English
Human-computer interaction studies how people use computing systems to accomplish tasks, and how those systems should be designed and evaluated. It includes the interface, the user’s knowledge and abilities, the surrounding activity, and the consequences of an action. A booking form can be technically correct while leaving people unsure whether a reservation exists or how to change it.
Start with a task and the information needed to perform it. Users must discover available actions, predict their effects, carry them out, and interpret the result. A design can fail at any of these steps. Making a button more visible will not fix an unclear refund policy; explaining the policy will not fix a control that cannot be reached with a keyboard.
HCI connects implementation to psychology, design, ergonomics, and the study of social activity. Research methods help distinguish a designer’s assumption from observed behavior. Visual appearance matters, but so do learnability, task success, error recovery, accessibility, and effects on people who did not choose the system.
Prerequisites: Basic programming (§2.1) helps when building interactive prototypes. You can begin user research and paper prototyping before mastering a software framework. Probability and statistics (§2.5) become important for quantitative comparisons.
From Batch Processing to Designing for Billions
Batch processing made turnaround time a major constraint: users often submitted work and inspected results later. Yet early computing also included interactive control, displays, and direct operation. Time sharing expanded access to interactive sessions; it did not invent all human-computer interaction. The important change was that repeated feedback became available within an ongoing task, allowing people to explore and revise their work.
Sutherland’s Sketchpad demonstrated direct manipulation of represented objects. Engelbart’s NLS and its 1968 demonstration combined pointing, structured documents, links, and collaboration in a system intended to augment intellectual work. These systems asked what activities the computer should support, not only how to make an existing command easier to enter.
Xerox PARC explored personal workstations, bitmapped displays, pointing, and graphical applications. The commercial Xerox Star appeared in 1981, before Apple’s Lisa in 1983 and Macintosh in 1984. These products developed related ideas under different cost and market constraints. Shared conventions for selecting, copying, opening, and saving reduced how much users had to relearn between applications. A graphical interface still required training and did not make every task self-explanatory.
Card, Moran, and Newell’s 1983 work connected task performance with cognitive models. GOMS described goals, operators, methods, and selection rules, allowing predictions for specified tasks and assumptions about skilled behavior. Shneiderman’s interface principles and Nielsen’s usability engineering provided practical evaluation methods. Their value lies in making designs inspectable and testable; neither a model nor a checklist substitutes for observing the relevant users.
Norman’s work clarified conceptual models, feedback, mappings, affordances, and signifiers. An affordance is a possible action in the relation between an actor and an object; a signifier communicates where or how to act. A handle may permit pulling while a label communicates whether pulling is the intended action. This distinction helps explain why an operation can be physically possible yet difficult to discover.
The Web extended design across devices, languages, connection speeds, and independently authored pages. Touch interfaces changed target sizes, available feedback, and the visibility of gestures. Mobile use also introduced interruption and changing physical contexts. Accessibility and inclusive design made it harder to treat one familiar user, device, or sensory ability as the default for everyone.
AI-assisted interfaces add uncertainty about what a system will produce and what actions it may take. Some questions are familiar—status visibility, correction, delegation, and recovery—while model-generated errors add new forms of difficulty. A useful design identifies the evidence available to the user, when review is possible, and how an incorrect action can be corrected. Conversational fluency alone does not answer those questions.
Cognition, Evaluation, and Designing for Diversity
How Human Cognition Shapes Interface Requirements
Working-memory demands depend on the task, familiarity, grouping, and available external cues. Miller’s discussion of seven items is not a rule limiting menus to seven choices, and later chunk estimates are not fixed interface capacities. Keep relevant information visible when people need it: show the previous selection, the amount being confirmed, or the steps already completed. Recognition can reduce recall demands, while a stable command language can be efficient for practiced users.
A mental model is the user’s account of how a system behaves. Consider an editor that displays “Saved” immediately after a network request starts. If the request fails, the interface has taught the wrong relation between feedback and persistence. Showing saving, confirmed save, and a recoverable failure helps users predict what survives a reload. Feedback should correspond to the state that matters for the task.
Mappings connect controls to effects. Consistent labels, grouping, and spatial arrangements can make those mappings easier to learn. Signifiers make available actions apparent, but visual appearance must be backed by interaction semantics: a control should have a meaningful name and a usable keyboard behavior. More minimal visual styling is not automatically less cognitive work if it hides distinctions people need.
Pointing has measurable costs. One common Fitts’s-law formulation is T = a + b log₂(1 + D/W), where D is movement distance, W is target width along the movement direction, and a,b are fitted for the task and device. For D=400, increasing W from 20 to 80 changes the index from about 4.39 to 2.58. This predicts a direction of improvement under the model, not a universal time saving or a complete account of touchscreen use.
Separate slips from mistakes. A slip executes an unintended action despite a correct goal; a mistake follows an incorrect understanding or plan. Spacing controls, constraining invalid inputs, and preserving undo can reduce the cost of slips. Clear explanations and better task structure can address mistaken understanding. Repeated confirmation dialogs can become habitual clicks, so use confirmations where their information helps a meaningful decision and provide recovery where possible.
Empirical Evaluation Methods
Choose a method for the question. A formative usability study observes representative people attempting realistic tasks to discover obstacles. Think-aloud can expose an interpretation, but speaking can change behavior and timing. Interviews investigate experience and reported reasons; analytics records selected behavior. Neither automatically reveals every cause.
Small iterative studies can uncover useful problems quickly. The familiar five-user recommendation is not a coverage guarantee. Under a simple independent-discovery model, a problem found by each participant with probability p is found by at least one of n participants with probability 1−(1−p)^n. With p=0.31 and five participants this is about 84%; with p=0.05 it is only about 23%. Real probabilities differ across problems, tasks, and user groups. Nielsen’s account explicitly discusses comparable users and iterative redesign.
Heuristic evaluation provides an expert inspection before or alongside user studies. Its findings are candidate problems, not measurements of how often users will fail. Test task completion, errors, recovery, and satisfaction with appropriate participants; measuring small differences in completion rates or time requires a study sized for that purpose, not automatically five people.
A randomized A/B test compares outcomes under two assigned alternatives. Define the assignment unit, primary outcome, exposure, observation period, and stopping rule. Randomization supports a causal comparison when implementation and analysis preserve it; interference between users or selective missing data can complicate interpretation. Report effect size and uncertainty, not only whether a threshold was crossed.
The chosen metric is part of the design claim. More clicks can mean engagement or confusion; fewer cancellations can mean satisfaction or an obstructed exit. Combine the metric with task outcomes, qualitative evidence, and checks for burdens on different groups. Record what changed, what evidence improved, and what remains uncertain.
Designing for Diversity: Accessibility and Inclusion
Accessibility concerns whether people with disabilities can perceive, understand, and operate a system. It must be evaluated in actual tasks, including assistive technologies and different input methods. WCAG organizes web requirements under perceivable, operable, understandable, and robust, with conformance levels A, AA, and AAA. A technical conformance assessment has a defined version and scope; it is not a claim that every person will find every task easy.
Use semantic controls, meaningful labels, visible focus, keyboard access, appropriate alternatives for images and audio, and sufficient contrast. Test reading order and status announcements with a screen reader. A custom control that looks like a button may lack the behavior a native button already provides. ARIA can communicate semantics when used correctly, but does not implement missing keyboard behavior.
Targets also require enough space to select. WCAG 2.2’s minimum target-size criterion specifies a 24-by-24 CSS-pixel requirement with defined spacing alternatives and exceptions. This differs from a platform’s recommended touch size and does not mean that the minimum is comfortable in every context.
Inclusive design also considers language, text direction, date and number formats, device capability, interruption, and connectivity. Captions or larger controls can help beyond the group they were first intended to serve, but benefits are not identical for everyone. Involve affected users, preserve meaningful choices, and test the complete workflow rather than only a page’s appearance.
What Studying This Changes
You should be able to turn “the interface is confusing” into a testable explanation: the action was not discoverable, its consequence was mispredicted, the result was unclear, or recovery was unavailable. Each suggests a different design change.
Choose one task, such as editing a reservation. Observe users without teaching the intended route, record where their expectations differ from the system, and change one cause. Test the revised flow, including keyboard use and a failed network request. Compare actual task success and recovery, not just preference for the new appearance.
This practice connects HCI to systems correctness: the interface communicates which actions are possible, which changes persisted, and which decisions remain with the user. A reliable backend and an understandable interaction are both needed for a task to succeed.
Resources
Books and Texts
Norman’s The Design of Everyday Things develops vocabulary for actions, mappings, feedback, and conceptual models. Interaction Design supplies a broader account of research and design methods; Krug and Nielsen offer practical approaches to web usability and evaluation. Read cognitive claims alongside their experimental scope, and use Inclusive Components and WCAG to connect design intentions to actual interaction behavior.
| Book | Role | Type |
|---|---|---|
| Norman, The Design of Everyday Things (revised ed.) | Foundational vocabulary and philosophy | Entry |
| Rogers, Sharp & Preece, Interaction Design (6th ed.) | Academic foundational text | Entry |
| Krug, Don’t Make Me Think (3rd ed.) | Web usability; accessible entry | Entry |
| Nielsen, Usability Engineering | Usability testing and evaluation | Depth |
| Cooper et al., About Face (4th ed.) | Interaction design depth | Depth |
| Lupton, Thinking with Type (3rd ed.) | Typography for designers | Depth |
| Wickens et al., Engineering Psychology and Human Performance | Cognitive science foundations | Depth |
| Pickering, Inclusive Components (free) | Accessible component patterns | Entry |
| Microsoft Inclusive Design Toolkit (free) | Inclusive design framework and design exercises | Auxiliary |
| Tidwell, Brewer & Valencia, Designing Interfaces (3rd ed.) | Interaction patterns reference | Reference |
| Brignull, Deceptive Patterns website (free) | Documented deceptive interaction patterns | Reference |
Courses and Lectures
Stanford CS147 offers a project-oriented sequence through research, prototyping, and evaluation. Practitioner articles and other courses can supplement a specific method. A subscription or certificate is not necessary for the first project: define a task, make a prototype, observe representative participants, and revise it. Check the provider’s current terms if enrolling in a taught program.
| Course | Platform | Type |
|---|---|---|
| Stanford CS147 HCI Design (free) | Stanford course site | Entry |
| Interaction Design Foundation courses | ixdf.org | Entry |
| Nielsen Norman Group articles (free) | nngroup.com | Reference |
| Google UX Design Certificate | Coursera; check access conditions | Entry |
| Baymard Institute UX research and training (free articles; subscription paid) | baymard.com | Reference |
Practice, Tools, and Current Sources
Paper prototypes or tools such as Figma can expose a proposed flow before full implementation. Use keyboard navigation and screen readers to inspect an implemented flow, with guidance from people experienced in those tools. Automated checkers such as axe or WAVE find some rule violations; their pass result and any advertised coverage percentage do not establish complete accessibility. Design systems provide examples to examine, not permission to omit testing in your context.
| Resource | Platform | Type |
|---|---|---|
| Figma (free tier) | figma.com | Practice |
| VoiceOver / NVDA / TalkBack (free) | macOS / Windows / Android | Practice |
| axe DevTools / WAVE (free) | Browser extension | Practice |
| Apple Human Interface Guidelines (free) | developer.apple.com/design | Reference |
| Material Design 3 (free) | m3.material.io | Reference |
| GOV.UK Design System (free) | design-system.service.gov.uk | Reference |
| WCAG 2.2 documentation (free) | w3.org | Reference |
| Nielsen’s 10 Usability Heuristics (free) | nngroup.com | Reference |
| CHI papers and author versions | dl.acm.org / author pages | Reference |
Traps
| Trap | Why it misleads | Better response |
|---|---|---|
| Designing from your own familiarity | A designer knows conventions and hidden state that new users may not. | Observe realistic tasks with the intended participants. |
| Treating appearance as the whole experience | Discoverability, task structure, feedback, and recovery can fail in an attractive interface. | Specify and test the full task outcome. |
| Using five participants as a universal sample size | Discovery probability and estimation precision depend on the problem and population. | Use iterative discovery studies and size quantitative studies for their own questions. |
| Adding accessibility only at the end | Interaction structure and control semantics can already exclude users. | Design and test keyboard, screen-reader, and other relevant paths early. |
| Optimizing one business metric as if it represented user benefit | A metric can improve when users are confused or prevented from leaving. | Check task outcomes, unwanted effects, and affected groups. |
| Treating a heuristic or cognitive law as a complete explanation | Its assumptions may omit context, skill, device, or social activity. | Use it to propose a change, then test the prediction in context. |
中文
人机交互研究人怎样使用计算系统完成任务，以及怎样设计和评价这些系统。研究对象包括界面、用户的知识与能力、周围的活动情境，以及动作的后果。一个预约表单可能在技术上正确，却让人无法确定预约是否成立，或该怎样修改。
可以从任务及完成任务所需的信息开始。用户需要发现可用动作、预测效果、执行动作，再理解结果。这些环节都可能出问题。让按钮更醒目，不能解决退款规则不清楚的问题；解释规则，也不能解决键盘无法到达按钮的问题。
HCI 将实现与心理学、设计、工效学和社会活动研究联系起来。研究方法帮助区分设计者的假设与实际观察到的行为。视觉外观重要，可学性、任务成功、错误恢复、无障碍，以及对没有主动选择该系统的人的影响也重要。
前置知识：基本编程（§2.1）有助于制作交互原型。在掌握软件框架之前，也可以开展用户研究和纸面原型设计。进行定量比较时，需要概率与统计（§2.5）。
从批处理到为数十亿人设计
批处理使周转时间成为重要约束：用户往往提交工作，稍后才检查结果。但早期计算也包括交互控制、显示和直接操作。分时系统扩大了交互会话的使用，并不是所有人机交互的起点。重要变化是，人们在同一项任务中能够反复获得反馈，探索和修订自己的工作。
Sutherland 的 Sketchpad 展示了对所表示对象的直接操作。Engelbart 的 NLS 及其 1968 年演示，将指点、结构化文档、链接和协作组合进一个旨在增强智力工作的系统。这些系统研究计算机应支持哪些活动，而不只是怎样让已有命令更容易输入。
Xerox PARC 探索个人工作站、位图显示、指点和图形应用。商用 Xerox Star 在 1981 年出现，早于 1983 年的 Apple Lisa 和 1984 年的 Macintosh。这些产品在不同成本与市场约束下发展相关思想。选择、复制、打开和保存等共享约定，减少了用户在应用之间需要重新学习的内容。图形界面仍然需要学习，也不会使每项任务都不言自明。
Card、Moran 与 Newell 在 1983 年的工作，将任务表现与认知模型联系起来。GOMS 描述目标、操作、方法和选择规则，在指定任务及熟练行为假设下预测表现。Shneiderman 的界面原则和 Nielsen 的可用性工程提供了实用评估方法。其价值是使设计可以检查和测试；模型与清单都不能代替观察相关用户。
Norman 的研究澄清了概念模型、反馈、映射、可供性和意符。可供性指行动者与对象之间可能发生的动作，意符则传达在哪里、怎样行动。把手可能允许拉动，而标签说明拉动是不是预期动作。这个区别解释了为什么一项操作可能做得到，却难以被发现。
Web 将设计扩展到不同设备、语言、连接速度，以及由不同作者制作的页面。触摸界面改变了目标尺寸、可用反馈和手势的可发现性。移动使用还带来打断和物理情境的变化。无障碍与包容性设计进一步说明，不能把一种熟悉的用户、设备或感知能力当作所有人的默认情况。
AI 辅助界面增加了对系统输出和行动的不确定性。有些问题仍然熟悉，例如状态可见、纠正、委托和恢复；模型生成的错误又增加了新的困难。有用的设计应明确用户能获得什么证据、何时可以审查，以及怎样纠正错误动作。对话流畅本身不能回答这些问题。
认知、评估，以及为多样性设计
人类认知如何塑造界面要求
工作记忆负担取决于任务、熟悉程度、分组方式和外部提示。Miller 关于七个项目的讨论，不是菜单最多只能有七项的规则；后来的组块数量估计，也不是固定的界面容量。在用户需要时展示相关信息，例如前一步的选择、要确认的金额、已经完成的步骤。识别可以减少回忆负担，而稳定的命令语言也可以让熟练用户高效操作。
心智模型是用户对系统行为的理解。假设编辑器在网络请求刚开始时就显示“已保存”，请求失败后，界面就传达了错误的反馈与持久保存关系。区分保存中、确认已保存和可恢复的失败，有助于用户预测重新加载后还会保留什么。反馈应对应任务真正关心的状态。
映射把控件与效果联系起来。一致的标签、分组和空间排列，可以让映射更容易学习。意符让可用动作显现出来，但视觉外观还需要交互语义支持：控件应有明确名称，并能通过键盘使用。如果过于简化的视觉样式隐藏了必要区别，它未必减少认知工作。
指点操作有可测量的成本。Fitts 定律的一种常见形式是 T = a + b log₂(1 + D/W)，其中 D 是移动距离，W 是沿移动方向的目标宽度，a,b 根据任务和设备拟合。当 D=400 时，把 W 从 20 增到 80，难度指标约从 4.39 降到 2.58。这是在模型条件下对改善方向的预测，不是普遍适用的节省时间，也不是触屏操作的完整解释。
应区分失误与误解。失误是在目标正确时执行了非预期动作，误解则来自不正确的认识或计划。拉开控件间距、限制无效输入和保留撤销，可以减少失误的代价；清楚的解释与更好的任务结构，可以处理理解错误。反复出现的确认框容易变成习惯性点击，因此应在其信息确实有助于决策时使用，并尽可能提供恢复途径。
经验评估方法
按问题选择方法。形成性可用性研究观察有代表性的人完成真实任务，以发现障碍。出声思考可以显露某种理解，但说话也可能改变行为和计时。访谈调查经历和人们报告的原因，分析日志记录选定的行为；两者都不会自动揭示全部原因。
小规模迭代研究可以迅速发现有用的问题，但“五个用户”的建议不是覆盖保证。在简单的独立发现模型中，若每位参与者以概率 p 发现某个问题，那么 n 人中至少一人发现它的概率是 1−(1−p)^n。当 p=0.31 且参与者为五人时，约为 84%；当 p=0.05 时，则只有约 23%。实际概率随问题、任务和用户群体而变化。Nielsen 的说明 也明确讨论了相近用户及迭代改进的条件。
启发式评估可以在用户研究之前或同时提供专家检查。检查发现的是候选问题，不是用户失败频率的实测结果。应让合适的参与者测试任务完成、错误、恢复和满意度；要测量完成率或时间的小幅差异，需要为该目标设计样本规模，不能自动套用五人。
随机 A/B 测试比较两种分配方案下的结果。需要定义随机分配单位、主要指标、实际接触到的方案、观察期和停止规则。实现与分析保持随机化条件时，它支持因果比较；用户之间的影响或选择性的缺失数据，会使解释变复杂。应报告效应大小和不确定性，而不只报告是否跨过显著性阈值。
选择指标本身就是设计主张的一部分。点击更多可能代表投入，也可能代表困惑；取消更少可能代表满意，也可能代表退出受阻。将指标与任务结果、定性证据及不同群体的负担结合起来。记录改了什么、哪项证据改善了，以及什么仍不确定。
为多样性设计：无障碍与包容性
无障碍关注残障人士能否感知、理解和操作系统。评估应覆盖真实任务，包括辅助技术和不同输入方式。WCAG 按可感知、可操作、可理解和稳健四项原则组织 Web 要求，并区分 A、AA、AAA 符合性等级。技术符合性评估有明确版本与范围，不代表每个人都会觉得每项任务容易。
使用语义正确的控件、明确标签、可见焦点、键盘访问、适当的图像与音频替代内容，以及足够对比度。用屏幕阅读器测试阅读顺序和状态播报。看起来像按钮的自定义控件，可能缺少原生按钮已经提供的行为。正确使用 ARIA 可以传达语义，但不会实现缺失的键盘行为。
目标还需要足够的可选空间。WCAG 2.2 的最小目标尺寸准则 规定了 24×24 CSS 像素要求，并定义了间距替代条件和例外。这与平台推荐的触摸尺寸不同，也不表示最低要求在所有情境下都足够舒适。
包容性设计还考虑语言、文字方向、日期与数字格式、设备能力、打断和连接状况。字幕或更大的控件可能帮助原定群体以外的人，但不是每个人都获得相同收益。让受影响的用户参与，保留有意义的选择，并测试完整流程，而不只检查页面外观。
学习这一部分会改变什么
学完后，应能把“界面令人困惑”转化为可检验的解释：动作没有被发现、后果被错误预测、结果不清楚，还是无法恢复？每一种解释对应不同的设计改动。
选择一项任务，例如修改预约。观察用户操作，不提前教预定路线；记录他们的预期与系统在哪些地方不同，再针对一个原因修改。测试修订后的流程，包括键盘操作和网络请求失败。比较真实的任务成功与恢复，而不只询问是否更喜欢新外观。
这种实践把 HCI 与系统正确性联系起来：界面传达哪些动作可用、哪些改动已保存，以及哪些决定仍由用户掌握。任务成功既需要可靠的后端，也需要可理解的交互。
资源
书籍与文本
Norman 的 The Design of Everyday Things 解释动作、映射、反馈和概念模型。Interaction Design 提供更广的研究与设计方法，Krug 和 Nielsen 的著作提供 Web 可用性与评估的实用思路。阅读认知结论时注意实验范围，并用 Inclusive Components 与 WCAG 将设计意图连接到实际交互行为。
| 书籍 | 作用 | 类型 |
|---|---|---|
| Norman, The Design of Everyday Things（修订版） | 基础词汇与哲学 | 入门 |
| Rogers, Sharp & Preece, Interaction Design（第 6 版） | 学术基础文本 | 入门 |
| Krug, Don’t Make Me Think（第 3 版） | Web 可用性；易读入口 | 入门 |
| Nielsen, Usability Engineering | 可用性测试与评估 | 深入 |
| Cooper et al., About Face（第 4 版） | 交互设计深度 | 深入 |
| Lupton, Thinking with Type（第 3 版） | 面向设计师的字体排印 | 深入 |
| Wickens et al., Engineering Psychology and Human Performance | 认知科学基础 | 深入 |
| Pickering, Inclusive Components（免费） | 无障碍组件模式 | 入门 |
| Microsoft Inclusive Design Toolkit（免费） | 包容性设计框架与设计练习 | 辅助 |
| Tidwell, Brewer & Valencia, Designing Interfaces（第 3 版） | 交互模式参考 | 参考 |
| Brignull, Deceptive Patterns 网站（免费） | 欺骗性交互模式记录 | 参考 |
课程与讲座
Stanford CS147 通过项目组织研究、原型制作和评估。实践文章与其他课程可以补充某种方法。首个项目不需要订阅或证书：定义任务，制作原型，观察有代表性的参与者，再进行修订。如果参加授课项目，访问条件以提供方的现行说明为准。
| 课程 | 平台 | 类型 |
|---|---|---|
| Stanford CS147 HCI Design（免费） | Stanford course site | 入门 |
| Interaction Design Foundation courses | ixdf.org | 入门 |
| Nielsen Norman Group articles（免费） | nngroup.com | 参考 |
| Google UX Design Certificate | Coursera；访问条件见平台说明 | 入门 |
| Baymard Institute UX research and training（免费文章；订阅付费） | baymard.com | 参考 |
实践、工具与当前资料
纸面原型或 Figma 等工具，可以在完整实现前展示流程。检查已实现流程时，使用键盘导航和屏幕阅读器，并参考熟悉这些工具的人的经验。axe、WAVE 等自动检查器可以发现部分规则违反；通过检查或某个宣传覆盖比例，都不证明完整无障碍。设计系统提供可研究的例子，不能免去具体情境中的测试。
| 资源 | 平台 | 类型 |
|---|---|---|
| Figma（免费层） | figma.com | 实践 |
| VoiceOver / NVDA / TalkBack（免费） | macOS / Windows / Android | 实践 |
| axe DevTools / WAVE（免费） | Browser extension | 实践 |
| Apple Human Interface Guidelines（免费） | developer.apple.com/design | 参考 |
| Material Design 3（免费） | m3.material.io | 参考 |
| GOV.UK Design System（免费） | design-system.service.gov.uk | 参考 |
| WCAG 2.2 documentation（免费） | w3.org | 参考 |
| Nielsen’s 10 Usability Heuristics（免费） | nngroup.com | 参考 |
| CHI 论文与作者版本 | dl.acm.org / 作者页面 | 参考 |
陷阱
| 误区 | 为什么会误导 | 更好的做法 |
|---|---|---|
| 按自己的熟悉程度设计 | 设计者知道的约定和隐藏状态，新用户未必知道。 | 观察目标参与者完成真实任务。 |
| 把外观当成全部体验 | 漂亮界面也可能在发现操作、任务结构、反馈和恢复上失败。 | 定义并测试完整任务结果。 |
| 把五人当作通用样本量 | 发现概率和估计精度取决于问题与人群。 | 用迭代研究发现问题，为定量问题另外设计样本规模。 |
| 最后才补无障碍 | 交互结构与控件语义可能早已排除了部分用户。 | 早期设计并测试键盘、屏幕阅读器等相关路径。 |
| 把单一业务指标等同于用户受益 | 困惑或退出受阻也可能使指标变好。 | 检查任务结果、不良影响和受影响群体。 |
| 把启发式或认知定律当成完整解释 | 其假设可能遗漏情境、技能、设备或社会活动。 | 用它提出改进，再在具体情境中检验预测。 |