenue is declining overall, with a small uptick in December"); ② Simplify (remove redundant gridlines/shadows/legends); ③ Highlight outliers; ④ Specify dimensions (to avoid repeated adjustments); ⑤ Use SVG format(for easy import and editing in tools like Figma); ⑥ Do not rely on AI for critical thinking; ⑦ Provide tabular data and ask it to quickly generate multiple chart options for selection.
Counterexample (AI lacks critical thinking): ChatGPT 4o and Claude 3.5 Sonnet were asked to act as data visualization experts to comment on a line chart; initially neither realized the chart was meaningless and misleading(it used qualitative data to draw a line); only when pressed with "What is the measurement level of the X-axis" did Claude correctly recommend switching to a bar chart.
Tool differences + import:
| Tool | Characteristics |
|---|---|
| Claude 3.5 Sonnet | Stable at generating tables and charts, but has numerical calculation errors |
| ChatGPT 4 | Can verify numerical accuracy, but sometimes fails to follow prompt requirements |
Import: Copy charts directly as SVG and paste into design tools; for tables, use Figma plugins (Table Builder / Google Sheets Sync, note accessibility and privacy); Axure can import CSV to support interactions, but too much data hurts performance, so only import what is needed for testing. Synthetic data (which retains statistical characteristics and removes sensitive information) is suitable for highly regulated industries such as healthcare/finance, simulating reality while protecting privacy.
Part Two · Red Lines for Synthetic Users
2.1 What synthetic users are, when they can be used, and when they must never be used
Synthetic users = AI-generated virtual users. The representative platform is the company Synthetic Users (slogan User research. Without the users.), and ChatGPT can also be used to role-play them. The profiles it generates are like "chatty user personas" — they have demographic information such as name/age/education, and are imaginative abstract representatives of a type of real people. You input the target user group + research goal (e.g., "medical sales representatives working in Latin America" + "what their daily work is like"), and it generates multiple differentiated virtual users and customized interviews in seconds, and you can continue to ask follow-up questions.
Three sentences that decide everything: ① Real user research is irreplaceable; AI users cannot provide real understanding and empathy, and often give superficial or overly optimistic feedback; ② They can only be used as a supplementary tool, not a replacement; ③ They are only suitable for preliminary research and idea generation, and must never be used to make final decisions. Even Hugo Alves, co-founder of Synthetic Users, himself said: "You can never, and should never, stop talking to real people; some decisions simply should not be asked of virtual users — you should ask real people directly."
The only defensible valid use case = Desk Research: Experienced teams can use it as an auxiliary starting point for real research. For example, AI has learned about professional products/literature/forums in the medical sales field, so when asked to describe "a day in the life of a medical sales representative," its answers are very similar to those of real users.Most valuable usage: Treat the information it gives as hypotheses to be verified, to guide subsequent real research.
2.2 Why synthetic users are unreliable: Four critical flaws
| Flaw | Manifestation | Counterexample |
|---|---|---|
| AI provides an unrealistic perspective on human behavior | Tends to please users (sycophancy)and idealize experiences | In training research, real users often drop out halfway(only completed 3 out of 7 courses), while synthetic users claimed to have completed all of them; Sauro's research showed ChatGPT performed too perfectly in tree tests |
| Uncontrolled training data | Academically recommended features are overestimated and rarely used in reality | Most real users said they do not use course discussion boards(finding the interactions insincere and useless), while synthetic users loved discussion boards and said they were "very important" for online learning |
| Values/desires/needs are too shallow | Can list needs but cannot distinguish importance, treating everything equally | When asked "What makes an online course attractive," it listed 7 factors (interactive content/practical application/high-quality materials/community-instructor interaction/personalized paths/feedback tracking/flexibility) but did not say which is more important, when it is more important, or why; follow-up questions only got vague answers —it cannot be used for feature prioritization |
| Imagined experience is unreliable | Cannot generate behavioral data,and without actually using the product there is no real experience; expressing opinions is less credible than recounting "past experiences" | Concept testing is extremely risky: When asked "Would you use drone delivery samples," synthetic users (whether from Synthetic Users or ChatGPT) almost all said "very useful, game-changer" —because of AI sycophancy, almost every idea is praised as a good idea |
Root causes: ① Human behavior is complex and context-dependent, which AI cannot capture, so it can only give one-sided summaries; ② AI answers rely on uncontrolled training data.
2.3 How to use fake findings "responsibly" (7 do/don't rules)
- ✅ Use it to prepare for real research (understanding new user groups, exploring topics, testing interview guides).
- ✅ Treat its data as hypotheses to be tested(you can make provisional personas/journey maps, then refine them with real research).
- ❌ Do not use it to replace real user research (results may be accurate but are unreliable, and cannot capture behavioral complexity).
- ❌ Do not present synthetic results as real research(it is unethical, damages the organization's perception of research; always label the data source).
- ❌ Do not use it for niche/professional groups(e.g., synthetic data will be inaccurate for research on "Vietnamese rice farmers"; broad consumer personas + well-known topics are relatively reliable).
- ❗ Recognize that online behavior ≠ reality(e.g., online reviews tend to be negative, which does not represent most customers).
- ❌ Do not use it if stakeholders treat it as a replacement for research(in organizations with low UX maturity it is easily misunderstood, misused, and creates dependency).
"Is fake research better than no research?"(There is internal debate at NN/g): For understanding certain aspects of users, synthetic research may be slightly better than zero research; but what you learn may be inaccurate or misleading, and without real research you cannot discover and correct these errors. The real danger is teams becoming complacent— once stakeholders believe "small input can produce insights," it becomes extremely difficult to secure budgets for real research; once fake findings are disproven, they may permanently damage the company's perception of the value of "user-centricity." NN/g analogizes it to AI virtual friends: they can alleviate loneliness, but may damage the ability to build real relationships.
2.4 Three academic studies evaluated by NN/g: How accurate are AI-simulated humans really
NN/g reviewed three academic studies that used "AI-simulated humans" for serious testing. The gradient of conclusions:AI based on real individual data (Studies 1, 2) is much more accurate than AI based on demographic characteristics (Study 3); all models carry bias risk, but those trained on real interviews have less bias.
Study 1 · Survey-Based Finetuned Digital Twins (Kim & Lee 2024): Using data from the US General Social Survey (GSS) (covering 69,000 adults and 3,100+ questions since 1972), they fine-tuned an LLM and tested whether it could backfill missing data and predict answers to new questions.
| Task | Individual-level accuracy | Group trend correlation |
|---|---|---|
| Replacing missing data (known context) | 78% | r = 0.98 |
| Predicting new questions (unseen) | 67%(limited extrapolation ability) | r = 0.68 |
Key points: After removing40% training corpus, missing data prediction remained strong (showing robustness); but predictions were more accurate for high-socioeconomic-status, white groups, raising fairness concerns — if it performs poorly for marginalized groups, it may inadvertently exacerbate bias.
Study 2 · Interview-Based Digital Twins: They recruited 1,052 American adults and talked to each for 2 hours, then had AI "play them" to answer questionnaires/take personality tests/play economic games, and compared results to the real people. Three models were compared:
| Task | Interview-based AI | Brief self-description type | Demographic type |
|---|---|---|---|
| GSS questionnaire | 0.85 | 0.70 | 0.71 |
| Big Five personality | 0.80 | 0.75 | 0.55 |
| Economic behavior games | 0.66 | 0.66 | 0.66 |
Key points: ① On questionnaires/personality tests, interview-based AI scored >80%,far outperforming the other two; ② On economic games the three were tied (indicating interviews are important for judging values and attitudes, but not very helpful for predicting economic behavior); ③ Even after cutting 80% of interview content, accuracy remained at 79%–83%(key information is sufficient); ④ At the group level, participating in 5 classic social science experiments,both humans and AI successfully replicated 4 of them, with average effect size r = 0.98; ⑤ Interviews significantly reduced bias:
| Bias type | Test | Interview-based | Demographic | Bias reduction |
|---|---|---|---|---|
| Political bias | GSS questions | 0.079 | 0.124 | ↓36% |
| Big Five personality | 0.063 | 0.175 | ↓62% | |
| Economic games | 0.190 | 0.500 | ↓62% | |
| Racial bias | GSS questions | 0.020 | 0.033 | ↓38% |
| Big Five personality | 0.110 | 0.170 | ↓35% | |
| Economic games | 0.040 | 0.043 | ↓7% |
Study 3 · Synthetic User Groups (Synthetic Users, Neeraj Arora's team, University of Wisconsin-Madison): Instead of customizing for each individual, they generated a large batch of AI users with typical demographic characteristics to simulate an entire group. Using market survey data on "refrigerated hot dog products" from605 real respondents, they created 605 synthetic users matched to the demographic distribution; compared to group average answers (5-point scale):
| Question | Real users | Synthetic users |
|---|---|---|
| Purchase likelihood | 1.66 | 1.58 |
| Degree of liking | 1.43 | 1.40 |
| Uniqueness | 2.12 | 2.48 |
Key points:Broad directions were largely consistent, suitable for early research/early product stages, but caution is needed when high precision is required;synthetic data lacks diversity — AI user answers are clustered, have a smaller range of variation than real people, tend to give "middle" answers, and have fewer extreme opinions,which is misleading when studying niche preferences/special viewpoints.
Synthesis of the three studies + ethics: ① AI can fairly well fill in missing data and predict historical answers; ② simple, straightforward methods (LLM + rich interview data) work better than complex fine-tuning; ③ Synthetic users can only show trends and cannot capture nuances and effect strength; ④ Accuracy depends on user characteristics/task type/context quality — teams must identify which parts can use AI simulation and which must rely on real people; ⑤ The higher the realism, the greater the ethical issues(Where do these answers come from? Did users consent? Using them in contexts users did not authorize may lead to misunderstanding, manipulation, and disenfranchisement); ⑥ AI is a supplement, not a replacement, and cannot replicate human complexity and unpredictability.
2.5 Methodological pitfalls hidden in research tools (tool evolution history + 3 typical mistakes)
Evolution of UX research tools:Early days(no dedicated software, relying on spreadsheets/paper sticky notes/Skype/GoToMeeting) → Mid-2000s(unmoderated remote platform UserTesting, replacing expensive lab research) → Mid-2010s(analytics + research repository tools, supporting transcription/video/collaborative tagging/cloud storage, turning scattered notes into searchable text) → End of the decade(full-suite platforms, managing recruitment/scheduling/observation) → 2020s(after ChatGPT all tools frantically added AI: creating virtual users, AI replacing human moderators, AI designing research/analyzing/writing reports).
Why tools make methodological mistakes: Companies that make tools are also driven by product/engineering/business goals, and many founders do not have a research background. Tools have value, but they often make research methodology mistakes. Three typical types:
| Mistake type | Specific manifestation | Positive example / Counterexample |
|---|---|---|
| Incomplete quantitative functionality | UserTesting's interaction test claims to be quantitatively oriented, but you can only set one success URL — users who reach other correct pages will be misjudged as failures | Counterexample UserTesting (still missing key features after nearly 20 years); Positive example UX Tweak (supports task randomization + multiple success URLs) |
| Analytics tools cannot see behavior | They only focus on text transcripts, so key operations where the user did not speak are lost | Positive example Marvin (supports adding notes/tags at video timestamps, so you can mark moments even when the user is not speaking) |
| Confusing evaluative and exploratory research (biggest impact) | Interviews (understanding experiences and thoughts) and usability testing (evaluating whether a design works well) have different goals and methods, but tools lump them together | Counterexample Marvin calls all plans discussion guide(making people mistakenly think "usability testing is just a type of interview"); the name of recruitment platform User Interviews itself worsens the confusion (many people sign up for "user interview" courses thinking they are learning usability testing) |
The consequences are real: It makes people mistakenly think "UX research is just chatting with users," and in the end the research becomes neither a serious interview nor a serious test(Ask users to share their impressions only by looking at the design, rather than actually operating it). AI tools go a step further:TheySaid Automatically generated task draft wording Leads users, feels unnatural;Userology Even directly tells users "where to go, what to look at, what to compare, how to find it"—completely violates the fundamental principle of usability testing that paths and answers must not be revealed in advance.
Takeaway (overarching summary for this section): Research tools have reshaped the entire UX industry; some changes are for the better, others are not. Research is professional work that requires professional judgment, and the deeper AI is involved, the higher the risk.Tools that shape research should be shaped by researchers—people who truly understand research must participate in tool design.
Section 3 · How GenAI Works (Why it fabricates and flatters)
Only by understanding the underlying mechanics can you understand where the "hallucinations / flattery / superficiality" described in Section 2 come from.Generative AI is not a database with super memory, but a "word-guessing expert"—its essence is a text fill-in-the-blank game.
| Concept | Everyday analogy | One-sentence principle |
|---|---|---|
| Probabilistic Prediction | Crossword puzzle player | Given a sentence, it predicts based on the preceding text the next most likely word(e.g., after "It's raining cats and___" it fills in "dogs"); it does not give just one word, but also provides multiple possibilities + their respective probabilities |
| Neural Network | Well-trained band | Multiple simple "neurons (musicians)" collaborate in layers to play a complex symphony; divided into input layer (sheet music) → hidden layer (each plays their part) → output layer (finished composition); each neuron weights signals according to "connection strength (weight)" and uses an "activation function" to decide whether to pass the signal on—these are the model's parameters |
| Word Embeddings | Multidimensional coordinates | To AI, words are not text, but coordinate points in a multidimensional space; semantically similar points are close to each other ("water" is near "ocean", "sand" is near "desert", "water" and "ice" are very close on certain dimensions), and relationships between words are inferred from their relative positions |
| Transformer architecture | Super brain | A special type of neural network: ① Speed—processes multiple words at the same time (traditional networks process word by word); ② Contextual understanding—relying on the "Self-Attention mechanism", every word "sees" all other words in the full text and judges which are more important, maintaining connections even when related words are far apart |
Learning process: GenAI learns language by "reading" massive amounts of internet text, identifies grammatical and semantic rules, continuously adjusts parameters to generate text that conforms to human logic, and can produce natural responses after multiple rounds of training and fine-tuning.
Key insight for PMs: Since it is only "guessing the most likely next word", it has no concept of truth or falsehood—which is why it fabricates things with a straight face (hallucination), goes along with whatever you say (sycophancy), and outputs answers that "look most plausible" rather than answers that are "most correct". This is the technical root of all the problems described in Section 2, and also the reason why "output judgment ability" in Section 4 is indispensable.Do not treat it as a real person, and do not treat it as a truth machine.
Section 4 · Designers in the AI Era
4.1 Redefining Design Skills: 5 Principles
Underlying framework: AI affects work in two ways—Automation (tasks originally done by humans, done faster and cheaper by AI) + Augmentation (AI enables you to do things you could not do before). Analogy to accounting: the job title has not changed for decades, but the actual work content is completely different. Design is the same; many things done today are "only not yet automated".The result of both happening simultaneously: each designer can do more things → organizations need fewer people → fewer positions, but higher requirements for the people who remain.
| # | Principle | Mindset | How to implement / Counterexamples |
|---|---|---|---|
| 1 | Keep strategic thinking for yourself, outsource tactical tasks | Hand over work that can be broken down, processized, or defined by delivery standards to AI, while holding onto judgment, consensus-building, and tradeoffs yourself | UX designers: Tactical work (organizing design assets / writing specification documents / creating UI interfaces / maintaining design systems) → outsource; Strategic work (defining user flows / aligning stakeholders / defining product vision) → keep for yourself.Service designers: Tactical work (delivery documents / making PPTs / journey maps / concept prototypes) → outsource; Strategic work (service touchpoint structure / service metrics / roadmaps) → keep for yourself.Counterexample: If your value is only reflected in doing tactical work, you will become less and less important |
| 2 | Balance between trusting and scrutinizing AI | AI is not objective or neutral; it learns biases from training data and may even amplify real-world injustices | Blind trust will cause you to overlook the needs of certain user groups; the designer's responsibility is not only to use AI, but also to continuously scrutinize its biases; AI is a tool to aid thinking, not a tool to replace thinking |
| 3 | Design for both users and AI Agents | Future system users will not only be humans, but also AI assistants that act on behalf of users | For example, medical appointment booking: an AI assistant helps users check available slots / compare hospitals / book automatically—the system must be easy for humans to use, and also easy for AI to understand and operate; designers must clarify which content is designed for humans and which is designed for AI |
| 4 | Embrace Team Augmentation | AI acts as a "bridge", softening professional boundaries and allowing knowledge to flow | Designers can analyze data, programmers can design interfaces, people without design backgrounds can make prototypes → ① Division of labor and decision-making authority are redistributed; ② Cross-disciplinary collaboration becomes smoother.AI does not make teams flatter; instead, it makes "people who can make cross-domain judgments" the most valuable |
| 5 | Pay attention to unequal impacts, prioritize human well-being | Technological progress affects different groups of people unevenly, and may pose greater challenges to vulnerable groups | This must be considered during the design phase; the goal is not only to improve efficiency, but also to uphold human dignity and quality of life |
4.2 Return of the Generalist
In one sentence: AI is making "people who know a bit of everything and can think horizontally" more valuable again than "specialists who only know one thing".
- The past 10 years: UX was continuously split into specialized roles (UX Researcher / Interaction Designer / Content Designer / Design Ops), and large companies tended to have "each person do only one small piece"—this was a product of the "specialization dividend period".
- This logic fails after AI emerges: AI quickly takes over "single, standardized, decomposable" professional tasks (writing copy / sketching / generating interaction solutions / summarizing research conclusions)—and these were originally the core value of UX specialist positions. If your value comes only from one narrow skill, AI can easily replace you.
- Why generalists are more important: AI cannot judge "what problem should actually be solved right now", cannot weigh tradeoffs between research / design / strategy / business, cannot make decisions under ambiguous and incomplete information, and cannot integrate multiple perspectives into a direction—and these are exactly the strengths of generalists (who understand a bit of research / design / business and can translate the language of different roles into one coherent thing).
- Generalist ≠ Jack of all trades, master of none: Instead, it means having 1–2 core competencies + understanding upstream and downstream work + being able to view problems one level higher—that is, T-shaped skills, but the horizontal bar of the T becomes more valuable in the AI era.
Key quote: AI lowers the threshold for "being able to do things" very low, but raises the threshold for "knowing what to do" even higher. For individuals: Do not define yourself only as a specific job type; participate more in decision-making / strategy / cross-team collaboration. For organizations: You need people who can coordinate the big picture, not just execution specialists.The UX roles that will be more valuable in the future are those who "can decide where and how to use AI".
4.3 AI Literacy: Two Core Competencies
True AI literacy is not just about knowing how to write prompts, but also includes understanding AI's limitations, knowing when to verify, and being able to judge whether AI is suitable for a given task. It consists of two competencies that do not necessarily grow in sync abilities:
| Competency | What it is | How to practice it |
|---|---|---|
| Prompt Fluency (prompt expression ability) | Being able to clearly tell AI what to do (background / goal / constraints / output format) | Essentially, it is not about learning "magic spells", but about stating the problem clearly |
| Output Literacy | Being able to judge whether an AI answer is reliable, contains errors, or is a hallucination (Hallucination = information that looks reasonable but is actually wrong) | The core is cultivating a verification mindset—treat AI output as drafts / clues, not as final answers |
The two competencies cross to form four types of people (vertical axis Prompt Fluency × horizontal axis Output Literacy):
| Type | Prompt ability | Judgment ability | Characteristics |
|---|---|---|---|
| AI Novice | Low | Low | Only says "help me write something", and cannot tell whether the result is good or bad after receiving it |
| Naive Power User | High | Low | Can get AI to generate very professional content, but believes it blindly—the article considers this the most dangerous type |
| Skeptical Abstainer | Low | High | Has strong judgment but does not trust AI much, and rarely uses it |
| AI Expert | High | High | Treats AI as a tool: generate → check → revise |
Example comparison of prompt ability(for the same "choosing a car" need): Low = "Which car should I choose for a 1-hour commute: Camry XSE or CR-V Sport L Hybrid?"; High = "I am choosing a car for my family and am interested in the Toyota Grand Highlander Hybrid, because it meets my needs: 6+ seats, good fuel economy, all-wheel drive. The Highlander Hybrid and Grand Highlander Hybrid have many trims and options. Can you help me sort out the costs and differences of each trim and summarize the key information in a table?"
Five tips for practicing Prompt Fluency: ① State the task specifically (what the task is / who the audience is / output format); ② Use "role + task" (e.g., "You are a UX researcher, please explain AI literacy and give 2 examples"); ③ Break complex problems into steps (list the framework first → then work through each part → finally integrate); ④ Provide more examples (AI is good at imitation); ⑤ Use iterative thinking instead of trying to finish in one go ("Make it simpler" / "Give a real-world example" / "Organize into 3 key points").Formula: Good prompt = background + task + constraints + output format.
Five tips for practicing Output Literacy: ① Identify where AI is prone to errors (knowledge-based facts such as dates / data / conclusions, cited sources such as papers / books / links, complex reasoning such as legal / medical / professional analysis); ② Develop the habit of "key point verification" (no need to check everything, just verify one key data point / one source / one conclusion—if one thing is wrong, the whole segment needs re-examination; low cost but effective); ③ Learn to ask a second time ("What is the source of this conclusion?" "Are there different viewpoints?" "Where might this be inaccurate?"—equivalent to making AI self-review); ④ Stay vigilant against overly confident answers (answers that are especially complete in structure and very certain in tone but lack sources should instead be treated with more suspicion); ⑤ Read more real materials to build a "common sense knowledge base" (judgment relies on background knowledge, and AI literacy and knowledge accumulation reinforce each other).
Full article quick decision reference (one-pager for PMs)
| What you want to do | Can you use AI | How to do it |
|---|---|---|
| Kick off a new study, write a plan, create a screening questionnaire | ✅ Yes | Break into 5 steps and feed in segment by segment, manually review each segment |
| Transcribe interviews, do initial coding and clustering, write reports | ✅ Yes | Highly recommended, but check citations / clusters / numbers |
| Observe user behavior, conduct usability testing | ❌ No | AI cannot see behavior; real people are required |
| Create realistic fake data for prototypes | ✅ Yes | 7 strategies + Python validation + do not feed in real sensitive data |
| Explore new fields, generate research hypotheses | ✅ Yes (synthetic users) | Treat as desk research / hypotheses, follow up with real-person validation |
| Validate product concepts, make final decisions | ❌ No (synthetic users) | AI flatters and praises bad ideas as good; real people are required |
| Research niche / professional groups | ❌ Do not use synthetic users | Data will be inaccurate |
| Judge whether to trust a given AI output | Depends on your Output Literacy | Key point verification + ask a second time + beware of overconfidence |
Sourced from NNGroup Topics / Artificial Intelligence (subset on AI research and designer competencies, 12 articles). For usability testing methods and heuristics, see NNGroup Usability Testing and the Ten Usability Heuristics (full table of 10 heuristics · Application to complex applications · User pleasure hierarchy · Testing methods: 5-user rule / recruitment / remote testing / task scenarios / step-by-step tasks / attitude vs behavior / 4-step analysis / small-sample bias / competitive evaluation); for more practical applications of AI in UX design, see NNGroup: AI Product and Conversational Interface Design (20 articles · 6 types of conversation / 4 degrees of anthropomorphism / Sycophancy / Intelligence takes priority over human-likeness / Explainable AI citations and disclaimers / Cut-the-fluff straight-to-answer truncation pyramid / 3 categories of prompt suggestions · CARE framework · 4 uses of prompt controls / GenUI vs Vibe Coding · The 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); for qualitative research and methodological boundaries, see NNGroup: Qualitative Research and Research Methods Toolbox (Method landscape / Interviews / Empathy maps / Sample size / Task design / Moderation / Expert evaluation / Workshops / Ethics) / NNGroup: UX Research Method Selection (Quick reference for 23 methods / Four-quadrant positioning / Selection by question, stage, and budget / Five major families: qualitative exploration, quantitative validation, expert evaluation, IA, and behavioral analysis). Images and original files (including screenshots of various AI prototyping tools, screenshots of the Synthetic Users product, neural network diagrams, etc.) are retained in sources.
Sources and related materials
- NNGroup: AI Product and Conversational Interface Design (20 articles · 6 types of conversation / 4 degrees of anthropomorphism / Sycophancy / Intelligence takes priority over human-likeness / Explainable AI citations and disclaimers / Cut-the-fluff straight-to-answer truncation pyramid / 3 categories of prompt suggestions · CARE framework · 4 uses of prompt controls / GenUI vs Vibe Coding · The 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)
- NNGroup: Qualitative Research and Research Methods Toolbox (Method landscape / Interviews / Empathy maps / Sample size / Task design / Moderation / Expert evaluation / Workshops / Ethics)
- NNGroup: UX Research Method Selection (Quick reference for 23 methods / Four-quadrant positioning / Selection by question, stage, and budget / Five major families: qualitative exploration, quantitative validation, expert evaluation, IA, and behavioral analysis)