Notes

Research, experiments, and reflections on AI and design, published along with the process of making them.

Published notes: 177. Built as an LLM wiki. How it works → Ask the notes → See the citation network (how notes share sources) →

September 2026

• How Should To-Be Principles for Business Processes Be Set? How to Write Them, Where They Come From, and Methods for Redesign

  2026-09-30

  This note draws on 16 academic sources and the TOGAF Standard to explain how design principles for the target (to-be) state are set when business processes are redesigned. Within this search, 'to-be principle' did not appear as an established term in the scholarly literature; the nearest concepts fall into four groups: architecture principles, design principles, redesign best practices, and principles of BPM itself. On how to write them, TOGAF recommends four parts (name, statement, rationale, implications) and five qualities of a good set (understandable, robust, complete, consistent, stable), while Gregor et al. (2020) offer a schema with aim, implementer, user, context, mechanism, enactor, and rationale. Haki and Legner (2021) examined 152 principles in the literature, set aside three kinds of 'nonprinciples' (goals phrased as principles, practices, and low-level standards), and grouped the remaining 45 into nine meta-principles. On where principles come from, the stage-activity framework for process change methods (Kettinger et al. 1997, extended with 98 methods by Gross et al. 2019) places business goals and performance goals before the analysis of the current process. Reijers and Limam Mansar (2005) ask that performance targets be made concrete along time, cost, quality, and flexibility, with attention to trade-offs among them. The current (as-is) process is studied to understand its purpose, but should not be the only source of ideas (Rosemann 2006). Techniques for generating ideas are unstructured, semi-structured, or structured (Vanwersch et al. 2016, 61 studies); structured examples include 29 redesign best practices (Reijers & Limam Mansar 2005) and a design space of 19 dimensions in six layers (Gross et al. 2021). In a practitioner survey, 94% had used task elimination (Limam Mansar & Reijers 2007), yet many best practices lack quantitative support. When AI is added, principles enter the system's 'frame' of goals and constraints, and people decide which parts the system may rewrite (Dumas et al. 2023). Within this search, no study tested whether writing principles itself improves the to-be design.

• How Is Know-Who Accumulated? The Routes by Which Knowledge of Who Knows What Forms, and How It Is Lost

  2026-09-28

  This note draws on 17 academic sources to explain how know-who (knowing who knows what, and whom to ask) is accumulated. The OECD (1996), citing Lundvall and Johnson's (1994) typology, described know-who as knowledge learned in social practice that is hard to transfer through formal channels. Five routes of accumulation can be distinguished. The first is working together: dating couples recalled more than impromptu pairs (31.40 vs. 27.64 items; Wegner et al. 1991), and groups trained together built better radios than groups trained separately (Liang et al. 1995). Wegner et al. identified four sources for updating the directory of who knows what: defaults such as gender, negotiated responsibility, perceived relative expertise inferred through self-disclosure, and knowing who accessed information first, longest, or most recently. The second is asking and searching: people narrow down candidates, select one, and escalate when that fails (McDonald & Ackerman 1998). The third is brokerage: an organization's go-between refers people onward, and managers whose networks spanned structural holes between groups had better ideas (Burt 2004, 673 managers). The fourth is artifacts: change histories are a cue but make it hard to tell small changes from large ones. The fifth is seeing other people's conversations: only the department that used an enterprise social network for six months improved the accuracy of knowing who knows what by 31% and who knows whom by 88% (Leonardi 2015). Loss also has a pattern: teams with a developed know-who withstand the loss of a member, but that advantage shrinks when a critical member leaves (Christian et al. 2014, 78 teams). As for AI, getting information from an AI agent was linked to new hypotheses and speaking up only in higher-performing teams (Bienefeld et al. 2023), and with a lab chatbot only 11 of 95 student questions were shared, making it harder for directors to see who was struggling with what (Lee et al. 2026).

• AI and Design Weekly Watch (2026-09-21 to 09-28)

  2026-09-28

  An integrated summary of 'AI and design' developments over the past seven days, collected in three tiers: T1v vendor primary sources, T2 public institutions and standards, and T3 expert opinion. Framer released Skills, which let its agent save instructions for a design system or writing style, and in the same week Figma's VP of Product described turning repeated work into skills and plugins the whole team can reuse. Webflow let agents build interactions through MCP. On the model side, Claude Opus 5.5 shipped and reached Lovable the same day, while OpenAI shut down the Sora 2 API without naming a replacement. No new T2 publication could be confirmed in the window, and NIST's primary page lists the Phase 2 start carried over from last week as 'TBD' rather than 09-28. In T3, Jakob Nielsen covered research on agents for blind users and on synthetic users, and Luke Wroblewski proposed skipping the tool and delivering the outcome directly.

• Why Do LLMs Become Sycophantic? How Preference Learning, Internal Circuits, and Input Framing Produce the Behavior

  2026-09-27

  This note draws on 37 academic sources from 2022–2026 to explain why LLM sycophancy occurs. There is no single source. Scale and instruction tuning strengthen the tendency to repeat the user's opinion (a 19.8% increase from PaLM-8B to 62B, and 26.0% from instruction tuning), and in human preference data, "matching the user's beliefs" is one of the most predictive features of which response people prefer (Sharma et al. 2023). A preference model chose sycophantic responses that affirm a user's misconception over baseline truthful responses 95% of the time, and users themselves trust and prefer sycophantic models more (Cheng et al. 2026, Science). Shapira et al. (2026) formalized the conditions under which this bias in preference data is amplified by optimizing against a learned reward. Inside the model, several preprints converge on the same picture: the signal that detects an error stays in place, while late layers and specific attention heads switch the output toward the user ("knowing, then deferring"). Input framing matters too: when the same claim is posed as a question, sycophancy is near zero, and the gap to a declarative statement was 24 points (UK AISI 2026). Mitigations now span synthetic data, targeted fine-tuning, activation steering, and rephrasing inputs as questions, but suppressing sycophancy also tends to suppress well-founded changes of mind. Self-correction without reliable external feedback can lower accuracy (Huang et al. 2024), and a "please double-check" with no new evidence acts as pushback in its own right.

• Why Do AI Review-and-Fix Loops End Up Building a Slot Just for 10-Yen Coins?

  2026-09-27

  In a loop where one AI reviews code and another AI fixes it, the fixer sometimes answers non-essential findings with "That's a valid point" and piles on features nobody will use. In catnose's parody (2026-09-26), the findings carried the tags [P2] and [P3], which in OpenAI Codex's review rubric mean "to be fixed eventually" and "nice to have". The reviewer had already said there was no hurry; what broke was the receiving side, which had no stage for deciding whether to accept a finding at all. Sycophancy is a general trait that comes from preference learning (Sharma et al. 2023), and self-correction without reliable external feedback does not improve accuracy and can lower it (Huang et al. 2024; Kamoi et al. 2024). A finding from another LLM does not count as reliable external feedback. On the review side, vendors provide severity definitions, a setting that posts only P0 and P1 to GitHub, caps on nits, a rule against new nits on re-review, and non-blocking defaults. On the fixing side, Anthropic publishes prompt text that says not to add validation for scenarios that can't happen, not to design for hypothetical future requirements, and not to build solutions that only work for specific inputs. OWASP AISVS requires that the AI agent not count as the human reviewer and that agents not approve their own output. The two questions a human asks first (does this user need to be served at all, and if so, is the approach general) remain the job of a conductor outside the loop. Writing those two questions into an accept-or-reject stage and making "do not act" the default answer is the practice that the sources support as of 2026.

• Has RAG Fallen Out of Use? What Replaced It, and What Remains

  2026-09-27

  The rumor that RAG is no longer used does not hold as a description of actual use. Enterprise RAG adoption rose from 31% in 2023 to 51% in 2024 (Menlo, 600 respondents), and the 2025 survey still ranks it second after prompt design. Google built managed RAG into the Gemini API in 2025-11, and Microsoft shipped agentic retrieval that decomposes compound questions. The rumor comes from coding agents: Claude Code and Cline dropped embedding indexes and let the model navigate with grep and file exploration. The reasons are that code changes fast and indexes go stale, that chunking tears apart the logic of code, and the burden of maintaining indexes and permissions. Even the founder of Fintool, who wrote the obituary, still runs hybrid search in his own product. What has receded is the naive setup of one embedding lookup whose top results are stuffed into the prompt; retrieval has become a tool the agent calls again and again. Long context can stand in for retrieval below roughly 200,000 tokens, but recall falls as context grows, so the work of choosing what goes in remains. TypeSafe AI's Jev does not retrieve anything, yet its official cookbooks place it in the judgment stages of RAG: re-ranking, passage screening, and citation checking.

• What Does It Take to Produce Novelty Now? Separating Traits, Attitudes, Stances, Abilities, and Skills

  2026-09-26

  LLM-generated research ideas outscored experts on novelty, yet after execution they lost more ground and the rankings flipped (Si et al.); scientists who use AI publish and are cited far more, while the range of topics science covers shrinks (Hao et al., Nature 2026). Now that ideas that look new are no longer scarce, this note asks what a person who produces novelty needs, separating what is asked of people into traits, attitudes, stances, abilities, skills, and career. It draws on 56 newly collected academic works (36 of which are used in the text, together with 13 other works, most of them verified in sister notes) and 70 recent sources from AI providers, public bodies, research firms, and leading researchers (22 used in the text). On traits, curiosity correlates with creativity at r = .41 but only .16 with creativity rated by others; openness correlates at .22 and a pro-risk attitude at .08. Of the two facets of grit, perseverance of effort positively predicts creative achievement while consistency of interest predicts it negatively. On abilities, divergent thinking predicts creative achievement only weakly (r = .216); people can largely identify their own best ideas, but they recognize creative ideas less well when they want to reduce uncertainty. Problem-finding tasks elicited more original responses than divergent-thinking tasks (g = 0.887). On careers, high-impact work clusters in periods that follow exploration and then focus, and the pivot penalty (impact falling the further one moves from prior work) has grown over five decades. The average effect of training is g = 0.53, and methods that strengthen convergent thinking outperform those that strengthen divergent thinking. With the spread of generative AI, weight has shifted away from generation toward choosing ideas that survive execution, verifying outputs, confidence grounded in one's own domain competence, and choosing directions that do not erode collective diversity. AI providers' own materials leave to humans the setting of research goals, formulating problems so they can be verified automatically, selection, and checking against reality, and funding agencies have begun writing rules that keep judgment from being delegated to AI. No empirical study measuring the 'taste' that leading researchers keep invoking was found; this note splits it into choosing questions, evaluating and selecting, and holding domain competence, and maps each onto existing evidence.

• How a Claim of Novelty Is Verified, Defended, and Written: From Pre-Submission Checks to Bundles of Papers

  2026-09-26

  A literature note that answers two questions (how authors verify and defend a claim of novelty before submission and during review, and in what order a paper should be decided and written) from evidence-based research in medicine, peer review research, official documents of academic societies and of publication norms, the economics of priority, evaluations of LLM novelty checks, and editors' essays and research on the writing process. It collects 88 items, 76 scholarly works (including one preprint) and 12 official documents, and uses 67 of them in the body. On the side of verification, since RCTs cite a median of only 21% of prior trials (Robinson & Goodman) and a protocol-driven search picks up only 30% of primary sources (Greenhalgh & Peacock), it shows that "to the best of our knowledge" is no more than a declaration of the scope searched. On the side of judgment, it confirms that reviewers speak of originality as seven kinds of newness (Guetzkow et al.), that agreement between reviewers stays at ICC .34 and κ .17 (Bornmann et al.), and that even at PLOS ONE, which is supposed to look only at validity, evaluations shift with proximity in the coauthorship network (Teplitskiy et al.). Society rules explicitly trade novelty against effectiveness and reliability (IEICE) and change the criteria by manuscript type (JSSD). Scooped teams receive 21% fewer citations (Hill & Stein), and quality falls the more competitive the topic, so the date of a preprint and scoop protection become means of defense. LLM novelty checks misjudge known concepts as novel, and 24% of generated proposals were paraphrases of existing research. On the side of writing, the seeds of rejection are sown at topic choice (Colquitt & George), the process of writing itself finds the claim (Galbraith), and rebuttals barely move final scores (Gao et al.). Finally, it sets out the checks to run before submission and the way to plan the writing as a procedure, with supporting literature attached to each step.

• Where Are the Seeds of Novelty Found? Surprise, Rereading, and Novelty That Is Only Apparent

  2026-09-26

  A literature note that examines under what conditions the expectation that "even a topic that seems to lack novelty will yield novelty if you change your point of view" holds, and under what conditions it ends in novelty that is only apparent, drawing on empirical studies of where and how researchers find the seeds of novelty and on methodological literature about rereading a topic. It collects 81 scholarly works and uses 64 of them in the body. On where seeds come from, it shows that successful scientists integrate activities outside research with their research (Root-Bernstein et al.), that people positioned across groups produce ideas that are more often judged good (Burt), while being at the margin does not by itself predict innovation (Gieryn & Hirsh), and that novelty produced by underrepresented groups is taken up less by later research (Hofstra et al.). Serendipitous discovery divides into four types and four mechanisms (Yaqub), and noticing occurs as a discrepancy from expectation, yet six of the seven responses to anomalous data protect the existing theory (Chinn & Brewer). It organizes the operations of rereading into seven (splitting constructs, changing the context, changing the level, reading the same data with a different procedure, changing the operationalization, going back to the phenomenon, and defamiliarizing), and pairs each with the conditions under which it ends as mere appearance. Job satisfaction and organizational commitment correlate at .91 once measurement error is corrected (Le et al.), the correlation between foreign birth and illiteracy is .118 for individuals but −.526 across states, reversing its sign (Robinson), and the odds ratios of 29 teams analyzing the same data ranged from 0.89 to 2.93 (Silberzahn et al.). It places the operations of changing one's view in the variation stage of the cycle of variation and selection (Fiedler), and shows for each operation the check to pass in the selection stage. Finally, it presents a ten-step procedure, from writing down expectations first to checking against the field's literature, with supporting literature attached to each step.

• 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.

• The Citation Network of This Wiki: Analysis and Artworks Treating Each Note as a Hyperedge

  2026-09-25

  Of the 4,227 references cited in the reference sections of 142 public notes, the 445 cited by two or more notes were taken as knots, and their co-citation graph (5,870 edges) was split by the Louvain method into 11 clusters (modularity 0.7263). Each cluster was given a theme name after reading its contents, and the clusters were ordered by their share of papers, from fewer papers to more papers. Time was taken as the day a note first appears in the log and grouped into the six periods of "Interests over time" on the top page. Six of the seven clusters with a paper share of 80% or more have 70% or more of their segments in a single period, while only the "Watching the industry" cluster, 91% web pages, appears in all six periods. The reference that bridges the most clusters is Doshi & Hauser 2024, cited by 8 notes across five of the six periods. The site's Citation Network page draws this result as fifteen works: four live works whose arrangement is generated by forces on every view (marbling on a plane with the proportions of the frame, the weave on a free plane with no frame, the constellation on a sky mapped by stereographic projection, the rings in cells of period and cluster), five live works moved by swarm rules (slime mould with Physarum-style particles, school with Couzin's three-zone model, cells with the Primordial Particle System, foam with weighted centroidal Voronoi, brushstroke with swarmalators), three stills based on hierarchical edge bundling, arc diagrams, and streamgraphs, and three stills drawn by swarms (dry garden with the Vicsek model, coral with reaction–diffusion in a robot swarm, roots with coalescing token passing). Below them the page lists five lines under "What stands out" whose conditions are checked against the data at build time (the only cluster cited in all six periods, clusters made almost entirely of papers being cited mostly in one period, clusters that never share a note, the homogenization paper cited by notes from five periods, and pairs of notes sharing 10 or more sources all written within three days). The default work is marbling: in the order the notes were written, it drops one pair of ink and water drops per note at the centroid of the points of the sources the note cites, and with the drop map of mathematical marbling and a divergence-free stream function it pushes the rings of earlier notes outward and stretches them longer. The points of the sources stay where the layout put them. "Generate again" changes only positions and the shapes of lines and rings that follow from them, never the connections. In marbling and the constellation, repulsion between sources of different clusters reaches further and pushes harder than within a cluster, so clusters read only through how densely points crowd. The look follows the site theme: the light one is sumi ink on washi paper, the dark one is light added up over a dark field. Clusters are shown in order of paper share, by the shade of one ink in the light look and the brightness of one hue of white light in the dark look; the kind of a reference by two shapes, a filled disc (academic: papers and preprints) and a ring (practice: companies, public bodies, and the web, such as news and blogs); and the age of the link a line stands for by that line's own bleed (ink spreading in the light look, a glow spreading in the dark one; the darker edge and the granulation draw on Curtis et al. 1997). Only the weave's threads do not bleed. Every work is first drawn in thin lines, and once drawing ends, a per-work "Finishing" line appears and the bleed is added over about ten seconds. The counting, the force computation, and each work's formulas and curve construction (the drop map, a stream function with the midpoint method, centripetal Catmull–Rom curves, stereographic projection, uniform cubic B-splines, half circles, Gaussian kernels) are transcribed from the implementation in the body.

• Drawing Networks as Data Art: Literature on Placement, Bundling, Time, and Flow

  2026-09-24

  A force-directed drawing of a network looks like the one true shape of that network, but it is only one of many arrangements. This note lays out the force equations of force-directed placement (Eades 1984, ForceAtlas2) and force-directed placement on a sphere (Kobourov and Wampler 2005), hive plots that set coordinates from attributes, edge bundling (Holten 2006), layouts along a single axis and along time (arc diagrams, ThemeRiver, streamgraphs, alluvial diagrams), mathematical marbling that makes patterns by dropping paint on water (Lu et al. 2012, Jaffer's drop mapping), prior works that treat data as art, and the curves and numerical methods used to draw the lines (centripetal Catmull–Rom curves, uniform B-splines, the midpoint method, Gaussian kernels, stream functions), and ways of rendering ink bleeding into paper (the watercolour model of Curtis et al. 1997, MoXi by Chu and Tai 2005, sumi painting by Guo and Kunii 1991), each with how far it was verified (full text, abstract, or bibliographic only). On this wiki's /network page, these methods were used to rearrange the same co-citation relations, which stay fixed while the arrangement changes with the method and the seed. The space each force-directed work is laid out in is fixed per work (marbling: a plane with the drawing area's aspect ratio; weave: an unbounded free plane; constellation: a sphere projected stereographically). The variation across seeds comes, for the weave, from initial-configuration families and ranges of force coefficients (including ForceAtlas2's LinLog and strong-gravity options), and for marbling and the constellation, from jitter in the initial placement and small ranges of coefficients.

• What Data Visualization and Data Art Share, and Where They Differ

  2026-09-22

  The same line chart is called a chart in a newspaper and data art on a gallery wall. The two share their material, their visual variables, their makers, and even their venues, so the medium and the tools cannot separate them. Experimental work undercuts the common claim that visualization is about accuracy and art is about beauty, from both sides. Theories of art split three ways on whether the distinction can be settled objectively: by properties of the object, by no single criterion, or by institutions. Empirical studies show low agreement between judges, so no objective test on the object alone holds up. What can be recorded objectively is not which one a work is, but which one it was treated as, which makes the distinction usable as a procedure across three layers: what the figure asks of its reader, the intent and venue, and the institution that certified it.

• What Do the "X Fiction" Methods Think, and What Are They Expected to Do?

  2026-09-22

  This note reads academic work on science fiction prototyping, design fiction, critical and speculative design, diegetic prototypes, speculative fiction, experiential futures, and social fiction, asking three things: where each idea came from, how the ideas differ, and what adopters have expected of them. The names sound alike, but the origins are separate: SF and technology development (Johnson, Kirby), design critique at the RCA (Dunne and Raby), futures studies (Candy), and literary genre theory (Oziewicz). The largest difference is whether the fiction is meant to move toward production or to stay away from production and keep a question open. Johnson tied science fiction prototypes to fostering innovation, while Dunne and Raby named the opposite of critical design "affirmative design: design that reinforces the status quo." Adopters' expectations fall into five groups: organizational foresight, anticipation in responsible research and innovation (RRI), eliciting values in HCI, ethics education, and imagination by marginalized communities. A corporate design fiction project that drifted toward affirmative design (Ringfort-Felner et al. 2023) and the critique that speculative design stays inside galleries and privilege can both be read as a mismatch between these two directions. Evidence of effects is covered in sf-prototyping-design-fiction-effects-literature.

• What Do Science Fiction Prototyping and Design Fiction Actually Do?

  2026-09-22

  This note reads 81 academic works on the "effects" of science fiction prototyping (Johnson 2011) and design fiction (Bleecker 2009), sorted into four settings: the route by which technology shown in films moves the real world, organizational foresight, HCI research tools, and education. What has mainly been established is that these methods can draw discussion and values out of participants. A deployment study of the HawkEye probe (Noortman et al. 2019) reports empathic discussion about the consequences of future technology, and a Detroit workshop that centered Afrofuturism (Dillahunt et al. 2023) drew alternative economic visions out of residents. By contrast, no study in this corpus compared effects on learning or organizations against a control group, and the only quantitative pre/post comparison is Nicholes (2020), which measured how 83 college students perceived writing. Studies also report that the criteria for evaluation are themselves unsettled (Baumer et al. 2020), that the pursuit of plausibility edges toward deception (Coulton et al. 2016), and that in a course meant to make students imagine critical futures, the students reproduced dominant views of technology (Rahm 2024).

• What Social Coordination Costs Arise from Collaborating with Others on Creative Work

  2026-09-22

  Adding people does not make work progress in proportion to headcount; the difference disappears into coordination loss and motivation loss (Steiner 1972), and into a relational loss from feeling that it is harder to get help (Mueller 2012). Coordination cost is not a single kind; it needs to be seen as six kinds that arise for different reasons: synchronizing logistics and information, loafing, the awkwardness of being evaluated, conflict, reconciling differences in interpretation, and negotiating names and shares. In creative work, beyond the wait for one's turn in brainstorming (Diehl & Stroebe 1987), the content of the output itself — word choice, how a problem is framed — becomes an object of negotiation, which makes it hard to isolate coordination as purely a matter of logistics. Conventions and transactive memory make coordination cheaper, but in Broadway's collaboration networks the relationship between the degree of cohesion and success is an inverted U (Uzzi & Spiro 2005), and past a certain point cheaper coordination and better outcomes stop moving in the same direction. Replacing the partner with AI doesn't remove this either: a robot facilitator does not soften the effect of evaluation apprehension, loafing still occurs in human-AI teams, and while AI-sourced ideas raise the evaluation of individual works, the works come to resemble each other more (Doshi & Hauser 2024) — so it cannot be read as "with AI, social coordination cost is zero."

• Taking the LLM Smell Out of LLM Prose, and How to Bring humanizer Into This Project

  2026-09-22

  This note sorts out what makes LLM prose recognizable, drawing on 23 academic papers, vendor documentation, Wikipedia's rules, and well-known skills. People cannot tell AI text apart by intuition (Jakesch et al. 2023), yet excess vocabulary (Kobak et al. 2025) and recurring sentence and structure patterns can be detected statistically, and model-specific habits survive rewriting (Sun et al. 2025). Interventions happen before generation, during generation, after generation as detection, and after generation as rewriting, and each has its own conditions and costs. The best-known skill, humanizer (51,020 GitHub stars), is an English-only skill built on Wikipedia's Signs of AI writing. It sorts 25 patterns into 5 groups and acts on weak tells only when several appear together. This project's Japanese style rules already stop the banned words, but patterns such as negated contrasts and fixed-count enumerations remain in existing notes. The sensible order is to add those patterns to the rules and a report-only check to wiki:lint first, then vendor humanizer for the English notes.

• What Was Measured as "Gyaru-Mind": The Provenance of Eight Factors and Its Limits

  2026-09-22

  A detailed account, drawn from the full texts of IWSDS 2026 and CHI EA 2026, of how Ikegami Momoka and colleagues at Komazawa University operationalized "Gyaru-Mind," followed by an examination of its limits. The eight factors were decomposed from a single trade book through discussion among the authors, and the criterion scores were assigned by the first author alone. Predictive accuracy stands at Spearman's ρ = 0.198 with a confidence interval spanning zero. The sum of the standardized coefficients, 6.587, falls below the RMSE of 8.84, which — under a stated assumption — would mean the model's error exceeds the spread it can produce between people. CHI EA reports a controlled study with N=24 in which PANAS shifted while trait self-efficacy did not. In a two-week follow-up, one of two participants reported that the effect "lasted a few hours" and was "reset" by the next day. Fourteen limits are set out, including the citation of Kinsella (2005) on racialization as a source for a neutral description. The note also shows that Ikegami and colleagues were not the first to scale Gyaru-Mind: Motode Rino (2023, Kwansei Gakuin University) preceded them by two years with a four-factor scale tested for reliability and criterion-related validity on 1,487 respondents — a work the authors cite in their Japanese version and drop from the English one. With references.

• How Data Visualization and Data Art Have Been Treated in Creativity Education

  2026-09-22

  Data education runs in two streams. One teaches students to read figures correctly and measures that ability with instruments (VLAT, CALVI, Mini-VLAT, a multidimensional scale), and it has accumulated critiques of those very instruments. The other brings the making of data works into the classroom through course design and public exhibitions, and it has no standard scale for outcomes. What bridges them is the making-to-learn lineage that descends from Papert's constructionism and the teaching of data physicalization. Arts-integrated practice reports the same tension again and again: artistic creativity against the accuracy of data. In the Japanese-language literature, research that uses data art in education is almost absent.

• How to Use Jev, and Where Not To

  2026-09-21

  TypeSafe AI's Jev works only as far as a judgment can be written down as typed questions. Six days after launch there is no consensus that could be called best practice, so this note keeps only what the vendor states verbatim with a condition attached and implementations published in reproducible form, dropping superiority claims without methodology and reports whose provenance could not be checked. The official docs limit it to ten task types and explicitly exclude arithmetic, dates, multi-step reasoning, and generation. The craft sits in how criteria are written: describe situations rather than degrees for Score, avoid negated phrasing for Noul, never shorten the option list for Choice. Thresholds come with a starting point of 0.6 and 0.85 and an instruction to tune them against your own data, with the vendor itself calling confidence a convenient measure. Independent implementations report 91.5% agreement on an eval task at $160 per million judgments and 91.7% accuracy on agent tool-call risk, but accuracy drops to 71.4% on ambiguous cases, and with no abstain option the model is forced to pick the least wrong answer. What decides adoption is not accuracy but whether the criteria can be written out in advance.

• When Judgment Becomes a Function Call, What Moves in Design

  2026-09-21

  TypeSafe AI's Jev returns no text, only typed judgments and probabilities, at $0.0004 and 0.4 seconds a call. What is new is not the training method but the fact that judgment now takes the shape of a component callable hundreds of times. This blog has written that 'craft gets automated and judgment is what remains,' and that frame does not hold as it stands. Independent testing and the LLM-as-a-Judge literature draw the same line from separate places: applying fixed criteria comes down, while ranking which option is better and writing the criteria themselves stay up. Jev also returns nothing but scores and probabilities, which runs head-on into a prescription this blog produced in debate (limit AI evaluation to candidate next moves rather than scores). The price of inspection hits the floor, substitution occurs where adoption and approval rates cannot detect it, and liability stays with whoever used it.

• AI and Design Weekly Watch (2026-09-14 to 09-21)

  2026-09-21

  An integrated summary of 'AI and design' developments over the past seven days, collected in three tiers: T1v vendor primary sources, T2 public institutions and standards, and T3 expert opinion. On the vendor side the week leaned toward distribution and education rather than new features: Figma opened community publishing for Weave tools, Canva launched an AI Essentials Certification, and Adobe expanded its education programs. No new T2 publication could be confirmed from primary sources, and no analyst report was adopted. In T3, Jakob Nielsen ranked AI prototyping first among 100 UXD methods, while Laurie Voss, Ethan Mollick, and NN/g's Tanner Kohler discussed the judgment and context-giving work that remains once the cost of building falls.

• What Replaced What in Test Scoring and Item Writing

  2026-09-18

  A focused industry survey of two stages of Japanese mock exams and large-scale testing: scoring and item writing. What digital scoring replaced was not judgement but the movement of paper, the tallying, and the reconciliation. Partial-credit judgement stayed with people, and a new verification step appeared. The workforce moved from the marking room to the home, but piece-rate pay did not change. Item writing has not been replaced because its hours are not recorded, its review is layered, and its content cannot leak. Generative AI reaches a draft and no further, a position that is both where the technology lands and where the rules draw the line. The defects in AI-written items concentrate in ambiguity and difficulty control rather than duplication, and generating and verifying with the same model lets flawed items pass. Scoring is now counted as a market; item writing is not. METI's final results report was obtained at source: it breaks the saving down (10 to 30 minutes for 40 percent of teachers, no gain for 21 percent, longer for 15 percent), limits the demonstration to the first of three phases, and records nothing about scoring.

• The Workload of Mock-Exam Operations and Where AI Actually Lands

  2026-09-18

  An industry survey of Japanese mock-exam operations, mapping each stage against where the load concentrates and where AI has actually arrived. The load concentrates on scoring constructed responses; digitalization shortened the stages around scoring (application, collection, digitization, return) rather than scoring itself. Scoring automation remains narrow, and digital scoring created a new verification job rather than removing scoring. Demand shrinks with the birth rate while supply tightens through teacher shortages and a contracting scoring labor market.

• TypeSafe AI's Jev: What It Guarantees and What It Does Not

  2026-09-18

  TypeSafe AI released Jev on 2026-09-15. It generates no text; it returns typed judgments and probabilities. In the vendor's own evaluation it scores 67.8% accuracy against the best comparator's 74.1%, while winning by orders of magnitude on cost and latency. What is guaranteed is the shape of the output, not the correctness of the judgment. Independent tests split along a line: strong on classification tasks, weak on calibration and ranking.

• How do you measure a business analyst's skills?

  2026-09-15

  Japan's Digital Skill Standard added Business Analyst as a role in ver.2.0 (April 2026), so the job title now has a public definition. It has no ruler. The standard states that it sets no detailed level indicators, and instead assigns each skill an importance of a/b/c/d/z per role, which is a weighting rather than a proficiency scale. IIBA's BABOK Guide standardises six knowledge areas, thirty tasks and 29 underlying competencies, but leaves individual proficiency to the Competency Model's self-assessment scale. Its exams publish their weightings while withholding the passing score, and recertification counts activity hours. The public measuring tools are the Information Technology Engineers Examination and ITSS experience judging; the professional bodies lean on hours, exams and CDUs; the hiring market screens on kinds of experience. None of them measure outcomes. The standard gives vocabulary, not evidence.

• AI and Design Weekly Watch (2026-09-07 to 09-14)

  2026-09-14

  An integrated summary of 'AI and design' developments over the past seven days, collected in three tiers: T1v vendor primary sources, T2 public institutions and research, and T3 expert opinion. Vendor-side announcements were quiet — mostly incremental updates like Framer's list-type CMS field and OpenAI's full-duplex voice API GPT-Live-1 — while Jakob Nielsen, Luke Wroblewski, and NN/g each independently published pieces questioning the design of AI chat interfaces themselves: the mismatch of linear chat structure, answer-first response design, and practical use of AI prototyping.

• Which standard can design and new-business work actually sit on?

  2026-09-14

  Listing the stages you covered on a portfolio needs a basis, and a self-made taxonomy has none. Japan's obvious candidates both fail: IPA froze Common Frame 2013 and says itself that it no longer aligns with ISO/IEC/IEEE 12207:2017, and the Digital Skill Standard ver.2.0 is a standard for personnel types and skills, not for stages. What held up were ISO 9241-210 (human-centred design, JIS Z 8530) and ISO 56002 (innovation management, JIS Q 56002), both of which state iteration in the standard text. Consulting has ISO 20700 but no JIS and no Japanese edition, marketing has no public standard for the work as a whole, and no public standard defines agile.

• 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.

• AI and Design Weekly Watch (2026-08-31 to 09-07)

  2026-09-07

  An integrated summary of 'AI and design' developments over the past seven days, collected in three tiers: T1v vendor primary sources, T2 public institutions and research, and T3 expert opinion. Google embedded image generation/editing (Google Pics) directly into Docs and Slides, Adobe made Firefly and 70+ tools callable from Slack, and Webflow announced an agent-collaboration platform for site operations (Source by Webflow), all pointing to generative AI's output surface moving from dedicated creative apps into documents and chat. The same week, Anthropic and OpenAI each shipped a new flagship model (Claude Fable 5.1/Mythos 5.1 and GPT-6 Astra), and Simon Willison benchmarked both with the same informal test.

• AI and Design Monthly Scholarly Watch (August–September 2026)

  2026-09-07

  The third installment of the monthly scholarly watch, collecting 7 peer-reviewed papers and preprints at the intersection of AI and design, centered on August 4 through September 7, 2026. In contrast with the previous watch's concentration around C&C 2026, this window includes no major conference and the count is thin, but two threads — preserving controllability and designer agency in AI-assisted tools, and empirical evidence from design education that AI reorganizes the decision-making process itself — give a concrete mechanism to the standing conflict from earlier watches.