# Designing Assistant Technology ## Review ### Verdict **Recommended as a design workbook, with an evidence caveat.** Noessel gives product teams a useful paired model: five places an assistant can help—perceive, know, plan, perform, and reflect—and a risk test for deciding how much judgment must remain with the person. The book is strongest when it turns that model into workshop questions and interface patterns. It is weakest when small studies, selected cases, or suggestive findings carry universal language. Use this book to generate and test design hypotheses. Do not treat its five assists as a proven exhaustive taxonomy, its **Human Goes First** pattern as an automatic default, or its quantitative research summaries as ready-to-repeat facts. One prominent number fails source checking: the cited medical study reports a 1.85 deterioration factor when AI advice was wrong, not that clinicians became “6.4 times worse.” The broader warning about overreliance still holds. The durable contribution is a better success criterion: an assistant should improve the joint task without quietly making the person less capable, less attentive, or unable to recover when the system fails. That criterion belongs beside speed, completion, and satisfaction—not beneath them. ### The book in one paragraph Noessel argues that products should be designed as assistants when a person must remain engaged with the work, rather than as tools that wait for direct control or agents that take over a delegated outcome. He maps assistance onto five recurring human activities: perceiving the situation, knowing relevant facts, planning, performing, and reflecting. He then shifts from opportunity to risk: AI assistance can create under-reliance, overreliance, and skill loss even when the product feels usable. For consequential work, he favors a **Human Goes First** sequence in which the user attempts the task, compares their work with the system’s independent result, and receives targeted learning. Cognitive forcing functions, pace-layer analysis, and attention to user differences help teams tune that sequence. The final chapters place the design problem inside labor, organizational, and social choices: the goal is not automation for its own sake but people who remain capable after the assistant is gone. ### What changed my view - “Usable” and “good for the user” are different judgments. A fast, pleasant assistant can still remove the practice that sustains skill. - Appropriate reliance has two failure directions. A design can cause blind acceptance or needless rejection, and an intervention that reduces one can worsen the other. - Preference and performance must be measured separately. People may dislike friction that improves judgment, but that does not excuse burdens that block disabled users or add effort where the skill has no lasting value. - The useful design unit is not “AI feature.” It is the human task loop, the harm if the system is wrong, the user’s opportunity to notice, and the recovery path. ### Best use Bring the book into discovery, risk review, and prototype planning. Run the five assists across one real task; identify which human capabilities should remain durable; choose an assistance pattern; and test joint quality, unaided performance, override behavior, burden, and accessibility. Pair the method with domain research, representative user testing, WCAG, cognitive-accessibility guidance, and a formal risk process when consequences are high. > “It’s just—usable is not enough.” — Christopher Noessel, `chapter2.xhtml#pg_16` (system-selected) > “The design of assistants needs to help users build appropriate reliance, avoiding both under-reliance and overreliance.” — Christopher Noessel, `chapter11.xhtml#pg_121` (system-selected) > “When the stakes are high, vibe is an irresponsible metric.” — Christopher Noessel, `chapter12.xhtml#sec10` (system-selected) > “Well-designed AI can make people smarter for having used it, even after they put the assistant down.” — Christopher Noessel, `chapter15.xhtml#pg_178` (system-selected) ## PROMPT 0 — Initialize the book record ### Source-of-truth record | Field | Result | |---|---| | Book | *Designing Assistant Technology: How to Build Products That Help People Think* | | Author | Christopher Noessel | | Edition | Rosenfeld Media ebook, published March 2026 | | Ebook ISBN | 978-1-959029-06-9 | | Paperback ISBN | 978-1-959029-60-1 | | Local source | EPUB 3.0; SHA-256 recorded in frontmatter | | Integrity | ZIP and XML passed; 1,593 internal references checked with 0 broken references | | Coverage | 36/36 spine items dispositioned; 15/15 chapters and 3/3 appendices read | | Anchors | XHTML filename plus fragment ID, such as `chapter11.xhtml#pg_121` | | Readwise | No matching Reader or Classic Readwise record found; no target highlights available | | Reader emphasis | 🔶 Unknown. Every excerpt and hotspot in this note is `system-selected`. | | Standards check | EPUBCheck was unavailable, so full EPUB standards conformance is unverified | The OPF identifies the ebook ISBN, title, author, year, and Lightning Source as publisher metadata. Rosenfeld Media’s title page confirms the March 2026 release, subtitle, formats, and ISBNs. This is a different book from Noessel’s 2017 *Designing Agentive Technology*. ### Layer and attribution rules used - **Layer 1:** local EPUB, used for the author’s argument, terminology, examples, and exact wording. - **External evidence:** primary studies, official standards, and official legal or risk-management texts used to test live claims. - **EA:** explicit author claim. - **IA:** inference from the author’s structure or examples. - **Q3P:** quoted or paraphrased third-party claim carried by the book. - **RI:** reader interpretation. None is attributed to the reader because no highlights or purpose statement were available. - **UI:** reviewer inference in this note. ### Accessibility and production check The EPUB declares EPUB Accessibility 1.1 and WCAG 2.2 Level A conformance, with textual and visual access modes, structural navigation, reading order, a table of contents, an index, page navigation, and alternative text. The package also says it was generally produced toward Level AA while noting that language shifts and color contrast were not fully checked. A direct audit found 112 images and one table. Most meaningful figures have detailed alternative text. Ten images use empty alternative text and presentation roles; most appear decorative, but the Greek-root image in Chapter 7 conveys content and should not be hidden. The single table lacks header cells and a caption. Footnote links are not consistently marked with semantic note relationships. These findings limit confidence in the package-level accessibility claim. ## PROMPT 1 — Establish the reading contract ### Inferred purpose 🔶 **Reader purpose was not supplied.** I infer a practical purpose: evaluate the book for product design, technical communication, AI-assisted workflows, and teaching. I use that assumption to favor reusable models, evidence quality, accessibility, and concrete experiments. If the purpose is historical study, literary criticism, or a purchase decision, the priority order should change. 🔶 **Prior beliefs were not supplied.** The questions below are reviewer-generated, not a record of the reader’s beliefs. ### Questions this reading should answer 1. What makes an assistant distinct from a tool, automation, or agent? 2. Are perceive, know, plan, perform, and reflect a useful and complete opportunity map? 3. Which human capabilities should a product preserve, and who gets to decide? 4. How can a team detect de-skilling before a product ships? 5. What does appropriate reliance look like in observable behavior? 6. When does forcing the human to act first improve judgment or learning? 7. When does “human first” create delay, cognitive burden, or an accessibility barrier? 8. How should teams measure under-reliance and overreliance separately? 9. Which examples establish only plausibility, and which provide causal evidence? 10. How strong is the empirical basis for the book’s quantitative claims? 11. How should assistance change with stakes, urgency, reversibility, and skill durability? 12. What role should explanations, confidence, and uncertainty play? 13. How do user differences affect the amount and timing of friction? 14. Does the framework account for disabled users and assistive technology? 15. Whose labor, autonomy, and accountability are changed by the assistant? 16. What must teams add from current governance and accessibility standards? 17. Which parts will remain useful as model capabilities change? 18. What small tests can turn the book’s patterns into local evidence? ### Prior-belief register | Topic | Starting position | Confidence | What could change it | |---|---|---:|---| | Usability | A usable assistant can still harm skill or judgment. | Medium | Longitudinal evidence that common assistance preserves or improves unaided capability | | Automation | More automation is not automatically better. | High | Reliable results showing transfer across tasks, populations, and failure cases | | Human oversight | A person in the loop is not proof of meaningful oversight. | High | Behavioral evidence of detection, override, and recovery under realistic pressure | | Explanations | More explanation does not always improve reliance. | Medium | Replicated domain studies showing a stable intervention across user groups | | Friction | Productive friction can help when stakes and learning needs justify it. | Medium | Evidence that the same benefit can be achieved with lower burden | | Accessibility | Cognitive forcing can exclude people if it raises memory, timing, motor, or language demands. | High | Representative testing showing equivalent access and outcomes | | Personalization | Adaptation may help, but hidden cognitive profiling is risky. | Medium | Transparent, consented, privacy-preserving trials with consistent benefit | | Skill | Short-term unaided performance is not the same as long-term de-skilling. | High | Longitudinal real-work studies across months or years | | Taxonomy | A five-part map can be useful without being exhaustive. | High | Comparative validation against competing task models | | Governance | High-stakes assistants need evidence beyond product analytics. | High | No plausible disconfirming observation; the question is which evidence is proportionate | ### Reading contract I will distinguish the book’s claims from my inferences, test the highest-impact empirical claims against primary sources, preserve author concessions, and not turn selected excerpts into reader emphasis. I will recommend one operating interpretation rather than list all possible readings. The output remains one file; no Reader publication, adjacent notes, or reciprocal vault links are authorized. ## PROMPT 2 — Inspect the structure ### Structural map | Section | Purpose | Key move | Hotspot | |---|---|---|---| | Part I: The de-skilling problem | Establish the cost hidden by conventional usability success | A navigation prototype becomes a warning about lost practice | Small, informal, lost-data study bears too much weight | | Ch. 1, “Usable Is Not Enough” | Open with Google Maps and IBM navigation work | Separates task ease from retained capability | Reported 48% slower decision time after assisted use | | Ch. 2, “The Question Is Whether the Labor-Saving Is Worth the Cost” | Show that feedback can improve the same task | Recasts the interface from answer-giver to coach | Reported 19% gain shares the first study’s limits | | Ch. 3, “Tools, Automation, Agents, and Assistants” | Define the product category | Distinguishes categories by where work and attention sit | Boundaries blur in mixed-initiative systems | | Part II: The universal assists | Supply an opportunity framework | Hangs five kinds of assistance on a human task loop | “Universal” is asserted more than established | | Ch. 4, “The Five Universal Assists” | Introduce the taxonomy | Perceive, know, plan, perform, reflect | Useful checklist; weak exhaustiveness proof | | Ch. 5, “Assist Perception” | Improve what users can notice | Reveal, filter, transform, and call attention | Salience can become manipulation or hide context | | Ch. 6, “Assist Knowledge” | Put relevant knowledge in the task | Recall, explain, contextualize, and surface missing knowledge | Users may not know what to ask; summaries can erase detail | | Ch. 7, “Assist Planning” | Help users compare paths before acting | Manifest likely outcomes and tradeoffs | Forecasts can create false certainty | | Ch. 8, “Assist Performance” | Support action without taking the task away | Guide, constrain, rehearse, and execute parts | Boundary between assistance and agency is unstable | | Ch. 9, “Assist Reflection” | Turn completed work into learning | Feedback and review build metacognition | Reflection must be timely, specific, and safe | | Ch. 10, “A Workshop for Finding Assists” | Operationalize the framework | Map tasks, generate assists, prioritize by value and risk | Workshop output remains hypothesis, not validation | | Part III: Appropriate reliance | Address AI error and human dependence | Pair opportunity generation with risk-shaped interaction | Strongest conceptual integration in the book | | Ch. 11, “The Human Goes First” | Define reliance and propose a central pattern | User acts, AI works separately, results compare, learning follows | The 6.4 statistic is unsupported by its cited study | | Ch. 12, “Cognitive Forcing Functions” | Present interaction patterns that resist blind acceptance | Delay, on-demand advice, partial explanation, and update choices | Performance, liking, and under-reliance trade off | | Ch. 13, “Different People Need Different Assistance” | Add individual differences | Need for cognition affects response to forcing | Small IBM study is nonsignificant; profiling raises privacy issues | | Ch. 14, “AGI, Claude, and the Design of Attention” | Test the argument through speculation | Imagines an assistant that protects human engagement | Thought experiment, not product evidence | | Ch. 15, “The Future Is What We Make of It” | State the ethical and political endpoint | Judge assistance by capability after use | Choice sits with institutions, not designers alone | | Appendices | Explain acknowledgments, AI use, and notes | AI colophon makes production provenance visible | Useful disclosure but not an audit of every generated contribution | ### Provisional inspectional summary The book’s argument has two halves. The first is generative: classify the product relationship and look for assistance across perception, knowledge, planning, performance, and reflection. The second is protective: decide whether the person needs to retain the skill, identify the cost of a missed AI error, and shape the interface toward appropriate reliance. The bridge is the claim that ordinary usability measures miss what happens to human capability. The workshop chapter makes the first half actionable; Human Goes First and cognitive forcing functions make the second half testable. The final chapters widen the frame from interface design to labor and institutional power. The structure is clear and teachable. Its weak point is evidentiary scale: compact prototypes, selected studies, and memorable examples often support a design possibility, but the prose sometimes moves from possibility to general rule without enough comparative or longitudinal evidence. ### Guiding questions after inspection 1. Is the five-assist model exhaustive or only memorable? 2. Can the classification survive products that shift between tool, assistant, and agent by task stage? 3. What outcome proves that assistance preserved capability? 4. How long must unaided retention last to count? 5. Which skills are worth preserving, and whose values control that choice? 6. Can Human Goes First work under time pressure? 7. Does cognitive forcing reduce harmful reliance without causing harmful rejection? 8. What is the accessibility cost of each forcing pattern? 9. How do teams validate assistance with affected people who are not direct users? 10. Which claims depend on unstable model capabilities? 11. How should confidence and explanations be tested rather than displayed by habit? 12. What legal duties now make parts of the argument mandatory in covered settings? **Reading decision:** full analytical reading was warranted because the practical framework and the empirical cautions are interdependent. A selective read would risk adopting the opportunity checklist without its reliance safeguards, or adopting friction patterns without their user and evidence limits. ## PROMPT 3 — Build the vocabulary | Term | Working definition in the book | Operational test | Risk or boundary | Example / counterexample | Source | |---|---|---|---|---|---| | Assistant technology | Helps a person perform while their attention stays on the task | Can the user still see, judge, and influence the work? | A product may shift categories within one flow | Copilot that supports drafting / agent that sends without review | `chapter3.xhtml#pg_26` | | Tool | Extends direct human action | Does output depend on continuous user control? | Smart tools may make recommendations | Hammer / autonomous procurement bot | Ch. 3 | | Automation | Performs a predetermined procedure | Is the sequence fixed once triggered? | Exceptions may require judgment | Scheduled backup / open-ended research agent | Ch. 3 | | Agentive technology | Acts toward a delegated goal with some autonomy | Does the system choose steps or timing? | Agency can be partial and reversible | Travel-booking agent / calculator | Ch. 3 | | Assistive technology | Established disability-centered field; not a synonym for all assistance | Is the term being used in its disability context? | Casual reuse erases a field and its users | Screen reader / generic writing helper | Ch. 3 | | See–think–do | Simplified loop linking perception, cognition, and action | Which stage does the product change? | Reflection and social context can fall outside a short loop | Dashboard alert → diagnosis → action | Ch. 4 | | Five assists | Perceive, know, plan, perform, reflect | Has the team considered each against a named user task? | Useful heuristic; universality unproven | Workshop checklist / claim that no sixth class can exist | Chs. 4–10 | | Assist perception | Make relevant conditions detectable | Does a user notice the right signal at the right time? | Attention steering can manipulate or obscure | Highlight anomaly / hide base rate | Ch. 5 | | Assist knowledge | Surface facts or explanations needed now | Can the user verify and use the information? | Summaries may be wrong or omit minority cases | Source-linked definition / unsupported summary | Ch. 6 | | Assist planning | Help compare goals, options, and effects before action | Can the user inspect assumptions and alternatives? | Forecasts can imply certainty | Preview route outcomes / one unexplained “best” plan | Ch. 7 | | Assist performance | Guide or support execution | Does help preserve meaningful control and recovery? | The assistant can become an agent silently | Step-by-step coaching / irreversible autonomous action | Ch. 8 | | Assist reflection | Help a person learn from results | Does feedback change later independent performance? | Surveillance or shame can masquerade as feedback | Private skill review / public punitive score | Ch. 9 | | De-skilling | Loss or weakening of a capability through reduced practice or displaced reasoning | Is later unaided performance worse, less persistent, or less transferable? | Short-term effort reduction is not proof of lasting loss | Immediate post-removal test / faster assisted work alone | Chs. 1–2, 11 | | Appropriate reliance | Accept correct assistance and reject incorrect assistance at suitable rates | Measure correct acceptance, incorrect acceptance, and missed useful help | Accuracy alone hides under-reliance | Override bad advice while using good advice | Ch. 11 | | Under-reliance | Reject or ignore useful system support | Does the user decline correct advice without good reason? | Can be rational when trust evidence is weak | Rechecking every correct result | Ch. 11 | | Overreliance | Accept system output when independent judgment should reject it | Does a plausible error survive review? | “Human in the loop” does not prevent it | Rubber-stamped diagnosis | Ch. 11 | | Human Goes First | User attempts the task before seeing a separate AI result, then compares and learns | Does the sequence improve detection or retained skill without undue burden? | Urgency, disability access, and low-value skill can defeat it | Draft then compare / forced delay in an emergency | Ch. 11 | | Pace layers | Different parts of a system change at different rates | Which knowledge, skill, rule, or model will drift first? | A slow institutional process may be a hazard, not protection | Stable safety principle vs changing model frontier | Ch. 11 | | Cognitive forcing function | Interaction constraint that prompts deliberate thought before reliance | Does it reduce wrong acceptance without excess wrong rejection? | Can increase load, delay, and dislike | On-demand advice / decorative confirmation click | Ch. 12 | | Need for cognition | Preference for effortful thinking used as one moderator | Do outcomes differ under measured, consented user characteristics? | Not a fixed ability; profiling can be invasive | Optional depth setting / hidden “low thinker” label | Ch. 13 | | Augmentation | Assistance that leaves a person more capable, not merely more productive | Is capability retained when the product is removed? | Some tasks should be automated, not preserved | Coach that teaches / answer service that replaces practice | Ch. 15 | ## PROMPT 4 — Inventory the claims Claim types: **EA** explicit author claim; **IA** inference from the book; **Q3P** third-party result as presented by the book; **UI** reviewer inference. | ID | Claim and attribution | Type / scope | Evidence expected and offered | Non-vacuous falsifier | Key / anchor | |---|---|---|---|---|---| | C-001 | Usability can coexist with lost skill. | EA; general design | Longitudinal or post-removal measures; Google Maps framing plus IBM prototype | Repeated representative studies show equal or better unaided retention under answer-giving assistance | Chs. 1–2 | | C-002 | Four assisted virtual commutes made IBM participants 48% slower at navigation decisions on day five. | EA/Q3P; tiny prototype | Raw times, sample, protocol; reported study, recordings later lost | Recovered data or replication fails to show the reported direction | Ch. 1 | | C-003 | A feedback version produced a 19% improvement. | EA/Q3P; tiny prototype | Same as C-002; reported intervention result | Replication shows no gain or shows it came from an uncontrolled difference | `chapter2.xhtml#sec3` | | C-004 | The key question is whether labor saved is worth the human cost. | EA; normative | Decision criteria and affected-party evidence; cases and ethical argument | A case has no meaningful lost capability, judgment, agency, or distributional cost | `chapter2.xhtml#sec1` | | C-005 | Assistants are distinguished by keeping human attention on the task. | EA; taxonomy | Comparative category analysis; definitions and examples | A product keeps attention on-task while functioning wholly as a direct tool or autonomous agent under the same criteria | `chapter3.xhtml#pg_26` | | C-006 | Assistance can create learning inside ordinary work. | EA; design possibility | Transfer or retention tests; feedback examples | Users improve only with the assistant present and show no later transfer | Chs. 2, 9 | | C-007 | Perceive, know, plan, perform, and reflect are universal assists. | EA; universal taxonomy | Broad cross-domain coverage and rival taxonomy comparison; conceptual mapping | A recurring, useful assistance class cannot be placed in any of the five without distortion | `chapter4.xhtml#fig4-2` | | C-008 | Manifesting outcomes in advance helps users choose correctly the first time. | EA; planning pattern | Controlled choice outcomes across domains; paired examples | Previews worsen choice quality through anchoring, false certainty, or overload | `chapter7.xhtml#sec3` | | C-009 | Reflection is key to improvement. | EA; broad learning | Feedback and metacognition studies; cases | Matched practice without reflection produces equal or better durable improvement | `chapter9.xhtml#sec1` | | C-010 | The five-assist workshop can reveal product opportunities. | EA; method | Team trials, coverage, and downstream validation; workshop procedure | Teams generate no useful hypotheses beyond ordinary task analysis or miss recurring needs | `chapter10.xhtml#sec1` | | C-011 | High-stakes assistants should aim for appropriate reliance, not maximal trust. | EA; safety principle | Correct/incorrect acceptance behavior; research and risk reasoning | Maximal trust yields better harm-adjusted results across realistic correct and incorrect cases | `chapter11.xhtml#pg_121` | | C-012 | A task should be screened by asking what happens if AI is wrong and the human misses it. | EA; risk heuristic | Hazard discovery and test coverage; design question | The question adds no material hazards or controls beyond standard task analysis | `chapter11.xhtml#sec3` | | C-013 | In a cited medical study, clinicians were “6.4 times worse” when AI confidently erred. | EA/Q3P; specific empirical | Original paper; book’s summary | Original result lacks 6.4 or measures a different factor | Ch. 11; **falsified by S2** | | C-014 | Human Goes First can protect judgment and learning. | EA; proposed pattern | Controlled comparison with retention, reliance, burden, and access measures; pattern reasoning | It produces no better error detection or retention, or harms urgent/access needs | Ch. 11 | | C-015 | Cognitive forcing can reduce overreliance. | EA/Q3P; bounded interaction research | Correct and incorrect advice trials; cited experiments | Wrong-advice acceptance does not fall, or overall harm rises through under-reliance | Ch. 12; S3, S5 | | C-016 | Designs that reduce overreliance may be disliked. | EA/Q3P; bounded experiment | Performance and preference measured separately; Buçinca et al. | Effective forcing is equally or more preferred without added cost | `chapter12.xhtml#sec9`; S3 | | C-017 | Partial explanations reduce overreliance. | EA/Q3P; task-dependent | Correct/wrong reliance and task performance; de Jong et al. | Wrong reliance stays flat, or reduced correct reliance erases net benefit | Ch. 12; S5 | | C-018 | Need for cognition can moderate the effect of forcing functions. | EA/Q3P; individual-difference research | Prespecified interaction and replication; studies plus small IBM test | Moderation fails across adequately powered replications or is explained by task familiarity | Ch. 13; S3, S5, S6 | | C-019 | “Treat everyone as high NFC” may be a safer high-stakes default. | EA; design recommendation | Representative safety and access trials; inference from mixed findings | It increases errors, abandonment, inequitable burden, or workarounds without net safety gain | Ch. 13 | | C-020 | Product categories and reliance needs are preferences shaped by context, not fixed human constraints. | EA; design framing | Cross-context behavior; author caveat | Stable constraints explain the choice regardless of context or user aim | `chapter13.xhtml#pg_150` | | C-021 | AGI design in Chapter 14 is a thought experiment, not a forecast. | EA; scope claim | Explicit framing; speculative screenplay | The chapter presents its scenario as empirical prediction | Ch. 14 | | C-022 | Good AI can leave people more capable after they stop using it. | EA; normative endpoint and empirical hypothesis | Delayed unaided tests and transfer; synthesis and examples | Assisted users show equal or worse durable capability with no offsetting justified benefit | `chapter15.xhtml#pg_178` | | C-023 | Narrow technical success can still fail socially. | IA; socio-technical | Affected-party outcomes and distributional analysis; labor and power discussion | All meaningful benefits and harms are captured by task metrics in a tested case | Ch. 15 | | C-024 | Teams should measure joint performance, retained skill, reliance, burden, and access separately. | UI; operating synthesis | Multi-outcome product study; derived from book plus external evidence | One metric predicts every material outcome without hiding tradeoffs | S1–S15 | ## PROMPT 5 — Set confidence and calibration rules ### Confidence policy | Band | Meaning | Use | |---|---|---| | High | Direct source, suitable design, replicated or authoritative standard; key limits known | May guide a decision inside the tested scope | | Medium | Credible direction but narrow setting, indirect measure, limited replication, or incomplete transfer | Treat as a design hypothesis and test locally | | Low | Anecdote, tiny informal study, speculative extension, missing data, or unsupported generalization | Use only to generate questions | | Rejected | Source check contradicts the stated fact | Do not repeat; replace with corrected claim | ### Prior-to-posterior updates | Claim | Prior | Evidence that moved it | Posterior | |---|---:|---|---:| | C-001: usability may hide skill loss | Medium | Short-term causal learning and unaided-performance findings across math, reading, and coding; no long-term workplace proof | Medium-high within learning tasks; medium generally | | C-007: five assists are universal | Low-medium | Strong conceptual coverage, no comparative validation or proof of exhaustiveness | Low as a universal claim; high as a checklist | | C-011: aim for appropriate reliance | High | Human–AI experiments, NIST guidance, and EU oversight duties align with the direction | High | | C-013: 6.4 medical deterioration | Medium before source check | Original paper reports 1.85 and better hybrid performance overall | Rejected | | C-014: Human Goes First protects learning | Medium | Answer-first alone did not produce learning; explanation-without-answer and guarded tutoring did better | Medium-low for the full pattern; medium-high for preserving cognitive work | | C-015: forcing reduces overreliance | Medium | Controlled studies support reduction but show dislike, under-reliance, and task effects | Medium-high in tested tasks; medium generally | | C-018: NFC moderates benefit | Medium | Several studies show interaction, but results vary and learning optimization did not consistently replicate | Medium | | C-022: good assistance can build durable capability | Medium | Education and coding studies show design-dependent retention; little long-term field evidence | Medium as a goal; low-medium as a common outcome | ### Calibration warnings - Do not confuse the number of examples with independent evidence. - Do not infer causation from a case description, survey, or author reflection. - Do not transfer effect sizes across domains, models, populations, or time horizons. - Do not treat “human review” as effective oversight without observing detection and override. - Do not collapse reduced effort, immediate learning loss, and long-term de-skilling into one outcome. - Do not count a claim as verified because a citation exists; inspect the cited result. - Do not treat preference as safety or dislike as evidence of effectiveness. - Do not promote a framework from useful to universal without rival tests and counterexamples. - Do not apply a forcing function without measuring under-reliance and accessibility costs. - Recheck model-capability boundaries, law, and NIST guidance because each can change. ## PROMPT 6 — Analyze the argument chapter by chapter ### Chapter analysis | Chapter | Purpose and thesis | Claims and evidence | Counterpoint, assumption, or open question | Role | |---|---|---|---|---| | 1 | Break the equation of usability with user benefit. Navigation help may trade immediate ease for later skill. | Google Maps experience; IBM virtual-commute prototype; reported 48% slowdown. | Sample, protocol, and raw evidence are not recoverable. How much navigation skill should a product preserve? | Premise; low-confidence empirical illustration | | 2 | Show an alternative: feedback can support both task and capability. | Revised prototype; reported 19% improvement; conceptual tradeoff. | Same evidence limits as Ch. 1. Could novelty or practice explain the result? | Repair to opening problem | | 3 | Define assistants against tools, automation, agents, and assistive technology. | Category definitions and examples. | Mixed-initiative products change category by phase; attention is hard to observe. | Conceptual foundation | | 4 | Claim five universal places to assist a task. | See–think–do bowtie; mapping of assists. | Exhaustiveness isn't tested against rival task models. | Main opportunity framework | | 5 | Expand perception assistance. | Product examples of filtering, translation, salience, and revealed signals. | Who chooses what is salient? What information is suppressed? | Framework elaboration | | 6 | Expand knowledge assistance. | Retrieval, definitions, contextual help, summaries, and prompting. | A person may not know what to ask; source quality and omission need explicit controls. | Framework elaboration | | 7 | Expand planning assistance. | Previews and comparisons that manifest outcomes before commitment. | Predicted outcomes can anchor decisions or hide uncertainty. | Framework elaboration | | 8 | Expand performance assistance. | Guidance, constraints, rehearsal, and partial execution. | At what point does the assistant take over the task or accountability? | Framework elaboration | | 9 | Expand reflection assistance. | Feedback, review, and deliberate learning after action. | Reflection can become surveillance; feedback needs psychological and organizational safety. | Framework elaboration | | 10 | Turn the five assists into a workshop. | Task mapping, idea generation, prioritization, and McGregor example. | Workshop productivity is not proof of user value; hypotheses still need field testing. | Method | | 11 | Introduce appropriate reliance and Human Goes First. | Reliance research, pace layers, risk prompt, consulting and medical studies. | Medical multiplier is wrong; pattern lacks integrated validation; emergencies may require AI-first help. | Argument pivot and primary pattern | | 12 | Offer cognitive forcing functions. | Buçinca and related studies; partial explanations; preference/performance split. | Reduced wrong acceptance can coincide with reduced correct acceptance. | Pattern catalog and evidence | | 13 | Account for individual differences. | NFC studies, IBM prototype, privacy and opt-out discussion. | IBM result is nonsignificant; NFC is one trait, not a full account of access or cognition. | Boundary and personalization layer | | 14 | Use a Claude-centered AGI scenario to explore attention and control. | Explicitly speculative screenplay and design reflection. | Scenarios surface values but cannot establish likely behavior or safety. | Provocation | | 15 | State the social and political choice. Augmentation should leave people capable. | Labor history, ethical reasoning, recap, and call to action. | Firms may profit from dependence; designers cannot settle labor governance alone. | Normative conclusion | 🔶 **Reader response:** unavailable. No reactions, highlights, or annotations were supplied. The analysis records source structure and reviewer evaluation only. ### Argument spine 1. Conventional usability measures task success while assistance is present. 2. Some assistance can remove the reasoning or practice that sustains later capability. 3. Teams therefore need an opportunity model that keeps the human task visible. 4. The five assists identify candidate help across the task loop. 5. Candidate help must be filtered by the consequences of system error and missed detection. 6. Appropriate reliance, not maximum use or trust, becomes the target. 7. Human Goes First and cognitive forcing are proposed ways to protect judgment. 8. User differences, accessibility, stakes, and pace of change determine whether the friction helps. 9. The final measure is joint task quality plus the person’s capability, agency, and recovery after use. ### Weakest link The weakest link is the move from **credible risk and useful examples** to **universal framework and default pattern**. The book establishes that skill loss and overreliance can occur. It also presents a memorable five-part checklist and several promising interventions. It does not establish that five categories exhaust assistance, that the full Human Goes First sequence outperforms alternatives, or that a single user-trait response transfers across domains. Repairing the link requires comparative taxonomies, controlled local tests, delayed unaided measures, and representative accessibility testing. ### Evidence ledger | Evidence | Independent unit? | Supports | Quality and limits | Disposition | |---|---:|---|---|---| | IBM assisted-navigation prototype | 1 | C-001, C-002 | Tiny informal study; recordings and detailed data lost | Illustrative only | | IBM feedback-navigation prototype | 1 | C-003, C-006 | Same provenance and evidence limits | Illustrative only | | Five-assist cases across domains | Many examples, one author synthesis | C-007–C-010 | Breadth shows applicability, not exhaustiveness or causation | Strong ideation value | | Dell’Acqua et al., 758 consultants | 1 | C-011, C-024 | Preregistered experiment; one firm, model, and task set; inside/outside-frontier reversal | Strong, bounded | | Reverberi et al., 21 endoscopists / 504 videos | 1 | General overreliance risk | Controlled repeated-video setting; reports 1.85, not 6.4 | Correct C-013; retain narrower risk | | Buçinca et al., 199 participants | 1 | C-015, C-016, C-018 | Simulated AI and low-stakes nutrition task; forcing reduced but did not end overreliance | Good bounded support | | Gajos and Mamykina, three experiments | 1 research program | C-006, C-014 | Answer-first alone did not create learning; explanation without answer did | Complicates full pattern | | de Jong et al., 474 across two studies | 1 research program | C-017, C-018 | Partial explanations reduce bad and good reliance; no overall performance win over full explanations | Mixed | | Buçinca et al. adaptive policies, 1,422 across stages | 1 research program | C-018, C-019 | NFC matters, but learning result inconsistent and task artificial | Medium | | Bastani et al., nearly 1,000 students | 1 | C-001, C-006, C-022 | Preregistered field experiment; guarded tutoring removed most learning harm | Strong in education | | Shen and Tamkin, 52 developers | 1 | C-001, C-022 | Randomized, small, one-hour unfamiliar-library task; 2026 preprint | Promising, narrow | | Liu et al., 1,222 participants | 1 research program | C-001, C-022 | Three experiments; immediate post-removal effects; 2026 preprint | Promising, not long-term | | WCAG 2.2 and W3C COGA | 2 official sources | Access limits on C-014–C-019 | Normative standard plus informative user-needs guidance; not AI-accuracy evidence | Required design layer | | NIST AI RMF and GenAI Profile | 2 related official sources | C-011, C-012, C-024 | Risk guidance, not causal UI evidence; RMF revision underway | Governance support | | EU AI Act Articles 13–14 | 1 official legal text | Oversight and automation-bias context | Scope depends on territory, role, risk class, and transition dates | Current legal context, not legal advice | The ledger counts studies or research programs, not every reported effect, as independent evidence. The book’s internal examples share one authorial selection process and cannot be counted as independent confirmation of the framework. ## PROMPT 7 — Test, grade, and repair the claims ### Evaluation key - **A:** well supported in stated scope - **B:** credible with material limits - **C:** plausible but under-supported - **D:** weak or overstated - **F:** contradicted by the relevant source - **N/A:** normative or taxonomic claim that evidence can inform but not settle alone | ID | Grade | Evaluation, strongest objection, and repair | |---|---:|---| | C-001 | B | Multiple causal studies now show worse immediate unaided learning or performance after answer-giving help. Long-term workplace de-skilling remains unproven. **Repair:** say assistance *can* impair near-term skill formation and test durability locally. | | C-002 | D | The number comes from a tiny informal study whose records are gone. **Repair:** retain as design origin story, not effect-size evidence. | | C-003 | D | Same provenance problem, with no usable replication. **Repair:** frame feedback as a prototype hypothesis. | | C-004 | N/A | Persuasive value judgment, bounded by cases where no valuable capability is lost and automation removes drudgery or harm. **Repair:** state who bears each cost and benefit. | | C-005 | B | Attention and allocation of work clarify categories, but mixed-initiative systems cross them. **Repair:** classify each task phase rather than the whole product once. | | C-006 | B | Feedback, scaffolding, and explanation-without-answer can support learning. Mere “human first” ordering is insufficient. **Repair:** specify the learning mechanism and test transfer. | | C-007 | C | The five classes cover much design work, but the universal label lacks rival comparison. **Repair:** call them a five-lens opportunity checklist. | | C-008 | C | Previews can improve choices, but can also anchor, overload, or misstate uncertainty. **Repair:** test calibrated forecasts against no-preview and alternative presentations. | | C-009 | B | Reflection has broad learning support, but its timing, quality, and safety matter. **Repair:** name the feedback loop and delayed outcome. | | C-010 | B | The workshop is coherent and ready to run as ideation. It has no evidence of superior downstream outcomes. **Repair:** judge it by opportunity coverage and validated user value. | | C-011 | A | Appropriate reliance is supported by controlled studies and current risk guidance. **Limit:** operational thresholds remain domain-specific. | | C-012 | A | The prompt is a sound hazard screen and aligns with current governance. **Limit:** add affected people, recovery, misuse, and systemic harm. | | C-013 | F | The cited paper reports a 1.85 deterioration factor when AI was wrong; no 6.4 result appears. **Repair:** replace the number and note that hybrid performance improved overall. | | C-014 | C | Components have support, but the integrated four-part pattern has not been independently tested. Answer-first alone did not produce learning in one research program. **Repair:** test it against hints, explanation-only, AI-first review, and no AI. | | C-015 | B | Controlled work supports reduced overreliance, with burden and under-reliance tradeoffs. **Repair:** measure correct acceptance, wrong acceptance, time, and access together. | | C-016 | A | The cited study directly separates preference from overreliance. **Limit:** dislike does not prove that friction is justified. | | C-017 | B | Partial explanations reduced reliance on wrong answers but also on right answers and did not beat full explanations on total performance. **Repair:** use only where information withholding is safe and net outcomes improve. | | C-018 | B | NFC moderates results in several studies, but effect patterns vary. **Repair:** treat it as one candidate moderator, not a stable persona. | | C-019 | D | “Treat everyone as high NFC” risks cognitive and accessibility burden and lacks strong validation. **Repair:** offer transparent, adjustable support based on task risk and observed behavior, with opt-out. | | C-020 | B | Context plainly shapes product relationships, but some access needs and human limits are constraints, not preferences. **Repair:** distinguish preferences, capabilities, rights, and environmental constraints. | | C-021 | A | The chapter labels its scenario as speculation. **Limit:** readers may still import its assumptions into forecasts. | | C-022 | C | Strong design goal; recent studies show both preserved and harmed learning. Long-term transfer is sparse. **Repair:** make delayed unaided capability a product metric. | | C-023 | A | Labor, distribution, and institutional incentives can defeat narrow task success. This is well supported as socio-technical analysis, though no single metric settles it. | | C-024 | B | The synthesis follows directly from observed tradeoffs. **Repair:** select measures proportionate to the harm rather than impose a maximal battery on every feature. | ### Grade distribution | A | B | C | D | F | N/A | |---:|---:|---:|---:|---:|---:| | 5 | 10 | 4 | 3 | 1 | 1 | The 24 dispositions match the 24 unique claims. The grade distribution is descriptive, not a score for the book. ### Adversarial checks - **Selection bias:** memorable failures and product examples may overrepresent visible outcomes. The book does preserve positive human–AI collaboration results, which limits but does not remove this concern. - **Survivorship:** successful design cases rarely reveal teams, users, or contexts where the pattern failed. - **Construct drift:** “de-skilling,” “reduced critical-thinking effort,” weak immediate learning, and lower later performance are related but not interchangeable. - **False binary:** assistant versus agent is a spectrum across a workflow, not always a product-level either/or. - **Universalizing from plausibility:** a taxonomy can organize thought without proving that every case fits or that no missing category exists. - **Metric substitution:** satisfaction, productivity, assisted accuracy, unaided retention, and appropriate reliance answer different questions. - **Automation-bias asymmetry:** reducing wrong acceptance is not enough if the intervention drives costly rejection of correct support. - **Individual-difference reduction:** NFC does not represent disability, literacy, language, expertise, confidence, culture, fatigue, or power. - **Appeal to friction:** effort is not inherently educational. It must cause useful cognition and remain accessible. ### Strongest fair version of the argument Assistant design should begin with the human task rather than the model’s capabilities. The five assists offer a compact way to find opportunities. Each opportunity should then be screened for error consequences, retained-skill needs, urgency, reversibility, affected people, and the user’s ability to detect failure. Where independent reasoning matters, teams should test patterns that preserve meaningful cognitive work—such as hints, source-linked explanations, separate attempts, comparisons, or targeted feedback—against lower-friction alternatives. Success requires evidence about joint performance, bad-advice acceptance, good-advice acceptance, later unaided capability, burden, accessibility, and recovery. The exact pattern is local; the obligation to test the human outcome is general. ## PROMPT 8 — Place the book in context ### Intellectual lineage | Lineage | What the book takes | What it changes or underplays | |---|---|---| | Human-centered and user-centered design | Start from human goals, tasks, and outcomes | Adds retained capability as an outcome but gives limited power analysis in the method itself | | Norman-style action and feedback loops | Perception, planning, action, evaluation | Recasts loop stages as places for assistance | | Automation and human factors | Automation bias, trust calibration, supervisory control | Uses “appropriate reliance” but needs more operational human-factors detail for high-risk domains | | Intelligence augmentation | Computers should extend human intellect | Makes “more capable after use” the explicit endpoint | | Learning science and metacognition | Practice, feedback, reflection, cognitive effort | Sometimes treats friction as a proxy for learning despite evidence that sequence and content matter | | Explainable AI and decision support | Explanations affect reliance | Correctly shows that explanation quality and timing matter more than mere presence | | Critical and socio-technical design | Products redistribute labor, agency, and power | Raises these issues late; workshop mechanics don't fully center affected nonusers | | Accessibility and assistive technology | Protects the established meaning of assistive technology | Product accessibility is not systematically integrated into every assist or forcing pattern | ### Before, during, and after publication - **Before:** Research on automation bias, explanations, cognitive forcing, and human–AI complementarity gives the book its empirical base. Noessel’s own *Designing Agentive Technology* supplies the agent comparison. - **During:** Rapid generative-AI adoption makes the assistant/agent boundary and skill-retention question concrete. The manuscript draws on workplace and medical studies whose scope differs sharply. - **After or near release:** The peer-reviewed “jagged frontier” paper preserves its central two-sided result: large gains inside a tested model boundary and worse accuracy outside it. 2025–2026 education, coding, and reasoning studies add causal evidence for immediate post-assistance costs, but still do not prove years-long workplace de-skilling. - **Current governance:** WCAG 2.2 and W3C cognitive-accessibility guidance add testable access and recovery requirements. NIST names overreliance, automation bias, anthropomorphism, oversight, and representative testing. EU AI Act Article 14 requires effective human oversight and awareness of automation bias for covered high-risk systems. ```mermaid flowchart LR A[Human-centered design] --> B[Five assists] C[Human factors and automation bias] --> D[Appropriate reliance] E[Learning science] --> F[Reflection and retained skill] G[Explainable AI research] --> H[Cognitive forcing] B --> I[Designing Assistant Technology] D --> I F --> I H --> I I --> J[Local task and harm tests] K[WCAG and COGA] --> J L[NIST and EU oversight] --> J M[2025-2026 learning studies] -. post-book refinement .-> J ``` ### Where the book sits The book is a bridge text between product ideation and human-AI safety. It is more concrete than a general ethics book and less evidentially complete than a human-factors or safety standard. Its distinctive move is to pair an opportunity taxonomy with a reliance filter. Its best readers are product designers, researchers, managers, and technologists who can run local studies rather than copy patterns directly. ## PROMPT 9 — Audit inclusion, power, and accessibility ### Inclusion audit | Dimension | What the book does well | Gap or risk | Repair | |---|---|---|---| | Disability language | Explicitly protects **assistive technology** as a disability-centered term | Disability is not a consistent design axis in the five assists | Add disability-led research and access criteria to every workshop stage | | EPUB access | Rich navigation and detailed alternative text for most figures | One meaningful image is hidden as decorative; table semantics and footnotes are weak | Fix `pg79.jpg` alt, header/caption markup, and note relationships; run EPUBCheck | | Cognitive access | Recognizes different preferences and cognitive effort | NFC narrows cognitive difference; forced effort can exclude | Test memory, timing, language, fatigue, and executive-function demands; make support adjustable | | Sensory and motor access | Some diagrams are designed with color-blindness in mind | Assistant interaction patterns rarely discuss keyboard, screen reader, speech, switch, or low-vision operation | Apply WCAG 2.2 and test status, focus, controls, timing, and alternatives | | Language and culture | Plain definitions and many familiar examples | Examples lean toward English-language, Global North commercial products | Research local mental models, translation, uncertainty language, and collective decision norms | | Race, gender, class, religion | Labor and power appear in the closing argument | These identities and structural effects receive little sustained analysis | Include affected communities, disparate error costs, access to recourse, and labor conditions | | Expertise | Novice and expert needs appear through skill-retention questions | Expertise can be treated as an individual property rather than institutional access | Study credentialing, team practice, tacit knowledge, and who gets training | | Privacy | Chapter 13 flags privacy and opt-out concerns in adaptation | Cognitive profiling may still invite hidden inference | Minimize data, explain adaptation, require consent, permit override, and set deletion rules | | Affected nonusers | Risk questions imply wider consequences | Workshop centers the direct user and task | Add people acted upon, downstream reviewers, communities, and those unable to contest output | | Labor | Final chapter makes automation a political and organizational choice | Product teams alone may lack authority to protect jobs or workload | Include workers, unions, procurement, policy, and incentives in governance | ### Accessibility acceptance criteria for an assistant 1. Every function works by keyboard and exposes programmatic name, role, state, and value. 2. Generated status and completion messages reach assistive technology without unexpected focus movement. 3. The user can inspect, correct, reverse, or stop consequential actions. 4. A cognitive-forcing step has an equivalent accessible route and does not depend on speed, memory, color, drag, speech, or fine motor input alone. 5. Context and prior inputs remain visible or recoverable; the flow does not force re-entry. 6. Voice interfaces allow pauses, correction, and simple responses. 7. Human help and recourse are easy to find through an accessible channel. 8. Tests include disabled users and measure both completion and reliance outcomes. 9. Personalization is transparent, optional, minimal, and reversible. 10. Error and burden results are broken down by relevant user groups without turning identity into a hidden risk score. ### Power questions the workshop should add - Who gains time, money, discretion, or status from this assistance? - Who supplies the hidden review, labeling, correction, or emotional labor? - Who is affected by the output but cannot see or contest it? - Can a worker refuse the assistant without penalty? - Who decides which skill is no longer worth preserving? - Does “efficiency” increase workload, surveillance, or pace expectations? - Who owns the interaction record and inferred cognitive profile? - What happens when the institution benefits from user dependence? ### Bottom line The book's inclusion posture is thoughtful but incomplete. It respects disability terminology, notices user differences, acknowledges privacy, and ends with a clear labor-and-power argument. Yet its central methods do not make accessibility or affected nonusers structurally mandatory. A team could run the workshop faithfully and still produce a cognitively demanding, screen-reader-hostile, or institution-centered assistant. The repairs above should be treated as part of the method, not a final compliance pass. ## PROMPT 10 — Distill notes without creating adjacent files This prompt is adapted to the one-file contract. The candidates stay embedded here. No atomic notes, MOC edits, aliases, or reciprocal links were created. ### Fleeting notes - The best distinction in the book may be product success versus human success. - The five assists and appropriate reliance form a useful generate-then-filter sequence. - Product friction needs a causal story: which cognition happens because of it? - A separate AI answer may reduce anchoring, but comparison design can still steer the user. - “Human goes first” sounds ethical but can become a ritual confirmation screen. - Keep correct-use and incorrect-use outcomes separate. - Test capability after the assistant is removed. - Accessibility changes whether a forcing function is productive or exclusionary. - AI provenance belongs in the artifact, not in a private production note. - The strongest claim is an evaluation question: what remains in the person? ### Literature notes 1. **Assistance is a task relationship, not a model class.** Noessel defines an assistant by how work and attention are distributed. One product may act as tool, assistant, automation, or agent at different moments. (`chapter3.xhtml#pg_26`) 2. **The five assists are an opportunity map.** Perceive, know, plan, perform, and reflect cover recurring moments around action. The model's practical value does not depend on proven universality. (Chs. 4–10) 3. **Usability can hide capability loss.** Immediate task success says little about later unaided work. The book's opening IBM numbers are low-confidence, but newer independent studies support the narrower risk. (Chs. 1–2; S8–S10) 4. **Appropriate reliance is bidirectional.** A good assistant helps people accept useful advice and reject harmful advice. Reducing acceptance alone can create under-reliance. (`chapter11.xhtml#pg_121`; S2–S5) 5. **Human Goes First is a design hypothesis.** The pattern preserves an independent attempt, comparison, and feedback, but no source validates the full sequence as a default. Answer-first alone did not produce learning in Gajos and Mamykina. (Ch. 11; S4) 6. **Friction has outcome and access costs.** Cognitive forcing can reduce wrong acceptance while increasing effort, dislike, and rejection of correct advice. (Ch. 12; S3, S5) 7. **NFC is one moderator.** It cannot stand in for expertise, disability, literacy, culture, confidence, fatigue, or preference. (Ch. 13; S3, S5, S6) 8. **Augmentation needs a post-use test.** Noessel's endpoint is a person who remains more capable after the assistant is removed. (`chapter15.xhtml#pg_178`) 9. **Current oversight duties are behavioral.** NIST and the EU AI Act focus on understanding limits, monitoring, override, interruption, representative testing, and automation bias—not the ceremonial presence of a person. (S13–S15) 10. **Provenance is part of design quality.** The book's AI colophon makes tool involvement inspectable and invites readers to judge how generated material entered the work. (Appendix B) ### Permanent-note candidates #### Assistance quality includes after-use capability **Candidate claim:** Evaluate an assistant both while it is present and after it is removed. Assisted speed or accuracy can coexist with weaker independent performance, so product quality includes retention, transfer, persistence, and recovery. - Confidence: medium-high for short-term learning tasks; medium for general work - Scope: tasks where human capability remains valuable - Support: C-001, C-006, C-022; S4, S8–S10 - Boundary: some dangerous, inaccessible, or low-value tasks should be automated rather than preserved - Proposed links: `[[Human-centered AI]]`, `[[Learning transfer]]`, `[[Appropriate reliance]]` - Do not attribute to reader; system-derived candidate #### Productive friction should follow risk and skill durability **Candidate claim:** Add cognitive friction only when missed errors matter, the skill should persist, the delay is acceptable, and an accessible route exists. The burden must trigger useful reasoning rather than ceremony. - Confidence: medium - Scope: decision support and learning-sensitive workflows - Support: C-012, C-014–C-019; S3–S6, S11–S12 - Boundary: urgency, cognitive access, and trivial skill value can favor direct assistance - Proposed links: `[[Cognitive forcing functions]]`, `[[Accessible interaction design]]`, `[[Risk-based design]]` - Do not attribute to reader; system-derived candidate #### Human-first workflows need outcome proof **Candidate claim:** A human-first sequence is not safe because the human clicks first. It is safe only if evidence shows better detection, calibrated use, learning, or recovery without unacceptable delay and burden. - Confidence: medium-high - Scope: AI-assisted judgment - Support: C-011, C-014, C-024; S2–S5, S13–S15 - Boundary: the best intervention may be hints, explanation without a final answer, expert escalation, or automation with controls - Proposed links: `[[Meaningful human oversight]]`, `[[Automation bias]]`, `[[Human Goes First]]` - Do not attribute to reader; system-derived candidate #### Preference and performance are separate metrics **Candidate claim:** A preferred assistant can cause more wrong reliance, while an effective safeguard can be disliked. Teams should measure satisfaction, good-advice use, bad-advice rejection, task quality, time, and access separately. - Confidence: high in the tested CFF setting; medium-high as an evaluation rule - Scope: assistance with fallible recommendations - Support: C-015–C-017, C-024; S3, S5 - Boundary: poor preference may reveal real burden and cannot be dismissed as resistance - Proposed links: `[[Product metrics]]`, `[[Trust calibration]]`, `[[Under-reliance]]` - Do not attribute to reader; system-derived candidate #### An AI colophon turns provenance into a design constraint **Candidate claim:** Recording where AI contributed makes provenance reviewable and pressures creators to decide which work requires human judgment, source checking, or disclosure. - Confidence: medium - Scope: published knowledge artifacts - Support: Appendix B and the book's own production disclosure - Boundary: disclosure does not prove accuracy, authorship quality, consent, or responsible use - Proposed links: `[[AI disclosure]]`, `[[Knowledge provenance]]`, `[[Editorial review]]` - Do not attribute to reader; system-derived candidate ## PROMPT 11 — Build a teaching and discussion pack ### Three-session sequence #### Session 1: From AI feature to human task **Goal:** Distinguish tool, automation, assistant, and agent; apply the five assists without assuming they are exhaustive. **Learning objectives** - Classify a product by task phase and allocation of work. - Map perceive, know, plan, perform, and reflect onto one user task. - Identify at least one affected person who is not the direct user. - Separate an opportunity hypothesis from validated benefit. **Preparation:** Chapters 3–5 and 10; vocabulary table above. **Exercise:** Choose a familiar workflow. Draw the human task before adding AI. Generate one assist in each category, then add a sixth “does not fit” column. Keep any item that resists the taxonomy rather than forcing it into a box. **Discussion questions** 1. At which step does the same product shift from assistant to agent? 2. Which proposed assist changes what the user can notice? 3. What disappears from the task when the assist succeeds? 4. Whose work or risk remains invisible? **Evidence of learning:** A task map with category reasoning, one counterexample, and one validation question per selected assist. #### Session 2: Appropriate reliance under failure **Goal:** Design and test for both overreliance and under-reliance. **Learning objectives** - Explain why trust and appropriate reliance differ. - Create correct, incorrect, uncertain, and unavailable-system test cases. - Compare Human Goes First with at least two other interventions. - Define behavioral measures for detection, override, and recovery. **Preparation:** Chapters 11–13; external sources S2–S6. **Case:** A writing assistant suggests accessibility acceptance criteria for a public-service form. Some advice is correct, one criterion cites an invented standard, and one correct suggestion creates an unreasonable cognitive burden. **Exercise:** Prototype three flows: AI answer first, human attempt then comparison, and explanation/hints without a final answer. Run a paper test in which participants do not know which advice is wrong. **Discussion questions** 1. What would count as harmful overreliance? 2. What would count as harmful under-reliance? 3. Did the human-first flow create real analysis or only delay? 4. How did confidence, citations, and interface polish change behavior? **Evidence of learning:** A measurement plan covering correct acceptance, wrong acceptance, time, confidence, accessibility, and recovery. #### Session 3: Capability, access, and power after launch **Goal:** Evaluate what the assistant leaves in the person and institution. **Learning objectives** - Design an unaided post-use measure. - Apply WCAG and cognitive-accessibility requirements to a forcing pattern. - Identify labor, privacy, and recourse effects. - Set a stop rule and review date for a product experiment. **Preparation:** Chapters 1–2, 9, 15; external sources S8–S15. **Exercise:** Return to the Session 1 workflow. Add a delayed no-AI test, keyboard and screen-reader acceptance criteria, a recovery path, an affected-party review, and a data-minimization decision. **Discussion questions** 1. Which capability should still exist after a month? 2. Is the forcing pattern accessible under fatigue, time pressure, and assistive technology? 3. Who benefits if users become dependent? 4. Who can stop, appeal, or reverse the system? **Evidence of learning:** A one-page launch gate with measures, group breakdowns, a failure threshold, and named decision owner. ### Facilitation notes - Do not demonstrate only polished success. Seed plausible errors and unavailable-system cases. - Give all materials in accessible digital form before the session. - Do not require speech, dragging, color discrimination, or timed reading as the only participation route. - Let participants opt out of disclosing disability, expertise, or cognitive-style information. - Separate critique of a design from judgment of the person who accepted or rejected AI advice. - Close by revisiting the five-assist map: which ideas survived risk, evidence, and access review? ### Short assessment rubric | Criterion | Meets | Needs revision | |---|---|---| | Task model | Human task and category shifts are explicit | Starts from an AI feature or treats the product as one fixed category | | Evidence | Claim scope and source limits are stated | Effect sizes or examples are copied without context | | Reliance | Correct and incorrect advice are both tested | Only overall task accuracy is measured | | Capability | Includes delayed unaided outcome where relevant | Measures only assisted completion | | Accessibility | Has concrete interaction and representative-test criteria | Says “accessible” without acceptance criteria | | Power | Names affected people, control, recourse, and data use | Considers only the direct user's convenience | ## PROMPT 12 — Convert the reading into action ### Recommended actions | Action | PARA home | Next step | Measure | Duration and stop rule | Review | |---|---|---|---|---|---| | Add a five-assist plus affected-party pass to one design workshop | Resources until tied to an active project | Map a real task and generate candidates across all five assists plus “does not fit” | New useful hypotheses; missed task stages; participant burden | One 60-minute workshop; stop if categories cause forced or duplicate ideas | 2026-09-15 | | Add appropriate-reliance cases to one assistant prototype | Active project only when a named product exists | Create correct, wrong, uncertain, and unavailable outputs | Correct acceptance, wrong acceptance, override, recovery, time | 5–8 formative sessions; stop if any seeded consequential error cannot be detected or reversed | 2026-09-30 | | Test one human-first pattern against a lower-friction alternative | Active project only when a named task exists | Compare attempt-then-feedback with hints or source-linked explanation | Joint quality, later unaided result, abandonment, access issues | Two prototype rounds; stop if no benefit or a group faces material access cost | 2026-10-15 | | Add accessibility gates to cognitive forcing | Resources / design checklist | Translate WCAG 2.2 and COGA needs into acceptance criteria | Keyboard completion, semantic announcements, correction, memory load, equivalent routes | One audit plus representative test; stop launch on a critical blocker | 2026-09-30 | | Preserve evidence provenance in design decisions | Resources / research practice | Record claim ID, source type, scope, and falsifier for consequential assumptions | Unsupported effect sizes caught before reuse | Apply for one decision cycle; stop any external claim lacking a checked source | 2026-09-15 | | Measure after-use capability where skill matters | Active project | Add a delayed no-assistant task | Retention, transfer, persistence, recovery | Pilot with a short delay, then set domain-appropriate interval; stop if the measure cannot represent real capability | 2026-10-31 | ### Most likely misuse The most likely misuse is turning a nuanced warning into a universal mandate: **force every user to work first because friction is good**. That reading would reproduce the book's own error target. It would substitute a pattern for evidence, ignore urgency and low-value labor, increase cognitive and accessibility burdens, and confuse ceremonial human action with meaningful oversight. Prevent it with four gates: 1. Name the capability worth preserving and who values it. 2. Name the harm if the system is wrong and the user misses it. 3. Compare at least one lower-burden intervention. 4. Test behavior and access with representative users, including wrong-output cases. ### Decision rules - **Use direct automation** when the task is dangerous, inaccessible, valueless to retain, reliably specified, reversible, and monitored. - **Use assistant-style support** when human judgment, learning, accountability, relationship, or adaptation remains central. - **Use cognitive forcing** when missed errors matter and independent reasoning improves detection or retention. - **Remove or reduce forcing** when it produces no measured benefit, drives under-reliance, blocks access, or creates unsafe delay. - **Escalate to a person or institution** when the system cannot provide adequate evidence, recourse, authority, or care. ### Archive decision This consolidated review belongs in `4. Archives/Book reviews` as a completed reference. The methods and candidate permanent notes remain proposals inside this file until a real project or later synthesis authorizes separate notes. No project is created from the book alone. ## PROMPT 13 — Build the visual models ### Visual 1: Five assists around a human task **Purpose:** Show that the taxonomy surrounds a person-led task rather than a model-led feature list. ```mermaid flowchart LR P[Perceive] --> T[Human task] K[Know] --> T L[Plan] --> T T --> A[Perform] A --> R[Reflect] R --> P X[Context, access, and affected people] --- T Y[System evidence and uncertainty] --- T ``` **Nodes:** five assist classes, human task, context/access/affected people, system evidence/uncertainty. **Connections:** assistance enters before or during action; reflection feeds later perception. **Groups:** task loop and cross-cutting constraints. **Labels:** verbs keep the human activity visible. **Uncertainty:** the diagram does not prove the classes are exhaustive. **Source:** C-005–C-010. **Layout:** left-to-right cycle with cross-cutting context below. **Prose alternative:** Perception, knowledge, and planning feed a human task; performance carries it into action; reflection returns learning to the next round. Context, accessibility, affected people, system evidence, and uncertainty constrain the whole loop. ### Visual 2: Argument and evidence dependency map **Purpose:** Separate the book's strong normative chain from its weak empirical generalization. ```mermaid flowchart TD U[Usability measures assisted success] --> G[Capability after use may differ] G --> O[Generate opportunities with five assists] O --> H[Screen missed-error harm and skill value] H --> R[Aim for appropriate reliance] R --> I[Test assistance interventions] I --> M[Measure joint quality, reliance, retention, burden, access, recovery] E1[Small IBM prototypes] -. weak support .-> G E2[Recent learning studies] --> G E3[Human-AI and CFF studies] --> R E4[WCAG, COGA, NIST, EU law] --> M O -. universality not established .-> H ``` **Nodes:** core argument steps and four evidence families. **Connections:** solid arrows show the defensible reasoning chain; dashed arrows show weak or unproven dependencies. **Groups:** book method, external empirical evidence, current standards. **Labels:** evidence strength appears on weak links. **Uncertainty:** long-term de-skilling and taxonomy exhaustiveness remain open. **Source:** C-001, C-007, C-011–C-024; S1–S15. **Layout:** top-down argument with side evidence. **Prose alternative:** Usability establishes assisted success, not retained capability. The five assists generate options; harm and skill value filter them; appropriate reliance defines the target; local tests determine the intervention. Small IBM studies weakly support the opening risk, while newer learning studies strengthen a narrower short-term claim. Standards expand the required measures. ### Visual 3: Reliance-intervention decision tree **Purpose:** Prevent automatic use of Human Goes First or cognitive forcing. ```mermaid flowchart TD A[Can an AI error cause material harm?] -->|No| B[Prefer low-burden help and ordinary quality controls] A -->|Yes| C[Can the user reasonably detect and recover?] C -->|No| D[Reduce autonomy, add expert control, or do not deploy] C -->|Yes| E[Should human skill remain durable?] E -->|No| F[Use review, override, monitoring, and recourse] E -->|Yes| G[Compare interventions] G --> H[Human attempt then comparison] G --> I[Hints or explanation without final answer] G --> J[On-demand or delayed advice] H --> K[Test correct and wrong advice] I --> K J --> K K --> L{Net benefit across quality, retention, burden, and access?} L -->|Yes| M[Deploy narrowly and monitor] L -->|No| N[Revise, choose lower burden, or stop] ``` **Nodes:** harm, detectability, recovery, skill durability, intervention choices, measures, deployment decision. **Connections:** each branch narrows a design response. **Groups:** risk screen, pattern comparison, evaluation. **Labels:** yes/no branches state the decision. **Uncertainty:** thresholds and time horizons are domain-specific. **Source:** C-011–C-019, C-022–C-024; S2–S15. **Layout:** top-down decision tree. **Prose alternative:** Low-harm tasks receive ordinary controls. High-harm tasks that users cannot detect or recover from require less autonomy, expert control, or no deployment. When detection is possible and skill matters, compare several reasoning-preserving patterns rather than selecting Human Goes First by name. Deploy only after net benefit across quality, reliance, retention, burden, and access. ### Visual 4: Product evaluation scorecard **Purpose:** Keep one attractive metric from hiding a harmful tradeoff. ```mermaid flowchart LR P[Assistant prototype] --> Q[Joint task quality] P --> CR[Correct-advice reliance] P --> WR[Wrong-advice rejection] P --> U[Unaided retention and transfer] P --> B[Time and cognitive burden] P --> AC[Accessibility and group equity] P --> RC[Recovery, override, and recourse] Q --> D[Decision] CR --> D WR --> D U --> D B --> D AC --> D RC --> D ``` **Nodes:** prototype, seven outcome families, decision. **Connections:** every outcome feeds the decision; none substitutes for the others. **Groups:** performance, reliance, durability, burden/access, control. **Labels:** observable outcome families. **Uncertainty:** weights must come from domain harms and affected people. **Source:** C-011, C-015–C-024; S1–S15. **Layout:** fan-out scorecard. **Prose alternative:** Evaluate the same prototype for joint quality, use of correct advice, rejection of wrong advice, later unaided capability, burden, accessibility and group equity, and recovery or recourse. A launch decision must reconcile all seven with weights tied to actual harms. ### Visual coverage decision | Requested visual type considered | Decision | Reason | |---|---|---| | Concept map | Included as Visual 1 | Clarifies the five-assist relationship | | Argument or dependency map | Included as Visual 2 | Makes the weak empirical step visible | | Decision tree | Included as Visual 3 | Prevents pattern-by-default use | | Evidence or measurement map | Included as Visual 4 | Keeps distinct outcomes from collapsing | | Timeline | Omitted | The sequence of publication and studies is simple enough in prose | | Matrix | Omitted as a diagram | The chapter, claim, evidence, inclusion, and action tables already serve this form | | Stakeholder map | Omitted | Power questions are actionable here, but a real map needs a named product and affected groups | | Process swimlane | Omitted | A specific team and workflow were not supplied | ## Sources and verification ### Primary book and publisher record - Christopher Noessel, *Designing Assistant Technology: How to Build Products That Help People Think* (Rosenfeld Media, 2026). Local EPUB, SHA-256 `6697dd6616528ce9dfbcaff54af6fa2f8d1e0ccdb3454aba65410fa692721878`. - Rosenfeld Media, [official book page](https://rosenfeldmedia.com/books/designing-assistant-technology/). ### External evidence 1. Fabrizio Dell’Acqua et al., [“Navigating the Jagged Technological Frontier”](https://doi.org/10.1287/orsc.2025.21838), *Organization Science* (2026). 2. Carlo Reverberi et al., [“Experimental Evidence of Effective Human–AI Collaboration in Medical Decision-Making”](https://doi.org/10.1038/s41598-022-18751-2), *Scientific Reports* (2022). 3. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos, [“To Trust or to Think”](https://doi.org/10.1145/3449287), PACM HCI (2021). 4. Krzysztof Z. Gajos and Lena Mamykina, [“Do People Engage Cognitively with AI?”](https://doi.org/10.1145/3490099.3511138), IUI (2022). 5. Sander de Jong et al., [“Cognitive Forcing for Better Decision-Making”](https://doi.org/10.1145/3710946), PACM HCI (2025). 6. Zana Buçinca et al., [“Towards Optimizing Human-Centric Objectives in AI-Assisted Decision-Making With Offline Reinforcement Learning”](https://arxiv.org/html/2403.05911v2) (2024). 7. Hao-Ping Lee et al., [“The Impact of Generative AI on Critical Thinking”](https://doi.org/10.1145/3706598.3713778), CHI (2025). 8. Hamsa Bastani et al., [“Generative AI without Guardrails Can Harm Learning”](https://doi.org/10.1073/pnas.2422633122), *PNAS* (2025), with [correction](https://doi.org/10.1073/pnas.2518204122). 9. Judy Hanwen Shen and Alex Tamkin, [“How AI Impacts Skill Formation”](https://arxiv.org/html/2601.20245v2) (2026 preprint). 10. Grace Liu et al., [“AI Assistance Reduces Persistence and Hurts Independent Performance”](https://arxiv.org/html/2604.04721v4) (2026 preprint). 11. W3C, [Web Content Accessibility Guidelines 2.2](https://www.w3.org/TR/WCAG22/). 12. W3C COGA Task Force, [Making Content Usable for People with Cognitive and Learning Disabilities](https://www.w3.org/TR/coga-usable/). 13. NIST, [Artificial Intelligence Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1) (2023). 14. NIST, [Generative Artificial Intelligence Profile](https://doi.org/10.6028/NIST.AI.600-1) (2024). 15. European Union, [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689), especially Articles 13–14. ### Evidence limits and drift watch - EPUBCheck was not installed; package claims of full standards conformance remain unverified. - No matching Readwise book or highlights were available, so no reader emphasis is inferred. - No independent source validates the integrated four-part Human Goes First pattern. - Current evidence supports immediate or short-term learning and performance effects; long-term cross-domain workplace de-skilling remains unproven. - The NIST AI RMF is under revision. The 2026 preprints may change after peer review. Model capability boundaries must be retested on the deployed model. - EU AI Act duties depend on territorial scope, system classification, role, and transition date; this review is not legal advice. ## Session completion check | Check | Result | |---|---| | Prompts 0–13 | Executed and present exactly once | | Source coverage | Full EPUB spine; all numbered chapters and appendices read | | Attribution | EA, IA, Q3P, and UI lanes used; no reader emphasis invented | | Unknown inputs | Reader purpose, prior beliefs, and reactions marked 🔶 | | Claims | Stable C-001–C-024 inventory with scope, evidence, falsifier, and evaluation | | External checking | 15 primary or official sources; book's 6.4 claim corrected | | Prompt 10 | Embedded candidates only; no adjacent notes or reciprocal links created | | Prompt 13 | Four Mermaid diagrams plus prose alternatives; omissions disclosed | | Reader publication | Not performed; no approval requested or implied | | Extra files in vault | None created | | Remaining work | None required for this reading contract | ### Final assumption and contradiction audit - **Assumption:** practical product-design and teaching use, because the reader purpose was not supplied. Marked 🔶 and kept visible. - **Assumption:** all selected excerpts and hotspots are system-selected because no reader highlights exist. Disclosed in frontmatter and Prompt 0. - **Contradiction corrected:** “6.4 times worse” is not supported by the cited medical paper; the paper reports 1.85 deterioration when AI advice was wrong and better hybrid performance overall. - **Tension preserved:** the book argues for friction while acknowledging preference and individual differences; this note adds accessibility as a required decision factor. - **Tension preserved:** the five assists are called universal, while the source provides a design synthesis rather than an exhaustiveness test. - **Evidence boundary:** recent causal studies strengthen the short-term skill-formation concern but do not establish lasting workplace de-skilling. - **Edition boundary:** metadata and publisher checks identify this as the 2026 Rosenfeld Media book, not the 2017 agentive-technology title. - **Omission check:** no requested prompt was skipped; no authorized publishing or link-editing action was inferred.