RESEARCH DOSSIER
Interface Choice Architecture and Digital Affordances
Abstract. Interface choice architecture and digital affordances describe the organization of decision opportunities in interactive environments and the action possibilities emerging between digital systems and their users. This fundamental research monograph develops a differentiated account of defaults, salience, disclosure, friction, navigation, consent interfaces, subscription pathways, and manipulative design patterns. It reconstructs ecological and human–computer interaction theories of affordances, decision framing, digital nudging, and experimental studies of privacy and commerce. The analysis compares field experiments, controlled consumer studies, interface audits, meta-analyses, contemporary causal methods, and 2026 research on consent and preference inconsistency. The synthesis finds credible, context-dependent causal evidence that particular interface structures change selections and decision processes; it also identifies variation in effect size, disputed aggregate nudging estimates, measurement limitations, and the unresolved relation between observable conversion and autonomy. The monograph organizes an empirical program centered on user comprehension, the practical ability to exercise or reverse preferences, heterogeneous access, and longitudinal evaluation of adaptive and human–artificial interfaces. Methodologically, it is a critical integrative narrative synthesis of identifiable primary research, scientific reviews, methodological literature, and official reports, with explicit distinctions among experiments, observational analyses, working papers, and normative frameworks.
Keywords: interface choice architecture; digital affordances; defaults; human–computer interaction; digital nudging; dark patterns; deceptive design; privacy consent; cookie banners; interface friction; visual salience; decision comprehension; choice quality; accessibility; causal inference; behavioral economics; manipulation; user autonomy; artificial mediation.
The Scientific Problem of Interface-Mediated Choice
Digital interfaces organize the conditions under which people encounter, interpret, compare, and execute alternatives. A subscription form, privacy dialog, mobile banking screen, search result, or artificial-agent control panel establishes a sequence of possible actions, a representation of their consequences, and a distribution of effort among them. Interface choice architecture is the systematic arrangement of those conditions. Its psychological importance lies in the fact that decisions are situated activities: the user's preferences, knowledge, attention, abilities, and goals interact with the arrangement of the environment at the moment of action. Interface research must therefore specify the arrangement, the decision process, the resulting behavior, and the meaning of that behavior for the user.
This monograph examines defaults, visibility, navigation, friction, disclosure, perceived possibilities for action, and manipulative interface patterns. The primary scientific questions are how an interface makes particular actions possible or salient; how its structures alter attention, comprehension, effort, and perceived control; how experimental studies identify changes in decisions; and how researchers distinguish higher interaction rates from improved decision quality. Its comparative scope includes privacy choices, online purchases, subscriptions, and digital service settings. The evidence is synthesized through a critical integrative narrative review of experimental human–computer interaction, behavioral decision research, communication theory, meta-analysis, and institutional investigation. It is neither a preregistered systematic review nor an original experiment.
The inquiry belongs to Media and Technology Psychology because the relevant environment is an interactive medium whose presentation can be changed independently of the user's underlying objective. It is connected to human factors, economic psychology, psychometrics, and causal inference. Its boundaries are equally important: reinforcement schedules and notification-driven habitual engagement, recommender-controlled media exposure, and generalized consumer choice-set complexity constitute adjacent research objects. Here they matter only when they clarify the design and interpretation of a concrete interface decision.
A decisive methodological principle follows: the same click can signify informed preference, compliance with an interface recommendation, confusion, resignation, or a wish to escape the interaction. The observable action and the participant's endorsed goal therefore require separate measurement. The history of decision framing demonstrates why the objective options alone are insufficient to explain selection (Tversky, A., & Kahneman, D (1981)); the privacy literature shows why knowledge and preference may fail to map neatly onto disclosure (Acquisti, A., Brandimarte, L., & Loewenstein, G (2015)).
Definition and Levels of Analysis
Interface choice architecture is the arrangement of selectable actions, information, defaults, visual emphasis, sequencing, and interaction costs in a digital environment that can influence how a decision is understood or implemented. A digital affordance is a context-dependent possibility for action arising through the relationship between a user's capabilities and goals and the material, informational, and social properties of a digital system. These concepts operate at different analytical levels. A preselected checkbox is an interface feature and part of choice architecture. The ability to inspect, reject, or revise consent is a set of potential actions available through the system. The eventual consent decision is an outcome.
Four levels are analytically useful. The first is the implemented interface: rendered elements, interaction sequences, input rules, and technical constraints that can be recorded reproducibly. The second is the perceived decision environment: which choices the participant notices, understands, trusts, and believes can be selected. The third is the process of selecting or abandoning an action, including attention, effort, inference, and emotional responses. The fourth comprises outcomes, including the selected option, accuracy, subjective autonomy, satisfaction, persistence of the choice, and later regret. A rigorous explanation identifies the path between levels rather than treating a screenshot as sufficient proof of psychological causation.
The distinction between feature, relational affordance, and outcome is developed systematically by Evans, S. K., Pearce, K. E., Vitak, J., & Treem, J. W (2017). Their criteria ask whether a proposed affordance is independent of a particular feature, whether it is distinguishable from the outcome, and whether it varies across user–technology relationships. Visibility, for example, can facilitate the location of information and may vary with interface design and user expertise; actually finding a document is the outcome. These distinctions prevent the common error of labeling every attractive button or every observed behavior an affordance.
Digital choice has a temporal dimension as well. A person may initially accept an option and later discover its recurring costs; a platform may save prior settings and make them the next encounter's defaults. An intervention may change the path into a decision without changing the options as conventionally described. Consequently, the relevant research unit is often a task trajectory rather than a single isolated tap. The analysis should record the initial state, the options realistically available, the number and type of steps required to reach them, and the downstream consequences observable within the study.
Historical Origins of Affordance Theory
James J. Gibson's ecological psychology framed affordances as opportunities for activity relative to an animal and its environment. In Gibson, J. J (1979), the perceptual environment is understood through action-relevant structures as well as visual description. The concept was subsequently extended into human–computer interaction, where designers needed to explain why an apparent control could be used intuitively in one situation and misunderstood in another. The migration of affordance language from physical settings to software produced a productive conceptual question: how does a rendered object constrain or enable an act when its apparent shape may be conventional rather than physically necessary?
Gaver, W. W (1991) emphasized that possible actions and perceptual information about those actions can diverge. A system may permit an action users do not discover, or appear to permit an action it cannot perform. Don Norman later differentiated conceptual roles of affordance and convention in Norman, D. A (1999). In graphical interfaces, the apparent clickability of a button frequently depends on learned design conventions, visual feedback, and signifiers. The resulting practical question is how a person recognizes a meaningful action, predicts its effect, and knows whether the system registered it.
These origins matter for modern empirical design. A high-contrast control may increase visual detection without making its function accurately comprehensible. A gray link may be available but rarely noticed. A disabled-looking button may signal impossibility despite being technically active. Likewise, low discoverability can be intentional: the ability to reject optional tracking can exist while being buried in a secondary menu. An experimental manipulation should specify whether it changes objective availability, perceived availability, interpretation, or the cost of exercise.
An affordance analysis also avoids technologically deterministic assumptions. Two users facing the same control may have different feasible actions because of motor abilities, language competence, prior experience, assistive technology, or available time. Evans, S. K., Pearce, K. E., Vitak, J., & Treem, J. W (2017) show why the relational status of affordances is central to comparing media contexts. An empirical instrument can record both features and perceived action possibilities, test differences between user populations, and relate those differences to eventual behavior.
Decision Framing and the Emergence of Digital Nudging
Behavioral decision research established that choice may depend on how equivalent outcomes are described. Tversky, A., & Kahneman, D (1981) demonstrated that framing can affect preferences, providing a theoretical foundation for investigating the labels, reference points, order, and contextual presentation of digital alternatives. Interface research extends that inquiry from verbal formulations to interactive decision sequences: the user may face a preselection, a salient path forward, a nested set of settings, or an apparent urgency message. These features can operate alongside ordinary valuation and deliberation rather than replacing them.
The term digital nudging has been used for the intentional guidance of behavior by user-interface design. Weinmann, M., Schneider, C., & vom Brocke, J (2016) formulated the concept for digital choice environments. It invites a comparison between different ways to structure decisions while preserving formal availability of alternatives. However, the scientific category covers heterogeneous interventions: changing a default, simplifying a form, making consequences visible, and presenting a social cue are psychologically different operations. Evidence for one cannot automatically validate another.
Two complementary traditions can therefore be traced. One analyzes decision biases and context effects through controlled experiments. The other examines action possibilities, usability, and meaning in human–technology relationships. Their convergence is especially apparent in consent dialogs and checkout processes, where a visual representation changes attention and effort while the underlying economic or privacy preference may remain stable. This suggests a joint research strategy: manipulate a specific design parameter, measure the psychological mediator it is thought to affect, and evaluate the choice against criteria broader than immediate compliance.
The history also clarifies an ethical distinction. Some design interventions make a goal easier for the person who holds it, such as highlighting a security-critical setting with clear explanations. Others exploit a mismatch between the user's priorities and the interface operator's incentives, such as making a costly commitment more prominent than its cancellation route. This distinction requires evidence concerning user intent, clarity, constraints, and consequence. Classification by interface appearance alone is a starting hypothesis, not a complete evaluation of psychological harm.
Relational Affordances and Discoverability
An interface may provide searchability, editability, reversibility, portability, persistence, or visibility. Each names a potential relationship between a user and the system rather than a single button. For choice architecture, the most important affordances often concern understanding and changing a decision: can a user identify the consequences, inspect alternatives, reverse a selection, revisit a setting, and obtain a record of the action? Those possibilities influence the practical scope of control before any explicit persuasion is applied.
Consider a privacy panel displaying a prominent acceptance control, a secondary button for managing preferences, and a link to reject optional collection several screens away. The objective alternatives may exist, yet the participant's actual opportunity to exercise them differs by cognitive, motor, and temporal resources. Discoverability can be measured as the proportion identifying an option without prompting, time to locate it, unsuccessful interaction attempts, and accuracy in explaining what it does. The experimental question is whether redesigning visibility and control placement changes these measures and the eventual expressed preference.
A related distinction concerns interface legibility and trust. Users must recognize not only that a control exists but also what its consequences will be. A deletion button can be unmistakable yet ambiguous about whether it removes a local record, server copy, or public post. Clearly labeled alternatives create more accurate mental models. Davis, F. D (1989) supplied influential measurements of perceived ease of use and usefulness in technology acceptance; those global constructs should be distinguished from task-specific discoverability and decision comprehension. A tool can be judged easy to use while its contractual implications remain obscure.
Affordance analysis becomes more rigorous when researchers independently code what the interface objectively permits, what users believe it permits, and what actions they successfully perform. Disagreement among these measurements reveals whether the obstacle is physical interaction, conceptual understanding, mistrust, or goal conflict. Repeated testing with assistive technology, varied literacy levels, and diverse devices is therefore part of investigating an affordance rather than an optional demonstration of inclusiveness.
Defaults and the Architecture of Initial States
A default is an option selected or applied in the absence of an active alternative selection. Digital defaults appear in consent controls, account privacy, recommended configuration, delivery settings, search preferences, and subscription renewals. Their causal effect can be studied by random assignment to different initial states while keeping the available options and substantive information constant. The primary behavioral measure is whether users retain or change the initial state; stronger evaluation adds comprehension, choice confidence, and preference stability.
Johnson, E. J., & Goldstein, D (2003) made default effects prominent in the context of organ-donation consent. Their work linked experimentally grounded questions about defaults to large institutional differences, though national comparisons cannot isolate defaults from legal, cultural, and administrative factors. The appropriate conclusion is that defaults warrant causal investigation in each decision environment. It would be an error to transport a country-level association directly to an estimate for a contemporary website.
A substantial synthesis by Jachimowicz, J. M., Duncan, S., Weber, E. U., & Johnson, E (2019) analyzed 58 studies with a pooled sample of 73,675 and estimated a mean default effect of d = 0.68, with a 95% confidence interval from 0.53 to 0.83. The same review documented important between-study variation, including results with weak, null, or negative effects. Its contribution is therefore twofold: default influence is a recurring empirical finding, and its strength requires contextual explanation. Effect size is not a universal forecast for any new interface.
Several causal paths deserve separate testing. Retaining a default can save effort, signal a recommendation, imply a perceived social norm, or establish a reference point. A user may also fail to realize that the initial state is selected. These mechanisms predict different consequences. If retaining the default reflects a trusted recommendation, explanatory labeling may preserve choice. If it reflects hidden preselection, greater visibility and an explicit decision requirement may reduce the effect. Researchers should avoid inferring a specific mechanism solely from a changed retention rate.
Visual Salience and Attentional Selection
Interfaces compete for limited attention. Position, size, contrast, typography, whitespace, and motion can determine which information is sampled before action. A prominent confirmation control can be noticed earlier than an equally available rejection link, influencing the path of decision making. In this context, salience is a design manipulation, attention allocation is a potential mediator, and selection is an outcome. Conflating these three produces explanations that simply restate the observed result.
Attention can be studied through eye-tracking measures such as first fixation, gaze duration, transitions between alternatives, and reinspection of disclosures. Cursor tracking and interaction logs may approximate aspects of exploratory behavior, though they have different measurement properties and cannot be treated as interchangeable with gaze. The most informative design tests whether a salience manipulation alters information sampling and whether information sampling predicts better comprehension, a more stable decision, or merely a faster click.
Salience has no fixed direction of welfare effect. Making essential risk information easier to locate may improve decisions; making optional purchases visually dominant may raise addition rates while leaving product understanding unchanged. Behavioral evaluation must therefore distinguish probability of choosing an option from the match between that option and the person's articulated goal. Tversky, A., & Kahneman, D (1981) supports the broader possibility that presentation matters, but each new digital cue requires direct study rather than analogy alone.
A methodological complication is that salience often changes several variables simultaneously. Color contrast may alter legibility, desirability, and perceived legitimacy. A larger button may also offer a larger touch target, benefiting people with motor limitations. Factorial experiments or carefully matched controls can help estimate the separate contributions of visual prominence, operability, and informational content. Such designs are especially important when claims about manipulation are made from visual screenshots.
Disclosure and the Timing of Information
Disclosure is the presentation of information that could matter to an impending choice, such as data recipients, subscription terms, recurring costs, cancellation conditions, or the consequences of changing a setting. Its psychological significance depends on more than factual presence. Disclosures must be visible, comprehensible, temporally relevant, and connected to a real opportunity to act. A privacy description available after the user has already agreed to tracking may have informational value while offering little prospective control over that particular decision.
Information architecture shapes whether a person encounters a disclosure while constructing a preference, after adopting an initial course, or only during a later attempt to reverse it. Researchers can vary position, level of detail, language complexity, and the proximity of explanations to associated controls. Behavioral outcomes can be combined with objective tests of whether people understand who collects data, for what purposes, how long collection continues, or what payments recur. Such measurements offer a stronger basis for judging informed decisions than simply recording whether a dialog was dismissed.
Acquisti, A., Brandimarte, L., & Loewenstein, G (2015) review behavioral processes in privacy decision making, including limited attention, social context, uncertainty, and the gap between expressed attitudes and observed disclosure. Their synthesis helps explain why merely adding information may fail to correct an unbalanced decision environment. Increased disclosure can create overload when essential terms are embedded in extensive technical language; selective disclosure can create false reassurance. The challenge is to improve meaningful comprehension while preserving a timely and executable choice.
A scientific evaluation should avoid presuming that more text always improves autonomy. Excessive legal prose can impose cognitive costs, and overly simplified summaries can omit decision-critical details. Experimental comparisons should record both objective comprehension and subjective confidence, allowing researchers to detect cases where an interface increases confidence without improving understanding.
Choice Sets, Complexity, and Decision Burden
Digital choice sets vary in size, structure, similarity, reversibility, and the degree to which attributes can be compared. Their effects depend on the participant's expertise, task difficulty, and motivation. Iyengar, S. S., & Lepper, M. R (2000) reported three field and laboratory experiments in which larger sets sometimes reduced selection, satisfaction, or motivation relative to smaller alternatives. These findings were influential in developing choice-overload hypotheses, but an interface with many alternatives is not automatically detrimental: relevant grouping, filtering, and searchability can make extensive catalogs manageable.
The distinction between number of options and organization of options is especially important online. A product listing may contain hundreds of items yet allow the person to filter by criteria that matter to them. Another interface may provide only three plans but hide crucial differences among them. Researchers should separately manipulate assortment size, comparability, presentation order, decision stakes, and filtering capabilities. Without such separation, a supposed effect of option count may arise from visual complexity or differences in information quality.
Decision burden can be measured through decision time, gaze alternation, search depth, subjective difficulty, error rates, postchoice satisfaction, and behavioral consistency when logically equivalent options are presented differently. No single measure captures all burden. An extended decision time can indicate effortful confusion or careful consideration; a short decision time can reflect confidence or impulsive acceptance. Interpretation depends on the task and its normative demands.
Choice-set complexity connects the present dossier to Consumer and Economic Psychology while retaining its distinct focus: the interface is the manipulable environment, including layout, navigation, and information presentation. Broader theories of preference construction and assortment effects remain independent research subjects that inform, but do not replace, interface-specific causal investigation.
Privacy Consent as a Controlled Decision Paradigm
Cookie and privacy-consent interfaces have become a particularly informative experimental setting because a range of nominally available alternatives can be displayed using different defaults, button colors, locations, and numbers of interaction steps. The core dependent variable is often whether the user authorizes optional collection. That outcome is readily logged, but its psychological interpretation requires knowing whether the participant understood the request and could exercise a contrary preference without excessive effort.
Nouwens, M., Liccardi, I., Veale, M., Karger, D., & Kagal, L (2020) examined 680 consent interfaces associated with major consent-management platforms in the United Kingdom and performed a field experiment with 40 participants. In that study, omitting an opt-out control from the initial page increased agreement by 22–23 percentage points, while presenting more granular choices on the initial page reduced agreement by 8–20 percentage points. These figures concern the authors' tested designs and population; their small experimental sample, limited design classes, and specific regulatory context restrict generalization.
Bauer, J. M., Bergstrøm, R., & Foss-Madsen, R (2021) report a randomized field study involving 1,493 users in which cookie-banner manipulations increased consent by 17 percentage points. The researchers distinguish architecture affecting whether a person makes a choice from architecture affecting which option is chosen. This decomposition is a valuable advance because an interface may raise the number of recorded decisions while also changing the distribution of those decisions. The two mechanisms should be estimated separately.
Consent research is methodologically attractive because control conditions can keep legal terms and feasible actions constant while varying placement or preselection. Yet consent is also unusually sensitive to local law, browser configuration, repeated exposure, and user fatigue. A field study should record the actual text and available controls, recruitment context, device types, prior consent history, and whether choices were real or hypothetical. The scientific claim must remain tied to that documented environment.
Subscription, Checkout, and Cancellation Systems
Commercial interfaces expose users to decisions about optional additions, renewal, commitment duration, shipping charges, and exit from recurring obligations. A checkout can make an extra product appear preselected; a subscription flow can conceal later payments or make cancellation difficult to discover. Such designs raise questions about preference formation, asymmetric effort, and whether acceptance persists after terms are accurately explained.
Experimental studies should compare the same product, price, and material terms under different visual and navigational arrangements. Where financial harm is possible, a simulated or fully reversible decision can protect participants. Outcome measures include purchase or signup, recall of essential terms, detection of optional additions, time to complete cancellation, and willingness to make the same choice when the interface is simplified. Studying a one-click conversion without checking comprehension cannot establish that the design improved the consumer's decision.
Luguri, J., & Strahilevitz, L. J (2021) provide comparative experimental evidence on interfaces intended to channel consumers toward an unfavorable offer. Their controlled studies are significant precisely because they compare designed conditions rather than merely cataloging suspicious patterns. They also illustrate that techniques differ: strong pressure can increase signups while provoking greater backlash, so an intervention's immediate conversion effect is not synonymous with long-term trust or user welfare.
A cancellation study can treat the user's prior stated desire to terminate service as a reference intention. The outcome is then whether the interface enables completion of that intention, not whether it succeeds in retaining a subscriber. This experimental framing clarifies why equal functional availability can coexist with unequal real-world accessibility. It also provides a more interpretable measure of whether an interface respects the direction of choice.
Dark Patterns as a Research Category
The label dark patterns describes interface arrangements that steer users through deception, obstruction, or other manipulative structures toward actions that may serve the operator's interests at the user's expense. Scientific classification requires greater precision than moral disapproval of a design. Researchers should specify the mechanism, the affected task, evidence concerning intention or predictable effect, and the possibility of misunderstanding or choice distortion. Not every persuasive visual element is deceptive, and not every increase in engagement is a harmful result.
Gray, C. M., Kou, Y., Battles, B., Hoggatt, J., & Toombs, A (2018) developed an influential taxonomy from a corpus of 118 interface examples. Its categories include nagging, obstruction, sneaking, interface interference, and forced action. These are design-level descriptors that support systematic coding and hypothesis generation. Their presence in an interface does not, by itself, quantify the probability of a user's behavior changing or identify a universally applicable legal category.
The distinction between taxonomy and mechanism matters. A repeated interruption can raise annoyance and increase abandonment. A misleading button label can alter the interpretation of available choices. Concealing a fee can change apparent value, while a hard-to-find cancellation control imposes effort on a previously formed decision. Classifying all four together helps detect patterns, but an explanatory model must recover the specific cognitive, motivational, or procedural path to each outcome.
A contemporary review by Bento-Silva, L. S., Cachero-Martínez, S., & García-Rodríguez (2026) examines dark-pattern scholarship in social-media consumer behavior. Such synthesis extends the domain beyond single checkout dialogs, but direct empirical tests remain necessary for any claim about mechanisms in a particular platform. Research should also record designs that appear manipulative yet have no measurable effect, because selecting only successful examples distorts scientific estimates.
Prevalence Research and Its Inferential Limits
Large-scale interface audits describe the presence and distribution of design patterns. Mathur, A., Acar, G., Friedman, M. J., Lucherini, E., Mayer (2019) inspected roughly 53,000 product pages from about 11,000 shopping websites and identified 1,818 instances of dark patterns spanning 15 types and seven categories. The authors documented instances associated with deception and commercial services offering such patterns. This provides evidence that interface manipulation is a measurable feature of online commerce rather than a collection of isolated anecdotes.
The strength of this research lies in scalable identification of observable design structures. Its limitation is equally clear: counting patterns cannot estimate how many users were affected, how much behavior changed, or how users understood those interfaces. Different website samples, detection rules, and definitions can produce different prevalence estimates. Automated crawlers may miss interaction paths, personalized variants, or designs that appear only after a particular sequence. A research claim about population exposure therefore requires information on traffic, interaction frequency, and who actually encounters each interface.
Prevalence and impact studies answer different questions. A high-prevalence design might have a small effect per exposure, while a rare but high-stakes cancellation obstruction could produce substantial harm for affected users. To estimate total consequences researchers need both exposure and causal effect, including the distribution across populations and contexts. The simple multiplication of a prevalence count by a laboratory conversion effect would be unwarranted without comparable units and valid transport assumptions.
A rigorous audit should preserve screenshots, screen recordings, interaction traces, collection dates, device configuration, account state, geographic location of the test session, and coding definitions. Independent coding and error analysis help establish the reliability of the classification. This also allows future researchers to understand changes in an interface over time instead of assuming a single screenshot represents a stable product.
Experiments on Manipulative Interface Effects
Experimental evidence can isolate effects more directly than observational taxonomies. Luguri, J., & Strahilevitz, L. J (2021) conducted two studies with census-weighted samples of American consumers. In their first experiment mild dark patterns more than doubled signup to an unfavorable service relative to the control condition, while aggressive patterns produced signup rates nearly four times as high. The authors also observed a greater backlash against aggressive designs. Their second study examined specific techniques and reported variation in impact, including stronger evidence for hidden information, trick questions, and obstruction than for some other devices.
These results establish causal effects under the study conditions, not a universal multiplier across all subscriptions or user populations. Experimental offers, stakes, participant awareness, and recruited samples affect generalizability. The backlash outcome demonstrates that a single interface can influence different endpoints in opposing directions: increased short-term signup alongside decreased trust. This makes it scientifically inadequate to describe success through conversion alone.
Privacy studies offer convergent but context-specific evidence. Nouwens, M., Liccardi, I., Veale, M., Karger, D., & Kagal, L (2020) and Bauer, J. M., Bergstrøm, R., & Foss-Madsen, R (2021) independently investigated how consent architecture shifts behavior, although their recruitment, manipulations, and outcome definitions differ. Such convergent findings support the broader proposition that some interface structures have causal effects on choices. They do not establish that every technique collected under the same taxonomic label acts through one mechanism.
The next generation of experiments should preregister specific manipulations, distinguish selection from comprehension, measure postchoice endorsement, and follow outcomes over time. A design may initially induce acceptance but later trigger withdrawal or complaint. Estimating both immediate and durable consequences provides a more useful account of interface effects than treating the first recorded click as the complete decision.
Mechanisms of Influence and the Measurement Problem
Several mechanisms may transmit interface effects: attentional selection, effort reduction, perceived recommendation, inferred norms, loss aversion, confusion, inertia, and avoidance of an obstructive interaction. A mechanism is a causal hypothesis about why a manipulation changes an outcome. To support it, researchers need more than a plausible narrative; they should measure the proposed mediator, manipulate it when possible, and examine whether the resulting pattern fits the explanation better than alternatives.
The default meta-analysis by Jachimowicz, J. M., Duncan, S., Weber, E. U., & Johnson, E (2019) helps organize explanations around how decision makers interpret preselected options. The central empirical challenge is that a person who keeps a default might be endorsing it, ignoring the decision, or trying to minimize effort. A questionnaire asked after the act may itself alter recall or rationalization. Complementary measures—attention tracking, decision time, prior preferences, comprehension tests, and experimental manipulation of advisory cues—can narrow the range of explanations.
Mechanisms interact with context. When users are under time pressure, the cost of inspecting alternatives may loom larger. In consequential decisions, a clear warning may elicit more deliberation. In repeated workflows, learned expectations may make a previously confusing control effortless. Researchers should therefore estimate moderation by task stakes, familiarity, age, accessibility needs, and device rather than assuming constant effects across populations.
A useful scientific model separates feature assignment, perceived affordance, intervening psychological processes, and outcomes. The model can express hypotheses about mediation but does not itself constitute evidence of the mechanism. This distinction protects the scientific synthesis from overinterpreting clickstream correlations as direct observations of cognition.
Meta-Analysis and the Limits of General Nudging Claims
A quantitative synthesis is valuable only when the underlying interventions, outcomes, and estimands can be interpreted meaningfully together. Mertens, S., Herberz, M., Hahnel, U. J. J., & Brosch, T (2022) combined evidence from more than 200 studies of choice-architecture interventions across behavioral domains and reported a moderate aggregate effect. This influential analysis advanced the debate by assessing many intervention classes, while also creating a difficult interpretive problem: a food-choice prompt, a retirement default, and a cookie-dialog redesign differ in setting, mechanism, and outcome. An overall estimate is therefore a summary of a literature, not the predicted effect of a particular digital control.
The dispute over publication bias is central. Maier, M., Bartoš, F., Stanley, T. D., Shanks, D. R., Harris (2022) reanalyzed the nudging evidence using publication-bias adjustments and argued that the corrected overall effect was unsupported. Mertens, S., Herberz, M., Hahnel, U. J. J., & Brosch, T (2022) replied, questioning the interpretation and generalization of those corrections. The resulting scientific disagreement concerns the evidence-selection process, sensitivity of bias models, heterogeneity, and the validity of a single average across interventions. It is not resolved by selecting whichever pooled number appears favorable to a preferred design philosophy.
Default-specific research provides a more focused, but still heterogeneous, benchmark. Jachimowicz, J. M., Duncan, S., Weber, E. U., & Johnson, E (2019) reported a sizable mean default effect alongside considerable variation between studies. That variation may reflect decisions about stakes, recommendations, user knowledge, recruitment, and outcome measurement. A digital consent study is more directly informative for a new privacy banner than the overall nudging average, yet even privacy studies differ in their operational conditions.
The defensible synthesis is hierarchical. Mechanism-specific randomized evidence supports concrete causal claims within specified contexts. Replications and systematic syntheses inform how widely those effects might extend. The most general meta-analytic statements must retain their uncertainty, study diversity, and sensitivity to analytical decisions. Effect sizes should not be presented as natural constants of user behavior.
Experimental Design and Causal Identification
The strongest basic study assigns participants or real user sessions randomly to clearly specified interface variants. Treatment might change a default while holding option descriptions constant, or move the reject control without changing its meaning. Random assignment makes baseline characteristics comparable in expectation and helps identify the effect of the manipulated design under the study conditions. The causal estimand must still be declared: the probability of selecting an option, the probability of making any active choice, the time to execute an existing intention, or a measure of preference-consistent behavior are different quantities.
Unit of randomization matters. When the same user returns repeatedly, randomizing each visit can expose the person to conflicting treatments and produce carryover. Randomizing accounts or participants may better estimate cumulative effects. If users discuss the interface or encounter shared social norms, interference between units can complicate inference. An adaptive service that modifies the displayed design based on prior clicks creates time-varying exposure, requiring more careful estimands than a one-shot A/B test.
Studies should define outcomes before examining the data, publish manipulation checks, retain all randomized units where appropriate, and report attrition and exclusion transparently. Small improvements in conversion become statistically detectable in very large platforms but may have little practical meaning; conversely, a low-frequency harmful action can be ethically significant even when an experiment has limited power. Confidence intervals, baseline rates, and absolute differences should accompany relative changes.
A critical confounding problem arises in observational comparisons: people who see different designs may differ systematically in device, country, prior engagement, or eligibility. Some platforms deliberately show more obstructive dialogs to particular segments. Regression adjustment can help describe patterns but does not automatically replicate randomized assignment. Researchers must articulate assumptions about selection, measurement, and omitted variables before making causal claims from log data.
Process Tracing, Eye Tracking, and Behavioral Logs
Choice-process research attempts to observe how information is sampled and how an action unfolds. Eye tracking can locate visual attention to prices, options, opt-out links, and disclosures. Mouse and touch logs can reveal navigation, hesitation, repeated clicks, reversal attempts, and abandonment. Screen recording can preserve the entire route through an interface. Think-aloud protocols and retrospective interviews can clarify perceived meaning, though verbalization can itself alter attention and is not a direct window into an unchanged internal process.
Each method has a different evidential status. An eye fixation demonstrates visual sampling under defined measurement conditions but does not establish that the information was understood. A cursor near a link may reflect exploration, hesitation, or an incidental movement. A repeated click on an unresponsive control may show a usability problem rather than resistance to a choice. Researchers should develop an explicit coding model linking raw signals to behavioral constructs and evaluate coding reliability across analysts.
An instructive multimethod design could measure a participant's prior choice goal, record their interaction with either a balanced or asymmetric interface, and ask them to explain the resulting decision. The combination can distinguish a user who intentionally authorizes tracking from one who clicks an accept button because the rejection route is difficult to find. Mediation analysis should be interpreted cautiously because post-treatment attention may itself depend on unmeasured factors such as motivation; experimental interventions on mediator-relevant features strengthen the argument.
Analysis of real-world logs also raises privacy and ethics concerns. Click histories and screen recordings may reveal sensitive preferences unrelated to the research question. Collecting only necessary telemetry, minimizing retention, and separating research identification from account identity are methodological safeguards as well as conditions for responsible research practice.
Construct Validity and Measures of Decision Quality
A central challenge is defining what a better choice means. A higher conversion rate measures success from an operator's perspective; it does not independently establish utility, autonomy, comprehension, or consent quality. A defensible psychological construct should be linked to observable indicators, a clearly stated measurement model, and evidence that the indicators discriminate among plausible explanations.
Decision quality can be studied through objective comprehension of consequences, consistency across equivalent presentations, concordance with preferences expressed before exposure, informed satisfaction, later endorsement, and successful completion of an intended action. These indicators are related but distinct. Preferences can be constructed during deliberation, and prior preferences may be unstable or inadequately informed. A valid assessment therefore cannot assume every change from an initial statement is manipulation. It must specify whether the study treats prior informed preferences, stable values, or logical consistency as its reference.
Recent work offers a more explicit operational strategy. Bogliacino, F., Pejsachowicz, L., Liva, G., & Lupiáñez-Villa (2026) propose a transaction test comparing whether interface modifications increase inconsistency relative to a plain-design condition. Their experiments on data-protection choices reported increases in inconsistent choices under several dark-pattern variants; the article also reports a second market experiment. The method is an important attempt to connect a normative concern about manipulation with experimentally observable inconsistency. Its results remain specific to the tested tasks, operational criterion, and designs.
Psychometrics becomes necessary where constructs such as perceived control, autonomy, decision fatigue, and trust are measured by scales. Investigators should examine reliability, dimensionality, convergent and discriminant validity, and measurement invariance across languages, devices, and user groups. A change in scale score could reflect an altered interpretation of a question rather than an altered underlying experience. Openly specifying the mapping from scale to construct makes replication possible.
Accessibility, Age, and Unequal Interface Opportunities
Affordances are relational, which implies that nominally identical interface options can provide unequal practical opportunities across users. Small controls, low contrast, unexplained icons, complex text, and tight response windows can disproportionately affect individuals with sensory, cognitive, or motor limitations. Access may also depend on language proficiency, device capabilities, connectivity, and familiarity with the type of interface. These are dimensions of the environment–person relationship and should be included in the research design.
To identify distributional effects, researchers can stratify samples, test interaction terms, and recruit participants using assistive technologies. However, subgroup comparisons require adequate statistical power and validated measures. Apparent differences between age groups may reflect education, device experience, or the tasks chosen rather than age-related capacities alone. Inclusive experimental procedures and direct usability observations provide richer evidence than generalizing from demographic labels.
A 2026 cross-sectional study by Zhang, Y., Chen, C.-F., & Hu, H.-L (2026) examined self-reported exposure to manipulative online shopping designs among 408 adults aged 60 and older. Its structural equation modeling explores associations involving cognitive overload, decision fatigue, trust bias, and digital literacy. Because the data are cross-sectional and exposure is perceived rather than experimentally assigned, the proposed mediation paths are model-based associations rather than established causal mechanisms. The study contributes plausible hypotheses and a relevant population, while leaving the causal question open.
An interface that is accessible in technical terms can still impose unequal decision costs. A screen-reader user may reach every control yet require substantially more navigation to locate an opt-out. Thus accessibility testing should include path equivalence, task completion, and successful comprehension, not only the formal presence of labeled elements.
Adaptive Interfaces and the Boundary of Personalization
Unlike a static form, an adaptive interface may reorder options or change text based on predicted user preferences, location, prior activity, or inferred vulnerability. In this setting choice architecture becomes dynamic. The design seen by one person may not be available to another, and even the same person may encounter a changing arrangement over time. This poses challenges for reproducibility because an archived screenshot describes only one state of a system.
Personalization can improve relevance and reduce unnecessary effort. It can also concentrate friction or persuasive cues on users predicted to be receptive. Evaluating these possibilities requires recording the decision policy, the features used for adaptation, and the distribution of treatments across populations. A system may maximize clicks without directly representing the user's informed preferences. A rigorous study must assess whether objective outcomes align with what the person was trying to accomplish.
Sequential experiments, controlled policy comparisons, and counterfactual evaluation can investigate adaptive interventions, but causal identification becomes difficult when the policy learns from the same behaviors it subsequently influences. Researchers should distinguish the interface's immediate presentation effect from longer-term effects of selection and feedback. General recommender ranking and habitual exposure are major adjacent research topics; this dossier focuses on adaptive presentation when it specifically structures an identifiable act of choosing.
The approach also motivates auditability. Analysts should preserve the applied treatment, decision time, interface state, and eligibility conditions, while limiting unnecessary personal data retention. Without such records researchers cannot reliably attribute observed differences to a feature or reproduce a high-stakes interface evaluation.
Human–Artificial Decision Mediation
Artificial systems increasingly mediate the presentation of choices through conversational interfaces, assistants, search summaries, and automated configuration suggestions. These interfaces can transform an explicit menu of options into a dialogue in which alternatives appear sequentially and are described in natural language. Choice architecture therefore includes which options are surfaced, how confidence is expressed, whether important conditions are omitted, and how the user can question, revise, or reject a proposed action.
An assistant can make an interface easier to navigate by translating technical conditions, showing comparisons, and preserving a record of the user's decision. It may also obscure the origin of a recommendation, treat probabilistic suggestions as certain, or make a proposed course of action feel socially endorsed. The scientific investigation should focus on observable properties: option coverage, factual accuracy, controllability, disclosure of automation, user comprehension, and successful override. Anthropomorphic impressions and trust may be measured as potential mediators, while claims about the artificial system's own psychology require separate evidence and definitions.
In the intellectual horizon of the Artificial Era, this constitutes a distinct human–artificial decision environment: humans encounter nonbiological systems capable of arranging information and proposing actions. The philosophical category can guide the selection of research questions, while empirical claims remain tied to experiments and logged interactions. A comparison between direct menu selection and conversational recommendation can measure changes in understanding, reliance, error correction, and perceived agency without assuming that either format is inherently superior.
This research intersects with human–artificial interaction as a conceptual neighbor. The present dossier retains a narrower unit of analysis: interface arrangements that influence a particular decision.
Recent Evidence on Real-World Consent and Privacy
Recent research has begun to connect short experimental treatments with repeated real-world interactions. Firullo, C., Decarolis, F., Cheyre, C., & Acquisti, A (2026) report a discussion paper using a panel of 789 United States Chrome users observed through a browser extension for up to three months. The authors report that users interacted with roughly one-third of consent banners they encountered, with substantial differences according to whether websites required a response. The working paper also explores choice stability across repeated encounters. As a discussion paper, this is research evidence with a distinct review status and should not be cited as a completed peer-reviewed journal publication.
A separate published field experiment analyzed by Athey, S., Catalini, C., Moehring, A., & Tucker, C (2026) investigates privacy choice architecture in a fintech setting involving distribution of Bitcoin to 4,494 MIT undergraduate students. Its abstract reports that minor navigation costs and reassuring but decision-irrelevant privacy information influenced privacy choices. This strengthens the case for studying procedural and informational distortions together. The particular recruitment population, technology, and decision context must be considered before extrapolation to other digital services.
Taken together with earlier randomized consent studies, these findings provide a program for ecological validation: observe repeated decisions, measure the cost of dealing with dialogs, and test whether apparently active consent reflects durable preference. Longitudinal participation and real consequences can improve external validity, but the use of browser-extension panels and service-specific experiments also introduces selection and instrumentation constraints. The most useful next studies should integrate randomized interface changes with repeated observation of comprehension and later choice revision.
Governance, Regulatory Evidence, and Normative Claims
Regulators investigate manipulative interfaces because a formally available choice may be difficult to understand or execute. The Federal Trade Commission (2022), surveys practices involving unwanted charges, undisclosed terms, difficulties canceling subscriptions, and the manipulation of privacy choices. The Organisation for Economic Co-operation and Development (2022) provides a comparative policy framework. Both are important institutional documents, though neither substitutes for a controlled experiment measuring a particular psychological mechanism.
Legal classification and psychological inference answer overlapping but different questions. A legal rule may prohibit misleading conduct or impose disclosure requirements based on specified standards, whereas a psychological experiment estimates how an interface changes behavior or understanding. A design can be unlawful without researchers having quantified a population-wide effect, and an experimentally established influence does not by itself resolve whether conduct violates a particular rule. A scientifically responsible monograph states the jurisdiction, effective date, and legal status when discussing binding requirements.
Policy is also dynamic. For example, an FTC announcement in October 2024 described a final click-to-cancel rule, but regulatory announcements must not be taken as an evergreen statement of current enforceability. Legal implementation, judicial decisions, and amendments should be verified against current primary regulatory records at the time of any subsequent legal analysis. The present research uses policy documents to identify relevant harms and assessment criteria, not to offer a comprehensive account of contemporary legal obligations.
Recent scholarship also interrogates how the concept of manipulation is operationalized. Sax, M (2026) analyzes dark patterns and autonomy from a consumer-policy perspective. The empirically oriented Bogliacino, F., Pejsachowicz, L., Liva, G., & Lupiáñez-Villa (2026) proposes experimentally assessable inconsistency, illustrating one possible bridge between normative arguments and behavioral measurement. Whether that criterion captures all relevant dimensions of unfairness remains a substantive research question.
Research Ethics and the Evaluation of User Autonomy
The ethical evaluation of interface architecture begins by specifying whose goal the design is intended to serve and what information the user needs to act on that goal. Conversion is a legitimate business metric, but it cannot by itself identify user benefit. A higher rate of permission granting may increase the operator's data access while undermining a person's previously expressed privacy preference. Conversely, a safe default can reduce accidental exposure without restricting meaningful alternatives. Context and evidence determine the character of an intervention.
Autonomy can be operationalized through the availability of alternatives, information comprehension, decisional competence in the defined task, and the practical ability to change or reverse a decision. These dimensions support transparent comparison across designs. They also prevent an overly narrow interpretation of autonomy as merely the existence of a button marked “No.” A choice can be formally selectable while imposing substantial and avoidable cost on the user.
Research ethics apply to experiments themselves. Participants should not incur unanticipated purchases, tracking, or account changes from a test without appropriate authorization. Deception, when scientifically justified and approved under applicable research standards, should be minimized and appropriately debriefed. Researchers using real platforms should consider power asymmetries, especially when sampling people who cannot easily exit a service. Where field randomization changes access to a necessary function, safeguards and independent review become particularly important.
A useful ethical test combines user-centered outcomes with distributional analysis. Investigators can ask who experiences increased understanding, who bears additional friction, whether accessibility differences amplify harm, and whether apparent benefits persist after a transparent explanation of the alternatives. This approach turns general values into concrete questions amenable to measurement while preserving the ethical dimensions that empirical measurement alone cannot settle.
A Comparative Method for Interface Evaluation
A reproducible interface study should begin with a task specification. Describe the decision maker, objective, alternatives, default, information available at each step, costs of selecting or reversing an option, and consequences of the decision. Capture the interface state precisely. Define an intervention that alters one or a manageable factorial set of parameters, along with a control that is interpretable and technically feasible.
The causal analysis should distinguish four outcomes. First, participation: whether the person notices and enters the decision. Second, selection: which available option they choose. Third, comprehension: whether they understand the action and its consequences. Fourth, endorsement: whether the result aligns with a preference or value assessed through a defensible procedure. Additional outcomes such as usability, regret, later reversal, trust, and task completion can explain longer-term effects. This framework is a synthesis of measurement needs, not a newly validated psychometric scale.
A robust research program combines laboratory control, field realism, and replication. Laboratory designs permit attention and comprehension measures but may create hypothetical incentives. Field experiments capture consequential behavior but often observe fewer internal mechanisms. Longitudinal studies show whether learned interface conventions and repeated exposures change effects, while cross-cultural and accessibility studies test transportability. Convergent results increase confidence when their assumptions and limits are explicitly recorded.
Study reports should preserve preregistered hypotheses when available, sample recruitment, exclusion decisions, device context, exposure history, software build, screenshots and interaction states, exact wording, manipulation checks, outcome definitions, absolute and relative effects, uncertainty intervals, data governance procedures, and plausible alternative explanations. For evidence reviews, authors should identify which sources were read in full and which were assessed through publisher abstracts. A critical narrative synthesis should not represent itself as a PRISMA systematic review merely because it uses many references.
Unresolved Problems and Research Priorities
The first unresolved problem concerns mechanism specificity. Many studies demonstrate that changing a design changes behavior, while fewer can isolate whether attention, inferred authority, effort, misunderstanding, or genuine preference revision transmitted that change. Careful mediator manipulation and process tracing are needed. Another problem concerns the stability of preferences: people often form preferences during deliberation, so a deviation from a pre-exposure answer cannot automatically be classified as exploitation. Researchers need normative and behavioral criteria suited to the decision.
A second challenge is transportability. Results from a national consent panel, a single commercial product, or an online experiment may not generalize across devices, cultures, disabilities, ages, regulatory regimes, and financial stakes. The cumulative evidence should preserve heterogeneity rather than hiding it under one pooled effect. Replication should include null results and designs expected to fail so the published record better represents real conditional effectiveness.
A third challenge is the adaptive interface. If a system learns which users yield to specific cues, treatment assignment is no longer independent of user characteristics or previous behavior. Transparent logging, sequential causal methods, and independent evaluation are necessary. Human–artificial interfaces add the further question of whether conversational framing alters the user's perceived set of available options before an explicit menu is ever displayed.
Finally, researchers require durable and comparable measures of decision quality. Consent rates, completion speed, satisfaction, and preference consistency capture different properties. Their divergence is not a statistical nuisance; it may be the central scientific finding. An interface that produces rapid agreement and later regret is psychologically different from one that produces a deliberate, informed, and stable choice. Scientific explanation advances when these distinctions remain visible in experimental design and theory.
Conclusion
Interface choice architecture is a measurable organization of digital decision conditions. Defaults, salience, disclosure, navigation, and friction affect what users notice, understand, and can execute. Digital affordances specify the relational possibilities for action that emerge between these structures and the capacities and aims of people encountering them. A complete explanation separates the implemented feature from the perceived opportunity, the psychological process, and the eventual outcome.
The most directly relevant experimental literature provides evidence that specific interface arrangements can change privacy consent, subscription decisions, and other online behavior. Large-scale audits establish that manipulative designs can be documented systematically. Meta-analyses demonstrate both recurring effects and major heterogeneity, with substantive disagreements over publication bias and generalization. These results support causal inquiry at the level of defined mechanisms and tasks, alongside rigorous recognition of the evidence's limits.
The scientific priority is to measure the quality and meaning of a decision as well as its occurrence. Interface evaluation should establish whether people can discover alternatives, comprehend consequences, exercise preferences, revise choices, and do so across differences in ability and context. As digital environments become adaptive and increasingly mediated by artificial systems, the organization of choice becomes a continuing subject of psychology, human–computer interaction, and the empirical study of human–artificial relations.
References
- Acquisti, A., Brandimarte, L., & Loewenstein, G. (2015). Privacy and human behavior in the age of information. Science, 347(6221), 509–514. https://doi.org/10.1126/science.aaa1465
- Arkes, H. R., & Blumer, C. (1985). The psychology of sunk cost. Organizational Behavior and Human Decision Processes, 35(1), 124–140. https://doi.org/10.1016/0749-5978(85)90049-4
- Athey, S., Catalini, C., Moehring, A., & Tucker, C. (2026). The digital privacy paradox and choice architecture: Evidence from an experiment in fintech. MIS Quarterly, 50(2), 801–816. https://doi.org/10.25300/MISQ/2025/19206
- Bauer, J. M., Bergstrøm, R., & Foss-Madsen, R. (2021). Are you sure, you want a cookie? The effects of choice architecture on users' decisions about sharing private online data. Computers in Human Behavior, 120, Article 106729. https://doi.org/10.1016/j.chb.2021.106729
- Bento-Silva, L. S., Cachero-Martínez, S., & García-Rodríguez, N. (2026). How dark patterns shape consumer behavior on social media: A systematic literature review and research agenda. International Journal of Consumer Studies, 50(5), e70240. https://doi.org/10.1111/ijcs.70240
- Bogliacino, F., Pejsachowicz, L., Liva, G., & Lupiáñez-Villanueva, F. (2026). The transaction test: An experimental method for assessing online interfaces. Journal of Economic Behavior & Organization, 244, Article 107469. https://doi.org/10.1016/j.jebo.2026.107469
- Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3), 319–340. https://doi.org/10.2307/249008
- Evans, S. K., Pearce, K. E., Vitak, J., & Treem, J. W. (2017). Explicating affordances: A conceptual framework for understanding affordances in communication research. Journal of Computer-Mediated Communication, 22(1), 35–52. https://doi.org/10.1111/jcc4.12180
- Federal Trade Commission. (2022). Bringing dark patterns to light. U.S. Federal Trade Commission. https://www.ftc.gov/reports/bringing-dark-patterns-light
- Firullo, C., Decarolis, F., Cheyre, C., & Acquisti, A. (2026). Consent and choice: Experimental evidence on user behavior at consent banners (CEPR Discussion Paper No. 22078). CEPR Press. https://cepr.org/publications/dp22078
- Gaver, W. W. (1991). Technology affordances. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 79–84). Association for Computing Machinery. https://doi.org/10.1145/108844.108856
- Gibson, J. J. (1979). The ecological approach to visual perception. Houghton Mifflin. https://search.worldcat.org/title/The-ecological-approach-to-visual-perception/oclc/4929611
- Gray, C. M., Kou, Y., Battles, B., Hoggatt, J., & Toombs, A. L. (2018). The dark (patterns) side of UX design. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems (Paper 534, pp. 1–14). Association for Computing Machinery. https://doi.org/10.1145/3173574.3174108
- Iyengar, S. S., & Lepper, M. R. (2000). When choice is demotivating: Can one desire too much of a good thing? Journal of Personality and Social Psychology, 79(6), 995–1006. https://doi.org/10.1037/0022-3514.79.6.995
- Jachimowicz, J. M., Duncan, S., Weber, E. U., & Johnson, E. J. (2019). When and why defaults influence decisions: A meta-analysis of default effects. Behavioural Public Policy, 3(2), 159–186. https://doi.org/10.1017/bpp.2018.43
- Johnson, E. J., & Goldstein, D. (2003). Do defaults save lives? Science, 302(5649), 1338–1339. https://doi.org/10.1126/science.1091721
- Luguri, J., & Strahilevitz, L. J. (2021). Shining a light on dark patterns. Journal of Legal Analysis, 13(1), 43–109. https://doi.org/10.1093/jla/laaa006
- Maier, M., Bartoš, F., Stanley, T. D., Shanks, D. R., Harris, A. J. L., & Wagenmakers, E.-J. (2022). No evidence for nudging after adjusting for publication bias. Proceedings of the National Academy of Sciences, 119(31), e2200300119. https://doi.org/10.1073/pnas.2200300119
- Mathur, A., Acar, G., Friedman, M. J., Lucherini, E., Mayer, J., Chetty, M., & Narayanan, A. (2019). Dark patterns at scale: Findings from a crawl of 11K shopping websites. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW), Article 81, 1–32. https://doi.org/10.1145/3359183
- Mertens, S., Herberz, M., Hahnel, U. J. J., & Brosch, T. (2022a). The effectiveness of nudging: A meta-analysis of choice architecture interventions across behavioral domains. Proceedings of the National Academy of Sciences, 119(1), e2107346118. https://doi.org/10.1073/pnas.2107346118
- Mertens, S., Herberz, M., Hahnel, U. J. J., & Brosch, T. (2022b). Reply to Maier et al., Szaszi et al., and Bakdash and Marusich: The present and future of choice architecture research. Proceedings of the National Academy of Sciences, 119(31), e2202928119. https://doi.org/10.1073/pnas.2202928119
- Norman, D. A. (1999). Affordance, conventions, and design. Interactions, 6(3), 38–43. https://cseweb.ucsd.edu/~goguen/courses/271sp03/jnd.html
- Nouwens, M., Liccardi, I., Veale, M., Karger, D., & Kagal, L. (2020). Dark patterns after the GDPR: Scraping consent pop-ups and demonstrating their influence. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (pp. 1–13). Association for Computing Machinery. https://doi.org/10.1145/3313831.3376321
- Organisation for Economic Co-operation and Development. (2022). Dark commercial patterns (OECD Digital Economy Papers No. 336). OECD Publishing. https://doi.org/10.1787/44f5e846-en
- Sax, M. (2026). The double darkness of dark patterns. Journal of Consumer Policy, 49, Article 34. https://doi.org/10.1007/s10603-026-09636-2
- Tversky, A., & Kahneman, D. (1981). The framing of decisions and the psychology of choice. Science, 211(4481), 453–458. https://doi.org/10.1126/science.7455683
- Weinmann, M., Schneider, C., & vom Brocke, J. (2016). Digital nudging. Business & Information Systems Engineering, 58, 433–436. https://doi.org/10.1007/s12599-016-0453-1
- Zhang, Y., Chen, C.-F., & Hu, H.-L. (2026). Choices under manipulation: Cognitive processes and behavioral biases of older adults under perceived exposure to dark patterns in online shopping. Acta Psychologica, 268, Article 107247. https://doi.org/10.1016/j.actpsy.2026.107247