{
  "id": "human-computer-interaction",
  "title": "Human–Computer Interaction",
  "lang": "en",
  "source_lang": "en",
  "summary": "Human–computer interaction studies and designs how people use interactive computing systems, combining technical, behavioral, and social approaches.",
  "category": "computer-science",
  "subcategory": "software",
  "type": "concept",
  "aliases": [
    "HCI",
    "Computer–human interaction"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/human-computer-interaction/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/human-computer-interaction/",
      "json": "https://aiwiki.page/zh/human-computer-interaction.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Human–computer interaction (HCI) is the interdisciplinary study, design, implementation, and evaluation of interactive computing systems for human use. It examines how people perceive and operate technology, how interfaces support or obstruct their activities, and how computing changes work, communication, and everyday life. Its subject extends beyond screens and keyboards to include speech interfaces, wearable devices, collaborative environments, immersive systems, and interactions with artificial intelligence.\n\nHCI connects computer science and software engineering with psychology, design, ergonomics, and the social sciences. Its central concern is not simply whether a system functions correctly, but whether people can understand it, accomplish meaningful goals with it, recover from difficulties, and retain appropriate control. The relevant unit of analysis may be an individual performing a brief task, a team coordinating over months, or a community affected by a large digital platform.\n\nThe field combines explanatory research with the creation of new technologies and design methods. Controlled experiments, field observations, interviews, prototypes, mathematical models, and critical inquiry each address different aspects of interaction. Consequently, successful interaction cannot be reduced to a single measure such as speed, satisfaction, or engagement: its quality depends on the people involved, their purposes, and the circumstances in which a system operates."
    },
    {
      "heading": "Scope, usability, and user experience",
      "text": "An interactive system includes more than its visible interface. It comprises input and output devices, software behavior, information organization, physical surroundings, and the procedures through which people use it. A hospital scheduling application, for example, interacts with staffing practices, professional roles, and other information systems. An interface that works well in an isolated demonstration may fail when interruptions, shared responsibilities, or unreliable connectivity become part of the setting.\n\nUsability concerns whether specified users can achieve specified goals with effectiveness, efficiency, and satisfaction in a specified context of use. This formulation is central to ISO 9241-11:2018. Effectiveness concerns the accuracy and completeness of goal achievement; efficiency concerns the resources used in relation to results; satisfaction concerns whether the experience meets users’ needs and expectations. These dimensions can diverge. A confirmation step may slow a transaction while reducing consequential mistakes, and an initially efficient interface may become difficult when unfamiliar cases arise.\n\nUser experience encompasses people’s perceptions and responses associated with actual or anticipated use, including expectations, emotions, preferences, and experiences over time. It therefore extends beyond immediate task performance. The experience of a banking application includes confidence before making a payment, comprehension during the transaction, and the ability to resolve a later dispute. Visual appearance and animation contribute to this experience, but so do reliability, pricing, customer support, and organizational policy. User experience is not a synonym for visual design.\n\nHCI also distinguishes users from other affected parties. A workplace monitoring system may serve a manager while imposing burdens on employees; an automated admissions interface may affect applicants who never operate its administrative tools. Requirements therefore cannot always be derived from the preferences of a single customer or purchaser. Accessibility, safety, privacy, learnability, and compatibility with collective practices can all be legitimate design objectives.\n\nThese objectives require explicit trade-offs rather than a universal definition of an “intuitive” interface. Familiarity can make a convention appear self-evident even when it must be learned, and expert performance may depend on features that initially increase complexity. HCI investigates such tensions by specifying the intended population, tasks, environment, and timescale before judging whether an interaction is successful."
    },
    {
      "heading": "Historical development",
      "text": "Early electronic computing offered little immediate interaction. Programs and data were commonly prepared on external media and submitted for batch execution, with results returned later. Interactive computing emerged through developments in real-time control, display technology, and time-sharing. These developments changed the relationship between people and machines: instead of preparing a complete computation in advance, users could inspect intermediate results and revise their actions.\n\nSeveral mid-twentieth-century projects established influential possibilities. Vannevar Bush’s 1945 essay “As We May Think” described the hypothetical memex, a device for organizing and associating recorded knowledge. J. C. R. Licklider’s 1960 paper “Man-Computer Symbiosis” proposed close cooperation between human judgment and computer processing. These were different proposals, but both treated computing-related technology as a medium for intellectual activity rather than solely as an automatic calculator.\n\nIvan Sutherland’s 1963 Sketchpad system demonstrated interactive graphical drawing with a light pen, including constraints and relationships among graphical objects. Douglas Engelbart and his colleagues developed the oN-Line System, or NLS, as part of a broader program to augment collective intellectual work. Their public demonstration in December 1968 combined a mouse, interactive text editing, linked information, and remote collaboration. Such systems showed that interaction techniques and organizational ambitions could develop together.\n\nDuring the 1970s, Xerox PARC’s Alto research computer and associated software helped establish bitmapped displays, windows, pointing-based interaction, and networked personal computing. Commercial systems subsequently brought related ideas to larger audiences, notably the Apple Lisa in 1983 and Macintosh in 1984. These developments should not be understood as a single invention transferred unchanged into products: interface conventions emerged through numerous research systems, engineering decisions, and commercial adaptations.\n\nHCI became more institutionally distinct in the early 1980s. The Association for Computing Machinery established SIGCHI, its special interest group devoted to computer–human interaction, in 1982, and the CHI conference became a major venue. Stuart Card, Thomas Moran, and Allen Newell’s The Psychology of Human-Computer Interaction appeared in 1983, presenting quantitative models of skilled interaction. Subsequent expansion of the personal computer, World Wide Web, and smartphone broadened the field from office productivity to domestic, social, mobile, and public settings. HCI consequently developed multiple intellectual traditions rather than replacing each earlier approach with one successor."
    },
    {
      "heading": "Perception, cognition, and action",
      "text": "Interaction depends on human perceptual and cognitive capacities. Visual displays must support recognition of relevant objects, distinctions, and changes; auditory interfaces must remain intelligible amid competing sounds; physical controls must accommodate movement and reach. These capacities vary across individuals and circumstances. A label that is legible at a desk may be unreadable on a moving vehicle, and a notification that attracts attention in a quiet room may disappear in a noisy workplace.\n\nAttention is selective, and interface elements compete with one another and with events outside the system. Increasing visual salience can improve detection but also create distraction. Frequent alerts can lose their effectiveness when users learn that most do not require action. Similarly, a display can contain all necessary information while making its relationships difficult to perceive. Grouping, alignment, contrast, and consistent placement help people discover structure without examining every item separately.\n\nMemory places additional constraints on interaction. Recognition from visible alternatives generally imposes different demands from recalling an exact command. Menus, examples, recent-item lists, and meaningful defaults can reduce the need to remember details, although excessive options create search and selection costs. Interfaces that require users to retain an identifier while navigating between screens impose avoidable burdens. Such burdens become especially important during interruptions, unfamiliar tasks, or cognitively demanding work.\n\nA mental model is a person’s working understanding of how a system behaves. It need not reproduce the implementation; it must support useful predictions. A folder metaphor, for example, can help explain organization while concealing technical differences between local files, shared documents, and remote shortcuts. Difficulties arise when the interface encourages predictions that the system violates. Clear feedback, stable terminology, and visible relationships help users revise their understanding.\n\nDonald Norman’s account of action distinguishes difficulties in determining what to do from difficulties in interpreting what happened, commonly described as gulfs of execution and evaluation. An unfamiliar control may create the first problem, while an unexplained status indicator creates the second. Affordances concern possible actions in a relationship between an actor and an environment; signifiers communicate where or how actions can occur. A clickable region may technically permit activation yet remain undiscoverable without a visible indication. These concepts help explain why adding functionality does not necessarily make that functionality usable."
    },
    {
      "heading": "Predictive models and interaction laws",
      "text": "Quantitative models allow researchers and designers to predict some aspects of interaction before building a complete system. Their value lies in making assumptions explicit and identifying measurable relationships, not in explaining every dimension of experience. Models of skilled pointing or command execution are especially useful for repetitive, well-defined tasks. They are less complete accounts of exploration, creative activity, collaborative judgment, or changes in motivation.\n\nFitts’s law, introduced by psychologist Paul Fitts in 1954, describes the relationship between movement time and the difficulty of reaching a target. A widely used HCI formulation is\n\n\\[MT = a + b\\log_2\\left(1+\\frac{D}{W}\\right),\\]\n\nwhere \\(MT\\) is movement time, \\(D\\) is movement distance, \\(W\\) is target width along the approach direction, and \\(a\\) and \\(b\\) are empirically estimated constants. Larger or nearer targets are generally faster to acquire under the model’s assumptions. This helps explain the benefits of sufficiently large controls and the special properties of screen edges for a mouse pointer constrained by display boundaries. Touch input introduces different constraints, including finger occlusion and the absence of a persistent pointer.\n\nThe Hick–Hyman law, developed from work published by William Hick in 1952 and Ray Hyman in 1953, relates choice reaction time to information uncertainty. For equally probable alternatives, a common expression is\n\n\\[RT = a + b\\log_2(n+1),\\]\n\nwhere \\(n\\) is the number of alternatives. Real interfaces also involve visual search, grouping, semantic familiarity, and unequal probabilities of use. The law therefore does not justify a universal rule that every menu must contain fewer items. A well-organized visible list can be easier to use than a deeper hierarchy that requires remembering where options are hidden.\n\nThe GOMS family of models represents goals, operators, methods, and selection rules in skilled activity. The related keystroke-level model estimates routine execution time by combining elementary actions such as keystrokes, pointing, hand movement between devices, mental preparation, and system response. These models can compare alternatives for a repetitive workflow, but typically assume that users know a method and do not make unmodeled errors.\n\nModel parameters must be calibrated for relevant devices, tasks, and populations. A prediction based on experienced desktop users may not transfer to children, switch users, or people operating a system while walking. Empirical fit is also not a complete design judgment: a faster interaction may be harder to learn, less accessible, or more error-prone. Predictive models are consequently used alongside observation and evaluation rather than as replacements for them."
    },
    {
      "heading": "Human-centered design and prototyping",
      "text": "Human-centered design organizes development around an explicit understanding of people, their activities, and their contexts. ISO 9241-210:2019 describes principles including user involvement, evaluation-driven refinement, iteration, attention to the whole user experience, and multidisciplinary work. The approach is not a requirement to accept every requested feature. It treats user accounts as evidence about needs and circumstances that must be interpreted alongside technical constraints and the effects of proposed changes.\n\nA typical design process moves among understanding context, specifying requirements, producing design solutions, and evaluating those solutions. These activities are interconnected rather than strictly sequential. Observing how employees coordinate a handover may reveal that an apparent data-entry problem is actually a problem of unclear responsibility. A prototype may then expose an exception that changes the requirements. Iteration is valuable because important properties of an interaction often become visible only when people encounter a concrete representation.\n\nPrototyping makes alternatives available for inspection before all implementation costs have been incurred. Paper sketches can reveal difficulties with terminology and task sequence. Clickable wireframes can test navigation and information organization. Executable prototypes can expose timing, gesture recognition, or integration problems. Fidelity has several dimensions: a prototype may look polished while containing little functional behavior, or look rough while accurately reproducing a critical interaction.\n\nWizard-of-Oz studies simulate a capability through hidden or partially disclosed human operation, allowing investigation before automation is reliable enough to implement. They have been used for speech and other intelligent interfaces. Their findings depend on the simulator’s consistency and on whether the simulated behavior resembles what a deployable system could achieve. Where participants are not initially told how the system works, consent and debriefing require particular care.\n\nDesign artifacts such as scenarios, journey maps, task models, and personas help teams communicate requirements. Their usefulness depends on their evidential basis. A persona derived from research can make recurring needs easier to discuss; an invented demographic profile can instead reinforce assumptions. Design rationale records why alternatives were accepted or rejected, preserving knowledge when teams change. Participatory design goes further by involving affected people in shaping design decisions, a tradition strongly associated with Scandinavian workplace projects of the 1970s. It raises practical questions about representation, decision-making authority, and compensation rather than merely adding a feedback session to an otherwise fixed development process."
    },
    {
      "heading": "Field research and the study of practice",
      "text": "Many interaction problems cannot be understood from a task specification alone. People adapt tools, coordinate informally, improvise around breakdowns, and rely on knowledge that is not written into official procedures. Field research investigates these practices in the settings where they occur. Its methods include observation, interviews, contextual inquiry, diary studies, and the examination of artifacts such as forms, shared documents, and communication records.\n\nEthnography contributes sustained attention to situated activity and to participants’ own understandings. In an operations room, for example, the physical arrangement of displays may allow colleagues to notice each other’s progress without explicit messages. Replacing those displays with isolated personal screens can remove an important coordination mechanism even if every individual function remains available. Field observations reveal such dependencies by examining what people actually do rather than only what procedures say they should do.\n\nInterviews can investigate goals, past events, expectations, and organizational constraints, but they do not provide direct records of all behavior. Participants may forget routine steps or describe the formally approved procedure rather than common practice. Contextual inquiry combines observation with questions about ongoing work, making it easier to connect accounts to specific actions and artifacts. Diary studies capture experiences distributed over time, although reporting itself creates effort and can influence behavior.\n\nResearchers commonly analyze qualitative material through coding, comparison, and the development of themes or explanatory accounts. Different traditions assign different roles to prior theory and researcher interpretation. Agreement between coders can be useful for some structured classification tasks, but it is not a universal test of qualitative quality. Clear documentation of sampling, analytic decisions, contradictory cases, and the relationship between evidence and claims is more broadly important.\n\nField findings require careful treatment of scope. A detailed study of one workplace can uncover a mechanism that is relevant elsewhere without establishing how frequently it occurs across all workplaces. Access arrangements may also exclude less visible groups, temporary workers, or people who have abandoned a service. Researchers must protect participants whose comments could affect employment or social standing. Field research is therefore both a source of design requirements and a way to challenge the categories through which a problem was initially framed; it can show that a proposed interface improvement addresses only a small part of the activity in question."
    },
    {
      "heading": "Evaluation, experiments, and measurement",
      "text": "Evaluation asks whether a design supports its intended purposes and what difficulties emerge in use. Formative evaluation informs changes during development; summative evaluation characterizes a system against specified criteria. Usability testing typically involves participants attempting representative tasks while researchers observe behavior, record outcomes, and investigate breakdowns. Realistic tasks specify meaningful goals without revealing the intended sequence of interface actions.\n\nCommon measures include task completion, time, errors, assistance required, and subjective ratings. Their definitions matter. A participant who reaches the correct screen but misunderstands the result may not have completed the task effectively. An average completion time can conceal a minority of very long failures. Think-aloud protocols help expose interpretation and expectations, but speaking can alter timing or workload, so performance measures require appropriate caution.\n\nThe System Usability Scale, developed by John Brooke in 1986, uses ten questionnaire items to produce a score from 0 to 100. That score is not a percentage of tasks completed or a direct percentage measure of usability. Other instruments assess workload, pragmatic and hedonic qualities, or specific experiences. A questionnaire must match the construct being investigated, and translation or substantial adaptation may affect its measurement properties.\n\nControlled experiments compare conditions while addressing alternative explanations. Between-subjects designs assign different participants to conditions; within-subjects designs expose the same participants to multiple conditions and must address order, learning, and fatigue. Randomization and counterbalancing serve different roles in reducing bias. Sound interpretation considers effect sizes, uncertainty, exclusions, multiple comparisons, and practical importance rather than treating a statistically significant result as sufficient evidence of a valuable improvement. Accessibility and safety issues may also be important even when they affect too few participants to dominate an aggregate measure.\n\nInspection methods complement user studies. Heuristic evaluation examines interfaces against general principles, while cognitive walkthroughs focus on whether a user can discover and understand actions in a task sequence. Neither method establishes how actual users will behave in all circumstances. Likewise, the claim that five participants are enough for usability testing is not a general sampling rule: discovery depends on problem frequency, task coverage, user diversity, and study purpose. Large-scale online experiments can estimate behavioral differences precisely, but a measured increase in clicks or time spent does not by itself demonstrate improved welfare or task success."
    },
    {
      "heading": "Graphical interfaces and interaction architecture",
      "text": "The graphical user interface organizes interaction through visible objects and controls such as windows, menus, icons, buttons, and text fields. Its effectiveness depends on the relationships among representation, action, and feedback. A useful control must indicate what it affects, communicate its current state, and produce an interpretable response. Visual polish cannot compensate for a hidden dependency or an ambiguous result.\n\nBen Shneiderman’s influential 1983 account of direct manipulation emphasized continuous representation of objects and actions, physical actions rather than complex syntax, and rapid, incremental, reversible operations with visible results. Dragging a document or resizing a shape illustrates this style. Direct manipulation can support exploration and reduce the distance between an intention and its visible effect, but some operations remain easier to express symbolically. Renaming hundreds of files according to a rule may be more efficient through a command or script than through repeated dragging and editing.\n\nInformation architecture determines how content and functions are grouped, labeled, and connected. Navigation systems must support orientation as well as movement: users need to understand where they are, what alternatives exist, and whether an action changes the current object or takes them elsewhere. Search and browsing complement each other. Search depends on knowing how to express a need, while browsing can expose unfamiliar categories but may require more navigation.\n\nInteraction architecture also governs modes and system state. A mode changes the meaning of an input, as when typing edits text in one state but invokes commands in another. Modes can provide expressive power, yet create errors when users do not notice a transition. Persistent indicators, constrained actions, previews, and reversible operations make state easier to manage. Undo is particularly valuable because it supports experimentation without requiring users to predict every consequence in advance.\n\nGraphical and textual interfaces are not mutually exclusive. A command-line interface can offer composability, precise expression, and repeatability, while a graphical interface offers discoverable choices and spatial feedback. Many professional applications combine both. Their implementation commonly involves event handling, focus management, application state, and asynchronous updates. Keeping displayed state consistent with underlying operations is essential: a progress indicator that reports only that a request was sent can mislead users into believing the operation itself succeeded."
    },
    {
      "heading": "Input modalities and embodied interaction",
      "text": "Input techniques translate human actions into system operations, while output techniques make system states perceptible. Keyboards and pointing devices remain important because they support precise, sustained work, but HCI also studies touch, stylus input, speech, gesture, gaze, motion, sound, and tactile feedback. Each modality offers particular possibilities and imposes particular constraints; no single modality is universally “natural” or effortless.\n\nTouch interaction reduces the separation between the action surface and displayed objects, but fingers obscure content and provide less precise pointing than a fine cursor in many tasks. Controls require adequate size and spacing, and gestures must be discoverable or learnable. A stylus can improve precision and support handwriting or drawing, while introducing questions about pressure, palm rejection, and the transition between marking and command invocation.\n\nSpeech interfaces combine speech recognition with language interpretation and dialogue management. They can support hands-busy activity and users who find typing difficult, but are affected by noise, accents, privacy, turn-taking, and uncertainty about available commands. Spoken output is transient and sequential: hearing a long list is different from scanning one visually. Repair mechanisms, such as clarifying a name or confirming a consequential action, are therefore central components rather than exceptional additions.\n\nMultimodal interaction combines channels, for example by pairing speech with pointing. The channels may supply complementary information: “move this there” becomes interpretable through accompanying gestures. Combining modalities requires temporal alignment, conflict resolution, and an account of what happens when one channel fails. Redundancy can improve access, but only if the alternative channels genuinely convey the needed information rather than reproducing the same barrier.\n\nHaptic interfaces communicate through vibration, force, or other tactile effects. Virtual reality and augmented reality introduce additional relationships among head movement, body position, rendered images, and the physical environment. Latency, tracking error, inappropriate depth cues, and mismatched motion can affect performance and comfort. Gaze is useful for estimating attention or selecting targets, but looking is not always an intention to activate; dwell times and explicit confirmation can address this ambiguity at the cost of speed. Embodied interaction also concerns how meaning arises through physical action and social context, not merely the replacement of buttons with sensors. A gesture acceptable in a laboratory may be tiring, conspicuous, or culturally inappropriate in public use."
    },
    {
      "heading": "Accessibility and inclusive interaction",
      "text": "Digital accessibility concerns whether people with disabilities can perceive, understand, navigate, and operate digital systems. It includes compatibility with assistive technologies and the design of alternatives to inaccessible interactions. Relevant differences include vision, hearing, mobility, speech, cognition, and combinations of conditions. Ability can also vary over time or with circumstances, making a fixed division between “disabled users” and an otherwise uniform population inadequate.\n\nThe Web Content Accessibility Guidelines, particularly WCAG 2.2 published as a W3C Recommendation in 2023, organize requirements under four principles: content should be perceivable, operable, understandable, and robust. Examples include text alternatives for meaningful non-text content, keyboard access, sufficient contrast, visible focus, and labels that communicate control purposes. Conformance provides an important testable framework, but it does not guarantee that every person can efficiently complete every task.\n\nAssistive technologies depend on information that may not be visible on the screen. A screen reader uses programmatic structure to announce headings, links, controls, states, and updates. In web applications, native semantic HTML elements provide many expected behaviors, while custom controls require additional implementation. An interface that looks like a button but lacks the appropriate role, keyboard behavior, or accessible name may be unusable to someone navigating nonvisually.\n\nAccessibility involves interaction sequences as well as individual components. A keyboard-operable form can still fail if focus disappears after an update, errors are announced without their locations, or a session expires before the user can finish. Captions must identify relevant speech and sound information, not simply display unreviewed recognition output. Visualizations may require descriptions, underlying data, or alternative exploration methods depending on their purpose. For cognitive accessibility, clear language, stable navigation, manageable steps, and understandable error recovery can be as important as technical interoperability.\n\nInclusive design considers variation from the beginning rather than treating accommodation as a final repair. It benefits from involving people with relevant lived experience and testing realistic end-to-end tasks with their preferred technologies. No single participant represents an entire disability category, and needs can conflict: dense information may support expert comparison while increasing cognitive burden for others. Internationalization introduces related challenges involving writing direction, text expansion, naming conventions, date formats, and reading proficiency. Accessible design therefore requires both standards-based checking and investigation of actual use across a deliberately varied population."
    },
    {
      "heading": "Collaboration and social interaction",
      "text": "Computer-supported cooperative work studies how groups coordinate activities through and around computing systems. Collaboration is not simply individual interaction multiplied by the number of participants. People must establish shared understanding, divide responsibilities, remain aware of others’ actions, and manage disagreements. A system can support each person’s local task while making collective work more difficult if it obscures dependencies.\n\nCollaborative systems are often described in terms of time and place: participants may work synchronously or asynchronously, and may be colocated or geographically distributed. These distinctions help identify design requirements but do not fully determine them. A shared document may combine simultaneous editing, delayed comments, and face-to-face discussion. A video meeting may connect several rooms whose internal conversations remain only partially available to remote participants.\n\nAwareness mechanisms communicate who is present, what has changed, and what others are doing. Shared cursors, edit histories, activity feeds, and status indicators can reduce coordination costs. They can also introduce distraction or unwanted monitoring. The appropriate amount of visibility depends on the work and its social relationships. A continuous presence indicator may be helpful for a tightly coordinated operation but misleading when workers perform important tasks away from their devices.\n\nTechnical consistency is only one component of collaborative editing. Distributed computing mechanisms can synchronize document state, but users still need to understand conflicting intentions and the meaning of revisions. Attribution, version history, comments, and permissions make changes interpretable and negotiable. Access control likewise has interaction consequences: an unexplained permissions error can interrupt a workflow, while an overly broad sharing action can expose information unintentionally.\n\nLarge-scale social media extends these concerns to audiences, moderation, identity, and community governance. Content created for one group may reach another with different expectations, a phenomenon often described as context collapse. Interface choices influence who can reply, how audiences are represented, which contributions become visible, and how harassment can be reported. The work of maintaining a community may fall disproportionately on moderators or particular members. Evaluation must therefore consider collective outcomes and the distribution of burdens, not only individual engagement. Systems for education, scientific collaboration, and mutual assistance similarly require attention to trust, incentives, and institutional arrangements that cannot be inferred from a button-level usability test alone."
    },
    {
      "heading": "Mobile, ubiquitous, and safety-critical settings",
      "text": "Mobile interaction occurs amid changing physical and social conditions. Small displays, intermittent attention, movement, battery constraints, and variable connectivity shape what users can do. A task completed while seated with both hands available may become difficult when carrying luggage or receiving directions outdoors. Responsive layouts address only part of this problem; mobile design also concerns interruption, resumption, offline behavior, and the timing of requests.\n\nMark Weiser’s 1991 article “The Computer for the 21st Century” articulated a vision of ubiquitous computing in which computation becomes embedded in everyday environments. Contemporary examples include wearable sensors, connected appliances, access systems, and environmental controls. Such systems may have no single interface or clearly bounded session. Users must nevertheless understand what is sensing, which device has authority, how actions propagate, and how to recover when connections fail.\n\nContext-aware systems adapt to information such as location, time, activity, or nearby devices. Adaptation can reduce effort, but incorrect inferences create uncertainty. Automatically silencing a device during a meeting is useful only if the meeting is identified correctly and exceptions remain manageable. Making inferred context inspectable and allowing straightforward correction helps users understand behavior. Shared environments introduce further complications because the preferences or permissions of one occupant may not represent those of everyone present.\n\nSafety-critical interaction includes aviation, industrial control, transport, and clinical information systems. Here, an error must be understood in relation to the wider technical and organizational system. A misleading display, poorly differentiated controls, an interrupted handover, or a delayed update can combine with other conditions. Alarm systems must distinguish urgency and support diagnosis rather than simply produce more alerts. Simulations and scenario-based evaluations are important because ordinary task completion measures do not establish performance under rare but consequential conditions.\n\nAutomation changes the operator’s role rather than removing human involvement. People may shift from direct control to monitoring and exception handling, creating problems of reduced vigilance, skill maintenance, and handover at precisely the most difficult moments. Situation awareness concerns perception of relevant elements, comprehension of their significance, and projection of their future status. Interfaces can support it through meaningful trends, clear automation modes, and visible system limits. Evaluation in these settings considers recovery, degraded operation, and coordinated response, not only routine speed."
    },
    {
      "heading": "Interaction with artificial intelligence",
      "text": "Artificial intelligence introduces systems whose outputs may depend on learned statistical relationships rather than only on explicitly specified rules. From an interaction perspective, this creates uncertainty about capability, consistency, and failure. A conventional command may have a predictable result within a known state, whereas an AI recommendation or generated answer can vary with context and may be plausible without being correct.\n\nHuman–AI interaction therefore concerns how capability is communicated, how users express goals, and how errors become detectable and correctable. Machine learning applications include recommendations, recognition tools, adaptive interfaces, and decision support. Appropriate feedback differs across these tasks. A photo-organizing system may tolerate some mistaken suggestions; a consequential administrative workflow may require a reviewable record of inputs, decisions, and overrides.\n\nLarge language model interfaces allow open-ended linguistic instructions and can support drafting, explanation, and information access. They also make the space of available operations less visible than a conventional menu. Users must infer what to request, how to specify constraints, and how to assess the response. AI hallucinations complicate this assessment because fluent output can conceal unsupported claims. References, source inspection, constrained workflows, and clear separation between generated content and verified records can support checking, but no presentation technique makes all generated statements reliable.\n\nTrust calibration concerns whether reliance matches actual performance in the relevant circumstances. Neither maximum trust nor constant distrust is an appropriate universal goal. Explanations may improve understanding, but they can also increase acceptance without improving judgment. Explainable artificial intelligence must therefore be evaluated for its effects on actual decisions, error detection, and understanding, rather than only whether users report liking an explanation. A confidence value is useful only if its meaning and calibration are appropriate for the task.\n\nSystems that take actions introduce additional issues of delegation. An AI agent that sends messages, modifies files, or makes purchases needs interaction mechanisms for permissions, scope, previews, interruption, and recovery. Requiring approval for every trivial action defeats delegation, while hiding consequential actions undermines control. Research consequently examines graduated autonomy and mixed-initiative interaction, in which initiative can shift between person and system. Evaluation must measure the combined human–AI process: assistance can improve an isolated output while increasing review effort, introducing new errors, or weakening performance when the assistance is unavailable."
    },
    {
      "heading": "Privacy, ethics, and research frontiers",
      "text": "HCI examines how interface choices distribute knowledge, effort, opportunity, and control. Privacy is partly a technical matter, but it is also an interaction problem: people must understand what information is collected, for what purposes, and with which consequences. A settings page does not provide meaningful control if choices are incomprehensible, defaults are obscure, or refusal prevents unrelated activities. Likewise, consent cannot be assessed solely by whether a button was clicked.\n\nUsable security investigates the relationship between protective mechanisms and human behavior. Authentication, warnings, permissions, and recovery procedures must operate under realistic conditions, including uncertainty and time pressure. If secure behavior requires remembering numerous arbitrary rules or repeatedly interrupting legitimate work, users may develop workarounds. Multifactor authentication, for example, involves not only enrollment and everyday approval but also device loss, account recovery, accessibility, and resistance to misleading requests.\n\nDeceptive design patterns are interface arrangements that steer or manipulate users toward outcomes they might not otherwise choose. Examples include making cancellation much harder than enrollment, presenting unequal visual emphasis for equivalent choices, and obscuring recurring charges. Analysis distinguishes legitimate guidance from designs that undermine informed choice. Measures such as conversion rate or retention can reward these practices unless evaluation also considers comprehension, regret, and the ability to refuse.\n\nResearch ethics concerns participant consent, confidentiality, compensation, and foreseeable harms. Large-scale online studies raise questions about what constitutes research participation when interface changes are deployed through ordinary services. Studies involving children, dependent relationships, sensitive records, or vulnerable communities require particular safeguards. Researchers must also consider people indirectly affected by a prototype, such as bystanders recorded by a wearable device or recipients of automatically generated communication.\n\nPersistent research challenges include understanding long-term consequences, comparing designs across cultures and institutions, and measuring benefits that do not appear in short laboratory tasks. Environmental impacts connect interaction design to device lifetimes, repairability, energy demand, and patterns of consumption. New sensing and generative systems intensify questions about agency, provenance, and surveillance. Across these areas, a central methodological difficulty is linking local evidence about an interface to wider consequences without assuming that short-term convenience, technical novelty, or increased use necessarily represents an improvement."
    }
  ],
  "links": [
    "computer-science",
    "software-engineering",
    "personal-computer",
    "world-wide-web",
    "smartphone",
    "hick-hyman-law",
    "ethnography",
    "system-usability-scale",
    "speech-recognition",
    "virtual-reality",
    "screen-reader",
    "html",
    "distributed-computing",
    "social-media",
    "education",
    "artificial-intelligence",
    "machine-learning",
    "large-language-model",
    "ai-hallucination",
    "explainable-artificial-intelligence",
    "ai-agent",
    "data-privacy",
    "multifactor-authentication"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}