Notes
Research, experiments, and reflections on AI and design, published along with the process of making them.
Design practice and AI (30)
How designers use AI and design systems in practice.
-
A Checklist for User Interviews: What to Prepare, What to Ask, and How to Report to the Client
2026-11-03
This note divides user interviewing with the people who use or buy something into nine stages, from planning through recruitment, consent, incentives, the interview guide and pilot, recording on the day, data storage, analysis, and reporting to the client, and turns each stage into a checklist of what to prepare and what to confirm. It draws on 28 academic works, 27 guidance documents from public bodies, guidance on laws, and industry codes of research ethics, 26 materials from research firms and consultancies, and 18 first-hand accounts by practitioners, and checks them against the full texts wherever these could be obtained. Gaps in preparation can ruin an interview study both before and after the questions are asked. In a study that recruited through social media, 271 responses arrived within 7 hours, and 94.5% of them were classified as fraudulent (Pozzar et al. 2020). When the researchers checked by telephone and email, only 8 of the 91 people who had completed the data were usable (Jean Louis and Thompson 2025). Screeners should mix foil options into the answer lists, ask about occupation with an open question, and plan for backup participants. In unmoderated studies with the most common incentive amount (60 dollars per hour), the no-show rate was nearly 30% (User Interviews 2025). Giving the incentive at the start of the session separates the incentive from how participants answer (GOV.UK). Consent should leave a record, and recording and observation should be announced both at recruitment and at the start of the session. Recording is the default, but notes written by experienced interviewers directly after the interview were comparable in detail to verbatim transcripts, and in some circumstances not recording is the best approach (Rutakumwa et al. 2020). Automated transcription produced sentences that did not exist in the audio in roughly 1% of transcriptions (Koenecke et al. 2024). Reports should distinguish facts, opinions, and interpretations, and should present recommendations together with how they were derived (MRS, JMRA). Data that can identify participants should not be passed to the client without the participants' consent (ICC/ESOMAR, JMRA). When a screen operation check is added, prototype responses, dummy data and test accounts, tasks that do not give away the answer, an explanation of thinking aloud, consent for screen recording and blurring of on-screen information, remote technology checks, and severity ratings must also be prepared. The average agreement between two evaluators on the problems found ranges only from 5% to 65% (Hertzum and Jacobsen 2001).
-
A Checklist for Group Interviews: What to Prepare, What to Ask, and How to Report to the Client
2026-11-03
This note divides group interviews (focus groups, round-table discussions) with people who use or buy something into stages, from planning through recruitment, confidentiality among participants, incentives, moderation, venue and observers, records, and analysis and reporting, and lists what to prepare and confirm in addition to what one-to-one interviews require. A final section covers what to add when a step for checking how participants operate screens is included. The evidence comes from 18 academic works, 16 documents from public bodies, standards bodies, and industry codes, 24 materials from research firms and trade journals, and 9 first-hand accounts by practitioners, checked against the full texts wherever these could be obtained. The data a group produces arise from interaction among participants. In a comparison that randomized 350 people, one-to-one interviews produced a wider range of items per person, but several kinds of sensitive disclosure were more likely in groups (Guest et al. 2017). In an analysis of 40 groups, more than 80% of themes appeared within two to three groups (Guest et al. 2017). Two groups per stratum gave a more comprehensive understanding of issues, and more groups added little (Hennink et al. 2019). Yet among 220 health studies, only 17% attempted to explain the number of groups (Carlsen and Glenton 2011). Participants should not know each other (JMRA), and people who attended a qualitative session within the past six months should be excluded (Government of Canada standards). Participants should be told, in the consent form and at the start, that researchers cannot guarantee that other participants will keep what they hear confidential (UC Berkeley). Ideally no more than half of the groups are observed, with one observer per group (MRS), and even non-interactive observers produced a physiological stress response and lowered some performance measures (Sonderegger and Sauer 2009). Reports should not include percentages (Government of Canada standards). Because a group discussion cannot show whether a screen is easy to use, operation checks should be run separately one to one or in an adapted form, with preparation for prototypes, dummy data, tasks, thinking aloud, screen recording and on-screen personal information, remote technology checks, and severity ratings.
-
Research Trends in HCI since Generative AI Became Usable (2023–2026): Many Systems Built, and the Checking Work Left to People
2026-11-01
This note reviews what HCI (human-computer interaction) research has asked about generative AI in the three years since the public release of ChatGPT in November 2022. It draws on 43 peer-reviewed papers and major preprints, mostly from CHI, together with counts of CHI papers. In OpenAlex, the share of CHI papers whose title or abstract contains LLM or generative AI terms rose from 1.9% in 2023 to 13.2% in 2024 and 23.0% in 2026. A systematic review of 153 CHI papers on LLMs from 2020 to 2024 (Pang et al. 2025) found that 62.74% used LLMs as components of systems, 90.85% acknowledged limitations related to research validity, and 84.98% used closed GPT-family models. By volume, the main current is research that builds systems with generative AI as a component. What HCI has newly learned, however, is written mostly in studies that measure what happens on the side of the people who use these systems: the difficulty of generative AI as a design material (failed prompting by non-experts, the gulf of envisioning), overreliance and appropriate trust, delegation to and oversight of agents, individual quality and collective homogenization, harm to learning, representational harm and the burden of verification for people with disabilities, sycophancy and the psychosocial effects of extended use, and the validity of LLMs as substitutes for research participants. Across these clusters, studies report that human work shifts from making toward verifying, integrating, and overseeing (Lee et al. 2025; Sarkar and Drosos 2025), and that performance and preference during use diverge from learning after use, collective diversity, and what benefits people (Bastani et al. 2025; Doshi and Hauser 2024; Cheng et al. 2026). Designs that narrow this divergence tend to be less preferred by users (Buçinca et al. 2021; Cheng et al. 2026). Other studies identify conditions under which harm does not appear (Kazemitabaar et al. 2023; the learning safeguards in Bastani et al. 2025; first-person expressions of uncertainty in Kim et al. 2024).
-
Conducting User Interviews: What Academia and Industry Share and Where They Diverge
2026-11-01
This note examines user interviewing for making design outcomes, from UIs to services and businesses. It draws on 32 academic works, 14 guidance documents from public bodies and codes of research ethics, 12 materials from research firms and consultancies, and 13 first-hand accounts by practitioners, and checks them against the full texts wherever these could be obtained. The conventions of asking are nearly the same on both sides: ask about specific past events instead of opinions or predictions of the future, do not lead, listen in silence, and read what people say together with observation. Academia has measured the limits of self-report and how strongly the interviewer shifts the answers (in Dell et al. 2012, participants were about 2.5 times more likely to choose the option they believed the interviewer had made, and 5 times more likely with a foreign interviewer who needed a translator). Industry has turned the same points into rules in its guidance and industry codes. The two sides diverge after the interviews. Academia decides when to stop by whether knowledge has become sufficient, using saturation or information power (saturation at 9 to 17 interviews, Hennink and Kaiser 2022), whereas industry repeats rounds of 4 to 8 people and stops when the team is confident in its design decisions. For analysis, academia keeps verbatim records and systematic procedures, whereas industry groups findings on sticky notes while memories are fresh and allots one hour of analysis to every two hours of research. A study that compared anthropology and design students at the same site found that the techniques were the same and that the differences lay in the purpose of the inquiry and the treatment of participants' voices (Segelström 2013). In industry, interviewing has spread beyond researchers (in a practitioner survey, 61% named product designers as people who conduct research). Ethics is checked by institutional review in academia and by organizational guidance and industry codes in industry. At the business layer, academia has tested the customer discovery method itself with a randomized experiment (Camuffo et al. 2020). On generative AI, both sides agree that synthetic users should not replace people. For AI as the interviewer, quality comparable to that of human interviewers has been reported for structured interviewing, and many assessments hold that exploratory interviews require a person.
-
Preparing a Review Paper on the Symbiosis of AI and Design: The Lineage of Symbiosis, Co-Creation Frameworks, Design-Specific Evidence, Philosophical Foundations, and a Map of Existing Reviews
2026-10-24
To prepare a review paper on the theme of the symbiosis of AI and design, we added 78 sources to the reading list on currents in the philosophy of design (49 sources) and organized them into six lineages: the lineage of the term symbiosis (Licklider 1960, Engelbart 1962, Negroponte 1970, Pask 1969), co-creation frameworks in HCI (from Horvitz 1999 to Rezwana and Maher 2023 and Weisz et al. 2024), design-specific evidence (from Dove et al. 2017 to Doshi and Hauser 2024), the philosophical foundations of symbiosis and their critiques (the extended mind, distributed cognition, mediation theory, the automation-augmentation paradox, and Sarkar's 2023 critique of 'collaboration'), more-than-human design, and 11 existing reviews. The existing reviews arrange the literature by type of tool, stage of design, and allocation of agency. Of the 13 reviews whose reference lists we examined through OpenAlex and Crossref, none cites Negroponte, Winograd and Flores, Verbeek, Ihde, Latour, Hutchins, or Clark and Chalmers, and only 2 cite Licklider. The provisional axis for rearrangement is which question in the philosophy of design each study inherits (who makes the judgment, what the designer can know about the material, how artifacts mediate action, and whose world is made). The full-text retrieval record (81 URLs) consists of 36 full texts, 38 abstracts or publisher descriptions, and 7 not retrieved.
-
How Many Times Have AI Agents Been in Fashion over 70 Years, and What Did Design Do Each Time? (Academic Research)
2026-10-24
This note traces how the idea of the AI agent emerged, became fashionable, receded, and returned between the 1950s and 2026, drawing on 47 works of academic literature in AI and HCI (human-computer interaction), and it examines whether design played a role in these rises and falls. There were two sources. Turing and McCarthy envisioned machines that think for themselves and improve when told, and Licklider regarded the long period before machines surpass humans as a time in which people and computers think together in close partnership. In the 1990s, the agent became a word of the interface. Maes explained learning interface agents through the metaphor of a personal assistant and placed their problems in two areas, competence and trust. Erickson at Apple separated adaptive functionality from the metaphor that portrays it as a living being and called for research on portrayal. In the same period, the strongest criticism also came from the design side. Norman wrote that the difficulties were social rather than technical, Lanier argued that people make themselves dumb in order to make agents look smart, and Shneiderman, in his debate with Maes, placed responsibility at the center and stated that anthropomorphic interfaces had failed. In the Lumière research that became the basis of the Office 97 assistant, a study in a usability laboratory had found that poor advice was costly for users, but the product placed a character in front and did not use the mechanism that decided whether to offer help on the basis of likelihood and utility. While the word receded from the interface, the tools remained on the side of design and human factors: levels of automation, calibration of trust, principles of mixed initiative, and 18 guidelines for human-AI interaction. The chatbot fashion of 2016 receded in 2018, and UX designers reported that machine learning was a difficult design material. LLM agents after 2022 returned from the machine learning side; their surveys cite definitions from the 1990s, and on web and OS tasks agents reached only 14.41% and 12.24%, against human success rates of 78.24% and 72.36%. Design research treated the level of autonomy as a design decision separate from capability, built interfaces in which people and agents share planning and execution, and showed that when 31 people tried advertised uses, they were obstructed by mismatches with their mental models. Design did not play the role of producing capability, but in each period it supplied the vocabulary for deciding how capability is presented and how far it is entrusted. Whether that vocabulary was followed in products has not been confirmed, except in the single case of the Office assistant.
-
What Should Creators Be Able to Take Over When Generative AI Does the Making? (Lessons from Aviation Automation Research)
2026-10-11
This note rereads, for creative work, a Japanese article that used aviation regulations to ask which skills software engineers should keep when AI agents write their code (Shimitaro 2026). What aviation regulations and research have protected is a state in which pilots use the autopilot but can take back the aircraft when it disengages. The FAA treats manual flight as the foundation of other flying skills and leaves the decision to fly manually to the pilot in command rather than to fixed quotas. In the research, the more successful an automated system is, the more investment in operator training it may require (Bainbridge 1983); people can sustain monitoring of a rarely changing source for only about half an hour; and automation bias cannot be prevented by training or instructions (Parasuraman and Manzey 2010). Cognitive skills decline more readily than manual skills: in simulator flights by 16 airline captains and first officers, instrument scanning and aircraft control were largely retained, whereas tracking position without a map display and recognizing instrument failures showed more frequent problems (Casner et al. 2014). Studies of generative AI report a shift from production to evaluation (Simkute et al.), a self-reported association between higher confidence that AI can perform a task and less critical thinking (Lee et al., 319 knowledge workers), and a 19% lower likelihood of correct solutions on a task outside the AI frontier (Dell'Acqua et al., 758 participants). In creative work, AI errors appear as plausible mediocrity without any alarm. Ideating with AI images strengthened fixation on an initial example (60 participants), and using AI ideas made individual works better but more similar to one another (Doshi and Hauser; Anderson et al., 36 participants). Based on this evidence, the note defines the capacity to take over in creative work as reframing the question, discerning the quality of options, and redirecting the work, rather than correcting errors in AI output. It then lists practices: generating the first ideas without AI, writing evaluation criteria before seeing AI output, returning to manual work at the core of a piece rather than following fixed quotas, trying arrangements in which AI serves as the monitor, bringing cases in which AI surprised the team back into critique and debriefs, and keeping hands-on practice. No study was found that measured skill decline among creative practitioners, and the listed practices are analogies from other fields whose effects have not been tested.
-
Does Prototyping Widen Design Exploration? Fixation on Early Examples, Parallel Prototyping, Rapid Prototyping with Generative AI, and Documentation in Research through Design
2026-10-10
This note examines whether making prototypes widens or narrows design exploration, drawing on 49 academic sources on prototyping, design fixation, prototyping with generative AI, tools for exploring design spaces, and process documentation in Research through Design (RtD). The benefits of making and testing depended less on making itself than on how prototypes were made. Under tight time constraints, participants who could test and revise outperformed those who could not (Dow et al. 2009). Creating several prototypes before receiving critique produced better results and more diverse designs than receiving critique after each prototype (Dow et al. 2010), and pairs who shared three designs incorporated more of their partner's ideas than pairs who shared one (Dow et al. 2011). When users saw only one paper prototype, they rated it higher and criticized it less than when they saw the same design among three (Tohidi et al. 2006). On the side of harm, designers who saw an example repeated its features even when told to avoid them (Jansson and Smith 1991). The effect of physical prototyping differed across studies, and one explanation holds that the more time and effort a model costs, the lower the novelty and variety of later ideas (sunk cost; Viswanathan and Linsey 2013). In a 60-participant experiment with Midjourney, the AI group repeated the most features of the example, fluency, variety, and originality tended to decline, and fixation shifted from the example to the AI-generated images (Wadinambiarachchi et al. 2024). Ideas from an LLM made individual stories more novel but made the stories more similar to each other (Doshi and Hauser 2024, 293 writers), and ChatGPT users produced ideas closer to those of other users (Anderson et al. 2024); several studies thus report individual gains alongside a narrowing at the collective level. In artwork data from more than 50,000 artists, adopting text-to-image tools raised productivity by 25% while average content novelty declined (Zhou and Lee 2024). Tools that expose the dimensions of a design space, turn generated output back into low-fidelity material, or lay out alternatives side by side were reported to support exploration in small user studies, whereas generating highly diverse ideas alone did not raise their value (Cai et al. 2023). RtD has treated documentation of the design process as a basis for research (Frayling 1993; Zimmerman et al. 2007), but no peer-reviewed study was found that directly examines how prototyping with generative AI changes the intermediate representations that such documentation records. As an untested interpretation, the note proposes that generative AI may weaken the conditions for sunk-cost fixation by reducing the effort of making, while strengthening the conditions for example-based fixation by presenting many finished-looking examples.
-
Settling a Business App's UI on One Design: What Narrows the Options, How to Compare Them, and Who Decides
2026-09-25
This note sorts out the material and steps for settling a business app's UI on one design out of countless options, drawing on public sources such as ISO 9241-210 and GOV.UK, research by NN/g and MeasuringU, primary sources from SAP, IBM, Salesforce, Atlassian, Basecamp, and GV, and testimony from Japanese business SaaS teams (SmartHR, Money Forward, RAKUS, LayerX, and others) and Sociomedia. Most options disappear before anyone draws them. The space is narrowed by which users are targeted and which are left out, how success is measured (for Carbon, task time and abandonment rate), devices and input methods (Fiori's cozy and compact), screen types chosen by purpose (floorplans) and business objects (OOUI), the time budget (Shape Up's appetite), and accessibility requirements. In the remaining space, three or more rough options are built in parallel. In NN/g's 1996 case, parallel design cost 73% more and the improvement was 70% (versus 18% for conventional iteration), and merging the good points of the losing designs scored higher than simply picking the best one. The material for comparison is small tests on representative tasks, and the problems found by novices and by experts overlapped only by about half. Detecting a 5-point difference from a benchmark with SUS takes 58 to 80 participants, so small tests suit grounds for dropping an option but rarely serve as grounds for choosing the final one. The final one is chosen by a person designated before the comparison, such as the Approver in DACI, the Decider in a Sprint, or Shape Up's betting table, relying on the product's principles. Decisions are recorded and become components, patterns, and review items that reduce the options for the next screen. As points where business apps tend to cause hesitation, the note covers information density, expert efficiency, the split between buyers and users, migration from existing work, and exceptions to public-facing principles on screens for staff (GOV.UK).
-
Asking an LLM for Design: Briefs, Skills, and a Process That Keep UIs from Converging on the Obvious
2026-09-25
This note sorts out how to ask an LLM for design rather than code, drawing on a Japanese company's case, 16 vendor primary sources, 7 research-firm sources, and 26 academic papers. In the starting case, Canary turned its design philosophy into a skill. The LLM then matched the product's tone, yet every screen converged on a generic UI. Vendors explain the pull as a fallback to the 'safe choices' that dominate training data, and academic work shows that typicality bias in post-training preference data sharpens the output distribution (Zhang et al. 2025). Anthropic's frontend-design skill was rewritten from a version that banned defaults such as Inter into one built on a brief grounded in the subject, calibration against known default looks, and a two-pass plan reviewed before building. The vendor itself notes that banning common fonts only makes generations gather on the next one. For divergence, experiments suggest that procedures (CoT) and diverse sampling cues work better than role labels alone (Meincke et al. 2024; Deng et al. 2026). Canary's 'designer personas' are called personas, but each is a procedure that fixes where to start, how to return, and when a proposal fails. Checking can be automated with screenshots and explicit criteria, but LLM critique of UIs still falls short of human evaluators (Duan et al. 2024), and people who see AI proposals tend to fixate on the first example (Wadinambiarachchi et al. 2024). The resulting pattern runs: brief, calibration of defaults, divergence placed before the design system, convergence through the design system, visual checks, and selection by a person.
-
How Much of the Evidence for Adjudicating the Six Design System Conflicts Is in Place?
2026-09-11
Decomposing the six adjudication criteria into units of measurement yields 37 pieces: 25 have some form of measurement and 12 have no research that could be confirmed. Most of the measurement, however, comes from proxy indicators in adjacent fields, and only two studies take design systems directly as their object, a comparison of reuse (n=8) and an evaluation of the artifacts of 4 DSs against 46 criteria. What is missing in common is an operational definition that splits a continuous quantity in two, a longitudinal time axis, attribution to the DS rather than to the side using it, blinding, and effect sizes buried under the coverage ceiling of automated audits.
-
Design System Best Practices: What the Evidence from Research and Practice Supports
2026-09-11
Weighs 23 academic and 46 industry sources on design systems against each other. 83% of organizations have a dedicated team (147 respondents), but only 5% measure ROI. Role-based token design, governing names, staged migration, automated verification in CI, and a defined contribution model each have first-hand testimony from multiple organizations. What blocks adoption is not technology but organizational factors, and research has identified four barriers. The controlled experiment on ROI is limited to n=8. WCAG 2.2 AA and DTCG 2025.10 set the preconditions for implementation from outside.
-
End-User Development (EUC/EUD): Genealogy and Limits — A Map of 152 Sources
2026-08-15
An integrative summary mapping the history, social demands, and limits of end-user computing (EUC) and end-user development (EUD) across 152 sources, from Licklider in 1960 to the LLM era of 2026. It shows how EUC, born as a framework of managerial control, and EUD, born as a framework of emancipation, have described the same phenomenon in opposite directions; how four distinct social demands were folded into the single word 'democratization'; how the limits take a different shape in accounting, medicine, and scientific computing; and how the limits the field documented for itself are being replayed in the same form in the LLM era.
-
What Has Generative AI Changed About Making for Oneself? A Map of 59 Sources
2026-08-14
An integrated summary mapping, through 59 sources, how generative AI (2023–2026) has changed making that starts from one's own need and serves oneself. It fills the one intersection absent from the two sister corpora (337 works on EUD history and learning outcomes): making-for-oneself × generative AI. Evidence that the entry barrier to making has dropped is accumulating across four areas (personal software, self-data, self-made assistive technology, hobbyist creation), but 32 of the 59 sources are unreviewed preprints and measurement is skewed toward immediate evaluation. Of the four gaps identified by the prior reviews, attachment to self-made artifacts has begun to move via IKEA-effect RCTs on AI co-created text, while delayed-test longitudinal tracking, tracking of private personal tools, and motive comparison remain unfilled in the generative AI era.
-
Design Systems as AI's Foundation Layer — Structured Design Knowledge Determines Agent Accuracy
2026-07-09
Analyzes the structural transformation whereby design systems (component libraries, design tokens, structured specifications) function as foundational infrastructure for AI agents, drawing on 27 academic papers and first-party information from 5 vendors. Provision of formal specifications reduces agent navigation by 33-44% and achieves 100% accuracy (Jin 2026); leveraging Figma JSON metadata outperforms pixel-only conversion (Gui 2026, ICLR); 5 companies simultaneously implement design systems as AI context layers. The meaning of investing in design systems has shifted from 'maintaining consistency' to 'determining the accuracy ceiling of AI agents.'
-
EU AI Act Article 50 and Design Practice — A Structural Analysis on the Eve of Enforcement
2026-07-09
Analyzes the impact of EU AI Act Article 50 (transparency obligations), effective August 2, 2026, on design practice through a corpus of 30 academic publications. Three structural problems are identified: (1) the scope of disclosure obligations is ambiguous, leaving the boundary between design editing and regulated manipulation undefined; (2) watermark and labeling implementation rates fall far short of regulatory requirements (38% / 18%); (3) greater disclosure granularity paradoxically erodes trust rather than improving user judgment. Designers confront a 'disclosure dilemma' — disclose AI use and lose client trust, or conceal it and incur legal risk — while gaps in copyright protection undermine the basis for billing.
-
90% Adoption x 10% Approval — The Paradox of AI Tool Diffusion and Evaluative Divergence
2026-07-09
Analyzes the structural divergence in creative industries where AI tool usage rates reach 90% while only 10% view the impact on their industry positively, drawing on 27 academic sources and 15 industry surveys. Three structural mechanisms are identified: (1) mandatory adoption generates symbolic adoption — using without endorsing (Heidenreich & Talke 2020); (2) AI triggers professional identity threat and psychological reactance (Mirbabaie et al. 2022; Jussupow et al. 2022); (3) technostress negatively moderates continuance intention across all pathways (Lu & Hu 2025, n=443 designers). A parallel trend appears among developers (Stack Overflow: trust declining from 40% to 29%, favorability from 72% to 60%). 'Using it but not endorsing it' is not individual irrationality but a structural collision between organizational pressure for efficiency and professional identity.
-
The AI Expectation-Capability Gap and UX — Industry Trends (2026)
2026-06-28
Organizes industry efforts to close the gap between AI's actual capabilities and user expectations through UX, drawing on 33 sources across three streams: vendor guidelines (Google PAIR / Microsoft HAX / Apple HIG / IBM / OpenAI / Anthropic / Meta), public surveys (Pew / Stanford HAI) and consulting surveys (McKinsey / BCG / Deloitte / Gartner), and practitioner patterns.
-
The AI Expectation-Capability Gap and UX — Academic Review (2026)
2026-06-28
A review of 32 academic sources on techniques for resolving through UX the divergence between what AI can actually do and what users expect (the expectation-capability gap). Organized into five areas: trust calibration, mental models, explainability, cognitive forcing functions, and expectation violation theory.
-
AX and Design: A Cross-Comparison of Academic and Industry Perspectives
2026-06-27
A cross-sectional comparison of the intersection between AX (Agentic Experience) and design, drawing on 109 sources: 36 academic and 73 industry. Three points of convergence between industry and academia are identified (agency distribution as the central problem, transparency alone being insufficient, and context setting the ceiling for accuracy), along with three points that industry omits but academia addresses (junior designers' learning pathways, loss of exploratory opportunities through efficiency optimization, and structural limitations of oversight UIs). While the academic standing of the term 'AX' itself remains unsettled, the underlying problem space is advancing rapidly under the rubric of human-agent interaction research.
-
Industry Discourse on Agentic Experience (AX): What Vendors, Consultants, and Practitioners Are Saying
2026-06-27
Drawing on 73 industry sources, this note examines how the concept of AX has been received, defined, and implemented across industry from 2025 to 2026. Four structural patterns emerge. (1) Every vendor proclaims the shift 'from UX to AX,' yet definitions diverge (Maeda = decision surface, Nielsen = Intent Layer, Salesforce = bidirectional design, Apple/Google = declarative primitives). (2) Design principles converge on 'transparency, control, and intent-first,' but the granularity of intervention differs. (3) The gap between hype and reality is quantitatively documented (Gartner: 40% cancellation forecast; Forrester: true autonomy remains rare; McKinsey: only 6% are high performers). (4) Developer testimony reveals structural constraints (93% approval fatigue, multi-agent 15x cost, context quality as the decisive factor).
-
The Academic Foundations of Agentic Experience (AX): Where Human-Agent Interaction Research Stands
2026-06-27
A review probing the academic underpinnings of 'From UX to AX' (Maeda 2026). As of 2026, peer-reviewed papers using the term AX itself barely exist, but adjacent research on human-agent interaction / AI delegation / agentic AI UX is surging, centered on CHI, CSCW, and DIS. Three structures emerge from 36 sources: (1) the design of agency distribution is the central problem (who controls what, and when); (2) the trade-off between trust and oversight (intermediate checking is optimal but does not guarantee full accuracy); (3) the gap between industry promotion and actual UX (demonstrated by an SR of 102 commercial agents plus testing with 31 users). Includes 7 review papers and 12 empirical studies from CHI/CSCW/DIS.
-
MCP and Design Systems: The Infrastructure Layer for Agent Integration
2026-06-27
MCP (Model Context Protocol) has rapidly gained adoption as a common connectivity layer across design tools from 2025 to 2026. With the Figma MCP server (beta June 2025), Figma Config 2026 Connectors, and Claude Design's design system import capability, the agentic loop between design and code is closing. Two structural findings: (1) design system maturity determines the upper bound of agent output accuracy (officially recognized by Figma); (2) the best agents operate with the fewest tools (Vercel experiment: reducing from 15 to 1 tool raised success rate from 80% to 100% and improved speed by 3.5x). MCP is simultaneously a protocol for inter-tool communication and a mechanism that transforms design systems into an 'agent context layer.'
-
Vibe Coding and UI Agents: Natural-Language Production and Its Boundaries
2026-06-27
A single phrase — 'vibe coding' — coined by Karpathy in February 2025 became an industry term within a year. UI-generation agents such as v0, Bolt, Lovable, and Replit Agent form its infrastructure, yet the distinctions drawn by Willison ('vibe engineering') and Osmani ('AI-assisted engineering') mark the fault line of production quality. While prototyping has been compressed to hours, design judgment, testing, and review remain as irreducibly human processes, shifting the bottleneck from 'making' to 'deciding.' In design production, the ability to evaluate the quality of generated UI is a prerequisite for effective agent use.
-
Design Agent Tools in 2026: The Current State of Autonomous Production
2026-06-27
In the first half of 2026, Figma, Adobe, Canva, and Anthropic simultaneously introduced agentic capabilities for design. All four companies espouse the principle that 'agents execute while humans judge,' yet they differ in the granularity of judgment and the timing of human intervention. Figma enables divergence and selection on the canvas; Adobe automates workflows across 60+ tools followed by review; Canva delivers a finished artifact from a single prompt for fine-tuning; Claude Design provides bidirectional code-design synchronization. The shared structural constraint is that the maturity of a team's design system sets the upper bound on agent output quality.
-
Ten Generative AI Industry Topics Ranked by Impact on Design Professions and Careers (June 2026)
2026-06-19
An analytical note ranking ten generative AI industry topics as of June 2026 by their degree of structural transformation of design professions and careers (roles, pricing, employment, organizational structure). The four evaluation axes are: (1) displacement of core tasks, (2) pricing power, billing, and name recognition, (3) employment, junior pipeline, and organizational change, and (4) structural certainty of impact. Top-ranked topics include generative UI / labor commoditization / vibe coding / agentic workflows / junior pipeline erosion; lower-ranked include image and video generation / regulation / tool market / enterprise adoption. Rankings reflect structural career change rather than tool novelty. Pricing and name-recognition dynamics connect to democratization-recommodification-paradox / design-pricing-vs-ai-commoditization / the gaps in livelihood viability and initial entry; role discussions connect to designer-role-ai-roundtable / ai-design-near-term-flashpoints. Most industry figures originate from vendor or listicle sources and require primary verification; unconfirmed claims are explicitly marked [unverified].
-
AI Flashpoints Certain to Arrive in the Next 1-3 Years — Three Design Leads' Near-Term Flashpoints (2026-2029)
2026-06-09
A debate log (hub) in which the leads of a design firm, a strategy consultancy, and an independent studio (business-only) independently enumerated the AI flashpoints 'almost certain to ignite within 1-3 years' (R1), cross-rebutted one another (R2), conceded the danger of deliverable-based pricing under a client counterargument and retracted the R2 claims (R3), and developed survival prescriptions for offices carrying large numbers of good-enough designers (R4). Cross-cutting findings are extracted into three atomic notes. Figures carry caveats as position talk and correlation-not-causation.
-
AI in the Design Industry: An Academic Review (2026)
2026-06-07
An integrative summary of 37 academic publications on AI/generative AI/LLM adoption in design (UX/UI, product, graphic, education), collected through a lightweight scoping review and organized into seven thematic clusters: co-creation, UI generation, evaluation automation, creativity effects, role transformation, and education.
-
AI Adoption in the Design Industry: Perspectives from Policy, Data, and Product Developers (2026)
2026-06-07
An integrated summary of AI adoption in the design and creative industries, compiled from three source tiers: official bodies and primary data (labor statistics, adoption rates, copyright and standards), the factual portions of think tank and consulting reports (with positional advocacy stripped), and first-person testimony from product developers (Figma, Adobe, Canva, Vercel).
-
Recent AI Applications in the Design Context (2026)
2026-06-07
A survey report on recent AI applications in the design domain as of 2026 (generative UI, agentic AI, Figma, synthetic users, accessibility)