Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Adaptive Digital Reward Environments, AI, and Addiction Risk

Abstract. Adaptive digital reward environments continuously alter the distribution of rewarding cues, social feedback, content, and monetized opportunities through user–system interaction. This fundamental critical narrative monograph examines how reinforcement learning, habit, incentive salience, emotion regulation, social influence, personalized recommendation, microtransactions, and responsive artificial interfaces relate to human engagement and clinically relevant risk. The evidence synthesis reconstructs the history of operant and computational accounts of reward, reviews original social-media learning experiments, controlled feed-ranking and notification interventions, algorithmic audits, prospective and intensive longitudinal research, clinical classification, and intervention evidence. Behavioral experiments establish reward-related changes in posting and measurable effects of algorithmic feed ordering, while surveys and emerging intervention studies point to potential harms and major causal-identification constraints. Existing evidence does not support inferring a clinical diagnosis from engagement frequency or the presence of artificial intelligence. The monograph evaluates measurement error, competing explanations, population heterogeneity, adaptation, long-term functioning, and ethical accountability; and specifies a causal research agenda for user-goal-aligned systems and rigorous evaluation of impairment. It is an original critical narrative synthesis rather than an independently performed systematic review, meta-analysis, or clinical study.

Keywords: adaptive digital reward environments; addiction risk; problematic digital technology use; reinforcement learning; social reward; habit formation; reward prediction error; algorithmic recommendation; personalized ranking; AI-mediated interaction; gaming disorder; ICD-11; social media; loot boxes; microtransactions; cue reactivity; self-regulation; ecological momentary assessment; platform experiments; algorithmic audits; risk prevention; Artificial Era

The Research Object: Adaptive Reward Environments and Risk

An adaptive digital reward environment is a technologically mediated setting in which the probability, timing, magnitude, visibility, or social meaning of rewarding events changes in response to a person's behavior, inferred preferences, or the behavior of others. The relevant environment includes the user interface, recommendation and ranking policies, notification system, social feedback, game economy, advertising, and pathways for stopping or continuing an activity. Artificial intelligence enters this account when machine learning is used to estimate a person's likely responses and modify what the person encounters. The scientific object of this monograph is the interaction among these adaptive arrangements, individual vulnerabilities, acquired learning, and outcomes that indicate clinically meaningful impairment or elevated risk. Neither use frequency nor the mere presence of AI constitutes addiction. The causal question is whether a specified configuration changes measurable behavior or harm relative to an appropriate counterfactual Brand et al. (2019) National Academies (2024).

A complete analysis must follow the reciprocal sequence from exposure to behavior and from behavior back to future exposure. A person's pause, click, purchase, share, return, or refusal can become an input to a ranking system. That system can change subsequent exposure even when the user's underlying preferences remain stable. In a live platform, the person and the system therefore form a feedback process whose observable dynamics can differ from a conventional experiment presenting fixed stimuli. The adaptation may improve access to valued material or systematically concentrate cues for recurrent checking. Identifying the effect requires the distribution of content actually served, the selection behavior of the individual, the learned changes in response, and the functioning that results. The system's objective function must be measured rather than inferred from a marketing description Guess et al. (2023) Milli et al. (2025).

The guiding research questions concern the learning mechanisms by which cues and variable outcomes affect return behavior; the influence of algorithmic selection and microtransactions on exposure; differences between reward responsiveness, habit, craving, and disorder; the evidence for causal platform effects; the role of adolescence and psychiatric vulnerability; and the evaluability of preventive interventions. This work is an evidence-grounded critical narrative synthesis of primary experiments, computational models, longitudinal and cross-sectional studies, randomized platform interventions, methodological reviews, and official clinical classifications. It reports existing evidence rather than a newly conducted experiment, systematic review, or meta-analysis. Every inference is constrained by the actual study design, outcome, and population Parry et al. (2021) World Health Organization (2024).

Scientific Definitions: Reward, Reinforcement, Habit, and Addiction

Reward is used in several related but different senses. In behavioral analysis, reinforcement is defined through its effect on the future probability of a response under specified contingencies. In computational learning, reward denotes an outcome signal that changes expected values or action policies; the relevant quantity may be social approval, novelty, money, information, or relief from an unpleasant state. In neuroscience, reward-related signals participate in valuation, learning, effort, and motivation. These meanings must remain operationally distinct. A surprising notification can attract attention without establishing a sustained reinforcement contingency, and an enjoyable gaming session can provide satisfaction without impaired control. Classical accounts of operant behavior and reinforcement schedules supply the historical conceptual vocabulary, while later computational and neural work supplies models of prediction and action selection Skinner (1938) Ferster & Skinner (1957) Schultz et al. (1997).

Habit refers to learned response tendencies elicited with relatively little deliberation by familiar contexts, especially when the response becomes less sensitive to current outcome values. In everyday discourse habit may also mean repeated action; experimental studies need a stricter distinction between repetition, automaticity, stimulus-response control, and goal-directed behavior. Smartphone checking was investigated by Oulasvirta and colleagues as brief inspections of rapidly changing informational content. Their work linked easily accessible informational rewards to recurrent checking and also emphasized that such behavior was often experienced as an annoyance rather than an addiction. Contemporary social-media models likewise find contributions from both reward learning and habit-like processes. A model's successful prediction of posting does not independently demonstrate clinical addiction Oulasvirta et al. (2012) Lindström et al. (2021) Turner et al. (2026).

Addiction risk in this monograph refers to the probability of emerging or worsening impairment associated with a behavior that competes with other valued activities, becomes hard to regulate, or continues despite substantial adverse consequences. It is an outcome to be measured across time, not a trait assigned to an interface. Problematic use is a broader research description that may capture distress, compulsive-feeling engagement, sleep interference, financial harm, or conflict without meeting a defined clinical threshold. This broader term is useful only when its indicators, thresholds, and validation procedures are explicit. Classification debates caution against treating a device, a service, and every activity accessed through it as interchangeable diagnostic objects Montag et al. (2021) Petry et al. (2015) World Health Organization (2024).

From Operant Conditioning to Computational Platforms

The historical foundation of adaptive reward design lies in experimental analysis of behavior and its extension to contingent reinforcement. Skinner's The Behavior of Organisms (1938) established an experimental program studying how consequences select and maintain actions. Ferster and Skinner's Schedules of Reinforcement (1957) examined patterned relationships between responding and the delivery of consequences. Variable-ratio schedules, for example, are often invoked when describing unpredictable online rewards. That analogy becomes an empirically testable hypothesis only after specifying the actual schedule, response requirement, reward probability, and changes in behavior. Many modern feeds use adaptive content selection rather than a fixed textbook schedule, and social rewards are delivered by other users as well as by platform policies Skinner (1938) Ferster & Skinner (1957).

Reinforcement learning later supplied mathematical tools for prediction and choice under uncertainty. Schultz, Dayan, and Montague's 1997 work linked dopamine neuron responses to reward prediction and unexpected outcomes, thereby helping establish a research program relating neural signals to computational prediction error. These findings do not justify the popular assertion that every digital notification produces an addictive dopamine hit. The relevant biological mechanism depends on task conditions, neural circuits, measurement methods, and species. Robinson and Berridge's incentive-sensitization theory distinguished motivational wanting from pleasure-related liking in drug addiction, offering a theoretically important caution against equating repeated pursuit with greater enjoyment. Extending that framework to particular digital behaviors requires behavioral and biological tests in those populations Schultz et al. (1997) Robinson & Berridge (1993).

With social platforms, learning and selection operate in two directions. The participant learns which actions obtain attention or useful information. The recommender learns which exposure patterns predict clicks, watch duration, or other events. This joint adaptation changes the opportunity structure for future actions. Lindström and colleagues modeled large collections of real posts with reinforcement-learning concepts and conducted a social-feedback experiment. Turner and colleagues subsequently used real posting data with a preregistered replication to distinguish reward-learning and habitual components. These studies mark progress beyond simple metaphors, while their outcome is posting behavior rather than diagnostic incidence Lindström et al. (2021) Turner et al. (2026).

Clinical Classification and the Domain of Generalization

The World Health Organization's ICD-11 recognizes gaming disorder, describing impaired control, increased priority of gaming over competing activities, and continuation despite negative consequences, together with significant functional impairment. The WHO specifies that the pattern is ordinarily observed for at least twelve months, with allowance in severe cases for a shorter period under the diagnostic requirements. Internet Gaming Disorder appeared in DSM-5 Section III as a condition for further study rather than a general diagnosis for all forms of technology use. These distinctions determine how this dossier interprets clinical evidence. A machine-selected short-video feed, a social networking application, and a game with paid randomized items must be investigated according to the behavior and harm actually measured World Health Organization (2024) Petry et al. (2015).

The term smartphone addiction is particularly susceptible to category error. Smartphones provide messaging, employment tools, games, music, maps, public services, and interpersonal relationships. High utilization of such a device provides little information about the specific behavioral process. Montag and colleagues' taxonomic analysis accordingly recommends greater precision about the online activity under study. Diagnostic labels also differ from self-report scales that include general feelings of excessive use or difficulty disconnecting. A scale can identify a high-scoring sample and still have uncertain positive predictive value for a clinical syndrome. A scientifically defensible monograph therefore tracks diagnostic cases, subclinical problems, and routine engagement as related but nonidentical outcomes Montag et al. (2021) Parry et al. (2021).

Biological and psychological theories of addiction add hypotheses about cue reactivity, craving, learning, affective regulation, and executive control. The I-PACE model describes interactions among predispositions, affective and cognitive responses, and executive processes across the development of addictive behaviors. It supplies a framework for selecting candidate mediators and moderators, not a guarantee that every relationship it predicts has been proven in digital settings. In research on social network use, investigators have found that automatic approach reactions can accompany common use while aspects of explicit wanting relate more closely to problematic use. The stronger claim that personalized algorithms cause a diagnosed addiction remains dependent on long-term exposure experiments, validated clinical endpoints, and representative samples Brand et al. (2019) Wadsley & Ihssen (2022).

Reinforcement Learning in Social Feedback Systems

Social rewards have a distinctive structure because they are generated partly by other people's responses. Likes, comments, replies, and reposts communicate attention and recognition; their distribution depends on audience size, posting context, prior social relationships, and platform visibility. Lindström and colleagues (2021) analyzed more than one million posts by more than four thousand users across several platforms, fitting computational models related to reward maximization. Their separate online experiment with 176 participants manipulated social feedback and supported a causal contribution of such feedback to posting decisions. The evidence establishes that social reward contingencies can shape behavior under the examined conditions. The work does not estimate the probability that reinforcement exposure produces a disorder, and the experimental simulation cannot capture all consequences of a long-term platform relationship Lindström et al. (2021).

Turner and colleagues (2026) modeled posts from 2,696 Twitter users and included a preregistered replication. The reported hybrid account combined reinforcement-learning and habitual components, with habitual signatures more apparent among frequent posters. Differences across age and gender groups were also reported; these observational subgroup findings call for replication with explicit tests of selection, population composition, and historical changes in the platform. Computational latent parameters depend on model specification and may map imperfectly onto everyday concepts of craving, automaticity, or loss of control. The scientifically useful advance is the operational separation of reward updating from repeating a learned response under similar conditions Turner et al. (2026).

A fundamental identification challenge is that a user may post more after receiving rewards because the rewards increased motivation, because the account became more visible, or because external social circumstances changed. Within-person sequences alone cannot distinguish these explanations. Randomizing feedback magnitude while holding post visibility constant would isolate one pathway, although ecological validity and consent become challenging. The preferred design links immediate contingencies, the subsequent action policy, and independent measurements of interference or impairment over time. In the absence of those features, reward-learning evidence informs the mechanism of engagement but supports only a conditional account of addiction risk Lindström et al. (2021) Brand et al. (2019).

Social Validation, Peer Influence, and Development

Quantified endorsement gives feedback a social comparative dimension. Sherman and colleagues (2016) used a simulated Instagram paradigm during functional magnetic resonance imaging to examine adolescents' responses to photographs displayed with different numbers of likes. Adolescents were more likely to endorse images presented as popular, including some depicting risky behavior; differences appeared in neural regions associated with reward, social cognition, and attention. This experiment shows that visible peer evaluations can influence short-term behavior and neural response to mediated social information. The research does not show that a larger number of likes causes dependency or that neural activity identifies a future clinical case Sherman et al. (2016).

Adolescence may alter the salience of affiliation, identity presentation, peer acceptance, and risk. However, adolescents differ in developmental stage, social support, vulnerability, and everyday opportunity for offline connection. A study that observes stronger response to displayed peer approval does not establish a uniform harmful effect across the adolescent population. Person-specific effects on well-being have been documented in intensive longitudinal research, suggesting that a population mean may conceal divergent psychological experiences. An adaptive platform may amplify an already important peer relationship, provide needed support, or repeatedly expose a vulnerable user to evaluative uncertainty. The relevant causal variable is the experienced reward and its consequences, not simply chronological age Sherman et al. (2016) Beyens et al. (2020) National Academies (2024).

Future developmental experiments should distinguish receiving feedback, anticipating feedback, viewing others' approval, and the opportunity to compare oneself with peers. These events have distinct cognitive and emotional functions and different consequences for behavioral reinforcement. A longitudinal design could estimate whether repeated uncertainty about peer responses predicts growing compulsive checking after controlling for baseline anxiety, depressive symptoms, offline stress, and prior checking. Such designs should also measure perceived autonomy and the possibility of leaving a platform without losing valued relationships. Clinical claims require sustained impairment in addition to evidence of feedback sensitivity Parry et al. (2021) World Health Organization (2024).

Cue Reactivity, Craving, and Contextual Triggers

A digital cue can signal an opportunity for information, contact, stimulation, monetary reward, or avoidance of discomfort. Its influence may arise from attention capture, learned expectancy, or anticipated relief rather than from a singular reward system. The I-PACE framework predicts interactions between personal predispositions, anticipated gratification or compensation, cue-induced affect, and self-regulatory capacities. In an adaptive interface, the frequency and content of cues may be selected from a history of previous behavior, making exposure endogenous to a participant's prior responsiveness. A user reporting strong urges to check after a notification demonstrates a subjective motivational state; the occurrence of an urge alone does not satisfy a diagnosis Brand et al. (2019) Montag et al. (2021).

Research on social networking sites has tested both implicit approach reactions and explicit wanting. The study of implicit approach motivation and reward in problematic social networking use reported that stronger automatic approach was associated with more frequent checking but did not characterize more problematic use in the same way; explicit wanting and reward-related experiences were more informative. This is a valuable falsification of simple assumptions that automatic attraction and pathology are synonymous. It also shows why mechanistic studies should include competing measures within one design and evaluate discriminant predictive validity Wadsley & Ihssen (2022).

Cue studies must separate immediate attention, return behavior, emotional state, and enduring functional impact. An ecological momentary assessment can ask about the cue, the person's goals before checking, subjective urgency, the nature of the material encountered, the length of the session, and any consequence for obligations or sleep. Passive logging can document app transitions and event timestamps, but the meaning of a cue still depends on context. Models should include prior mood, environmental demands, and competing tasks; otherwise a predictor interpreted as craving may actually index stress-driven coping or urgent interpersonal communication Parry et al. (2021) Oulasvirta et al. (2012).

Uncertainty, Intermittent Outcomes, and Session Continuation

Intermittent outcomes can sustain repeated responding when some actions yield salient consequences and others do not. Classical schedules-of-reinforcement experiments provide an analytic foundation for this proposition. Yet a social feed mixes many contingencies: clicking may yield a useful message; scrolling may produce humorous novelty; continuing a game may unlock progress; an online purchase may deliver an uncertain virtual item. These events differ in the control a platform exercises over outcome probability. Some are programmed directly, while others depend on social relationships or user-generated content. The phrase variable reward should therefore identify the concrete mechanism: probabilistic delivery, uncertain value, variable timing, or unpredictable social reception Ferster & Skinner (1957) Lindström et al. (2021).

Infinite scrolling and autoplay change the cost of continuing relative to stopping by reducing the need to make an explicit next-choice decision. That design can extend sessions through convenience, inertia, interest, or perceived relevance. A credible test must randomize a meaningful stopping affordance or presentation rule and measure actual continuation, without presuming that longer sessions indicate loss of control. Behavioral choice depends on what would have happened under a different interface, and stopping friction can interact with attentional load, fatigue, and user motivation. This is a plausible risk mechanism whose quantitative contribution varies across contexts and remains an empirical question rather than an established universal constant Guess et al. (2023) National Academies (2024).

The strongest study design decomposes the total exposure into components: reward outcome distributions, cue timing, user control over the stream, and algorithmic selection. Factorial experiments could evaluate these components independently and in combination; subsequent follow-up could examine changes in sleep, priorities, financial expenditure, or clinically assessed impairment. An immediate increase in watching time is mechanistically informative but insufficient to establish progression to disorder. Equally, a null effect on one immediate outcome cannot rule out longer-term adaptation. Experimental claims should state the period, platform, population, contrast, and actual outcome Dekker et al. (2025) Guess et al. (2023).

Emotion Regulation, Escape, and Negative Reinforcement

Digital engagement may be maintained by positive experiences such as amusement and approval, but also by relief from loneliness, boredom, uncertainty, or distress. Negative reinforcement describes the strengthening of a behavior when an aversive condition is reduced following that behavior. A person who checks a feed repeatedly after uncomfortable silence may be seeking reassurance; another may be responding to genuine responsibilities. The same observable act can have different learning histories and outcomes. The I-PACE model treats gratification and compensation as pathways embedded in a broader person-affect-cognition-execution interaction, encouraging measurement of the immediate function of use Brand et al. (2019).

Distress is also capable of reversing the apparent association. A student with sleep disruption may use social media longer because of preexisting insomnia; stressful social interactions may increase exposure and impairment simultaneously; an individual facing social exclusion may rely on digital communities for meaningful support. An analysis correlating screen duration with psychological symptoms without modeling these circumstances leaves causal direction open. Contemporary large-sample studies and national reviews consistently emphasize that broad use measures conceal different activities and contexts. The task is to identify an exposure-related mechanism beyond baseline distress, not simply count its co-occurrence with symptoms Orben & Przybylski (2019) National Academies (2024).

Researchers can combine ecological sampling with experimentally varied stress and recovery contexts where ethically appropriate, while distinguishing instrumental coping from loss of behavioral control. Sensitive measures should ask whether the activity served the participant's own goal, whether alternatives were available, and whether use created subsequent costs. Harmful cycles may arise when short-term relief is followed by social conflict, lost sleep, or greater avoidance, which in turn increases the appeal of further short-term relief. This feedback loop is a theoretically testable process; a particular platform should be named as a causal contributor only when longitudinal timing and a credible comparison support that inference Brand et al. (2019) World Health Organization (2024).

Recommendation Algorithms as Adaptive Exposure Policies

A recommender system estimates which material to present and in what order. The model may use observed clicks, viewing duration, reactions, prior searches, account relationships, item features, or inferred similarity across users. The resulting interface is an exposure policy: it influences which opportunities for reward or distress are encountered, the sequence in which they occur, and how much effort it takes to continue. Engagement optimization measures behaviors that are straightforward to observe, whereas welfare, deliberate preference, learning, and functional impact are substantially harder to capture. An algorithm can predict a click without determining that the click advances the person's long-term goals. This distinction is central to the difference between ranking success and human benefit Milli et al. (2025) Jahanbakhsh et al. (2026).

Personalization also creates a dynamic measurement problem. Users who are initially drawn to a category can be shown more of it; their further engagement then strengthens the system's estimate of preference. Such feedback can make exposures more concentrated over time, although the direction and clinical importance depend on the ranking objective, exploration rules, and available material. Some personalization reduces search costs and promotes useful contacts. To identify an addictive risk effect, researchers must compare otherwise comparable exposure policies and follow users long enough to observe clinically meaningful change. A snapshot of one highly engaging feed identifies content distribution, but only under a counterfactual can an attributable algorithmic effect be estimated Guess et al. (2023) Burton et al. (2024).

Algorithmic transparency in this context requires more than disclosing that machine learning is present. Researchers need operational documentation of ranking targets, event definitions, model-update intervals, the balance between exploration and exploitation, and the treatment of sensitive or vulnerable user groups. Systems can be changed during a study, complicating its reproducibility. Versioned algorithm descriptions, auditable feature logs, independent ethics approvals, and appropriate privacy safeguards are foundational to evidence that is transferable beyond one implementation National Academies (2024) Milli et al. (2025).

What Randomized Feed Experiments Establish

Guess and colleagues (2023) conducted a platform-integrated randomized experiment during the 2020 United States election. Consenting Facebook and Instagram users were assigned to reverse-chronological feeds instead of default algorithmic ranking. The intervention substantially reduced time on the platforms and user activity, while also changing exposure to particular types of content. For the examined three-month period, it did not significantly change measured political polarization or several other key attitudes. The design provides strong evidence that ranking policy can causally change engagement in a real environment. It is an experiment in feed ordering, not a direct test of addiction onset, withdrawal, or clinical recovery Guess et al. (2023).

Milli and colleagues (2025) compared engagement-ranked and reverse-chronological Twitter content in a preregistered audit involving 806 users. The study reported amplification of emotionally charged and out-group-hostile political material, along with a mismatch between some algorithmically selected content and what users said they wanted to see. Such evidence directly challenges the assumption that revealed engagement always represents endorsed preference. Yet it concerns a particular platform, historical ranking system, and political-content subset. Its relevance to addiction risk is an intermediate mechanism: an optimizing system may prioritize stimuli that induce engagement even when the experience is unwelcome, thereby motivating studies of whether similar mismatches extend to sustained compulsive behavior Milli et al. (2025).

Burton, Herzog, and Lorenz-Spreen (2024) used a controlled collaborative-filtering experiment to evaluate changes in curation metrics and their influence on beliefs. Their findings show that alternative ranking goals can generate substantively different cognitive consequences even when the broad task is similar. The experiment supplies evidence for the general significance of optimization targets, but changes in accuracy and polarization are not addiction outcomes. Across these studies, the most defensible synthesis is causal specificity: machine ranking affects identifiable exposures and actions; a claim about disorder requires an additional, separately measured causal chain Burton et al. (2024) Guess et al. (2023) Milli et al. (2025).

Human Preferences, Algorithmic Objectives, and User Control

The objectives of a recommendation engine may be heterogeneous, combining clicks, revenue, retention, content diversity, safety rules, or creator incentives. An engagement metric is often a proxy rather than a complete expression of a person's goals. This proxy is especially vulnerable to misinterpretation when a user watches something out of concern, disgust, social obligation, or fatigue. A system that minimizes the probability of stopping can improve an operator's engagement statistic while moving the user away from a desired activity boundary. Such misalignment is a testable proposition about objectives, behavior, and stated preference; it should be assessed in particular implementations rather than assumed as a property of all AI Milli et al. (2025) Jahanbakhsh et al. (2026).

Jahanbakhsh and colleagues (2026) developed ranking procedures that use users' expressed human-value priorities as controls. In two experiments involving 141 and 250 participants, they showed that participants could meaningfully distinguish value-aligned versions of their feeds from engagement-based versions. These studies demonstrate technical and behavioral feasibility for alternative objectives. They do not establish a population-level reduction in addiction, since clinical functioning and long-term compulsive use were not primary endpoints. Nonetheless, they make an important intervention target empirically accessible: the degree to which users can understand, choose, and revise the criteria governing their exposure Jahanbakhsh et al. (2026).

Evaluation should distinguish a control's presence from its usability and real consequences. A menu offering chronological order is ineffective if few users can locate it, if the setting resets, or if the alternative policy still concentrates the same harmful cues. Relevant outcomes include objective time spent, concentration of content types, satisfaction with what was encountered, willingness to return, accuracy of user understanding, and measures of perceived agency. Follow-up is necessary to determine whether early preference-aligned choices endure or become displaced by habit and convenience. The research agenda therefore treats transparency, stable controls, and exit affordances as variables for experiments rather than labels that automatically guarantee well-being Guess et al. (2023) Jahanbakhsh et al. (2026).

Loot Boxes, Microtransactions, and Financially Contingent Rewards

Microtransactions are monetary exchanges embedded within a game or platform, and randomized loot boxes supply uncertain virtual outcomes after a purchase or other action. Their functional similarity to gambling depends on payment requirements, outcome uncertainty, monetary or exchange value, marketing, and the person's behavioral history. A survey by Zendle and Cairns (2018) found a positive association between loot-box expenditure and problem gambling. The study raised legitimate concerns about a potentially vulnerable subgroup, while its observational structure could not establish whether purchases contributed to gambling problems, whether prior gambling vulnerabilities led to purchases, or whether both reflected other determinants Zendle & Cairns (2018).

Incentives in commercial digital games span several distinct types. Cosmetic purchases offer a relatively specified benefit; season passes encourage repeated participation across time; scarcity promotions generate perceived urgency; randomized paid items introduce uncertainty; and targeted offers may adjust presentation according to purchase histories. The legal and clinical implications differ across these designs and jurisdictions. A comprehensive research program must distinguish real-world expenditure, opportunities to lose money, refund rights, and the behavioral meaning of the reward. The user's age, household resources, access to payment methods, and history of gambling-related harms are potential moderators. No causal estimate of addiction can be derived merely from a platform's sale of virtual items Zendle & Cairns (2018) World Health Organization (2024).

To isolate harmful design features, an experiment might vary probability disclosure, expenditure caps, or default purchasing friction while holding game content constant. A natural experiment could examine verified sales before and after a regulatory change or a platform-level restriction, with comparable untreated groups and pre-trend checks. Researchers should measure financial harm, diminished control, and persistence of use separately from user satisfaction or game enjoyment. A clinically meaningful endpoint may be relatively rare; power planning and longer follow-up are therefore essential. The ethical imperative is especially clear when studying minors or users with preexisting gambling problems Zendle & Cairns (2018) National Academies (2024).

AI-Generated Social Interaction and Reward-Mediated Relationships

Generative AI adds a distinctive research setting: responses, style, timing, and apparent interpersonal attentiveness can be produced dynamically rather than chosen from a fixed catalog. The system can react to a user's messages, maintain conversational continuity, or supply personalized praise and reassurance. Such interactions may support knowledge acquisition, practical coping, accessibility, or companionship. They may also alter expectations of responsiveness and encourage repeated consultation. The psychological object is the user's observed relationship to an adaptive interface, including agency, perceived reciprocity, and the consequences for offline functioning. These effects require direct human-participant studies; the mere anthropomorphic appearance of a model is insufficient evidence of dependency National Academies (2024) Brand et al. (2019).

Research should distinguish engagement that is beneficial and deliberate from a repetitive pattern associated with interference, distress, or displacement. Continuous availability and personalized affirmation are candidate moderators, but claims about addiction risk from AI companionship cannot be transferred automatically from gaming-disorder research or social network experiments. A conversational system differs from both a slot-machine schedule and a social feed: its feedback is linguistically contingent, its goals may include user assistance, and its simulated interpersonal qualities can be interpreted very differently across users. A theoretically sound design must measure the interaction's function, the amount of voluntary choice, social network circumstances, and clinically relevant impairment independently World Health Organization (2024) Montag et al. (2021).

The Artificial Era introduces a deeper institutional question about whose objectives govern a learned interaction. If an artificial system is evaluated solely by session duration, the observed relationship may be optimized toward persistent contact; if it is evaluated by successful completion of a user-defined task, shorter interaction could signal better service. This comparison creates a testable design principle: reward policies should be evaluated against user-specified purposes and long-term outcomes rather than retained attention alone. It is a conceptual extension from AI-alignment and human–computer interaction research, offered here as a hypothesis and program of measurement, not as an already established causal clinical result Jahanbakhsh et al. (2026) Milli et al. (2025).

Measurement: Exposure, Problems, and Clinical Functioning

The validity of addiction-risk research rests first on its measurement model. Exposure can mean opening an application, receiving a notification, seeing a ranked item, continuing to watch, responding to a social endorsement, or purchasing an uncertain reward. These events should be recorded as separate variables with clear temporal resolution. Engagement measures such as total screen time, session length, and post frequency are useful behavioral descriptors but cannot by themselves represent craving, impairment, or a diagnosis. The best instrument depends on the hypothesis. A test of recommendation policy needs served-impression logs, while a test of impaired control needs reports or observations of unsuccessful regulation, persistence despite costs, and functional consequences Parry et al. (2021) World Health Organization (2024).

Parry and colleagues' preregistered systematic review and meta-analysis (2021) assembled 106 effect sizes comparing self-reported and logged digital media use. Self-report and logs correlated only moderately, and the agreement of problematic-use measures with usage traces was weaker. The implication is methodological, not that subjective experience is useless. Reports are indispensable for urges, motives, and distress, whereas server logs and device traces are more appropriate for reconstructing event timing. Neither source is complete: logs omit the meaning of an action, while retrospective questions are influenced by memory, self-evaluation, and response framing. Multi-method alignment requires predeclared windows and event definitions Parry et al. (2021).

A clinical measurement battery should record domain-specific symptoms, their duration and severity, potential comorbid problems, and actual interference with relationships, employment, study, finances, or sleep. When the investigated activity is gaming, diagnostic interpretation should follow the applicable ICD-11 requirements. Other digital activities require equally explicit operationalization but should not be silently labeled with the gaming-disorder diagnosis. It is appropriate to analyze dimensional problematic use alongside clinical outcomes if the constructs remain separate and the limitations of each scale are disclosed. Estimating prevalence requires a validated case definition, sampling frame, and specified region and period World Health Organization (2024) Montag et al. (2021) Petry et al. (2015).

Study Designs and Their Causal Targets

Cross-sectional surveys can identify groups with higher reported symptoms and explore candidate predictors. They cannot establish whether an adaptive interface caused those symptoms. Prospective cohorts add temporal sequence, but confounding and reverse causality persist when users select platforms and platforms select stimuli conditional on behavior. Intensive longitudinal designs repeatedly sample each person, permitting examination of change after particular exposures. However, time-varying confounders such as stressful events, fatigue, and social obligations may jointly influence both platform use and distress. Each design contributes evidence about a different estimand, and a research synthesis should retain those differences rather than ranking all findings by statistical significance alone Orben & Przybylski (2019) Parry et al. (2021).

Randomized platform experiments can identify effects of a defined change to exposure policy. The chronological-feed experiment by Guess and colleagues establishes an effect of feed ordering on time and activity during its studied period. It does not identify the effect of receiving a diagnosis or explain which elements of chronological ordering produced that change. The manipulation may alter content quality, predictability, relevance, and the effort of locating material simultaneously. Factorial interface experiments can isolate mechanisms more closely, but their simplified conditions may fail to reproduce social pressures and long-term adaptation. Trial registration, fidelity monitoring, intention-to-treat analysis, and transparent outcome definitions strengthen inference Guess et al. (2023) Dekker et al. (2025).

Quasi-experiments can exploit phased platform changes, geographic rollouts, or rule changes. Their credibility turns on whether exposure assignment is plausibly independent of unmeasured outcome changes after conditioning on an appropriate design. Difference-in-differences requires defensible counterfactual trends; instrumental approaches require an exclusion restriction; interrupted time series must separate intervention effects from concurrent changes. In adaptive systems, treatment may spill over to untreated users through content competition, peer behavior, and recommender retraining. Cluster randomization and network interference models are therefore relevant. The choice of analysis must follow the real assignment process rather than the attractiveness of a causal label National Academies (2024) Guess et al. (2023).

Algorithmic Audits and Dynamic Identification

An algorithmic audit can compare the items delivered under different ranking conditions, estimate selection patterns, and evaluate the divergence between algorithmic predictions and user-articulated preferences. Milli and colleagues compared ranking outputs with reverse-chronological baselines, while Jahanbakhsh and colleagues tested value-responsive alternatives. Such designs offer important access to the policy layer between content inventory and lived experience. However, an audit of material selected for display does not measure the full psychological effect of being exposed to that material. Exposure, attention, interpretation, later action, and long-term consequences form separate stages. A defensible audit identifies which stages were actually observed Milli et al. (2025) Jahanbakhsh et al. (2026).

Personalized systems violate a convenient assumption of static treatment. An intervention at one point in time may alter subsequent recommendations because the model updates on the very responses affected by the intervention. User-level histories can influence later choices, and changes for one user may change distribution for other users. Researchers must specify whether they estimate an initial ranking effect, a sustained policy effect, or the total effect of a recursively adapting system. Off-policy evaluation using logged propensities may help under appropriate support and identification assumptions, but opaque systems and unrecorded decisions severely constrain what can be recovered. An analysis that omits model versions and exploration probabilities cannot reliably reconstruct the exposure mechanism Guess et al. (2023) National Academies (2024).

AI-driven safety evaluations also require challenge tests beyond average engagement. Researchers may probe simulated and consented trajectories of repeated use, distinguish high-vulnerability from low-vulnerability strata, and evaluate whether the system responds appropriately to explicit attempts to stop. Performance should be assessed for unintended exposure to gambling-like rewards, compulsive checking triggers, self-reported regret, and the reliability of access controls. Such tests can establish local properties of the system, but clinical claims require corroboration in actual populations with appropriate safeguards. Algorithmic behavior and human disorder are distinct levels of observation linked by an empirical research program Brand et al. (2019) National Academies (2024).

Statistical Quality, Reproducibility, and Heterogeneity

Evidence quality depends on more than sample size. Exposure measurement error can attenuate associations, selective attrition can distort longitudinal patterns, and uncorrected repeated testing can generate spurious findings. A large sample may detect associations that explain little variance in well-being. Orben and Przybylski (2019) used specification-curve analysis across three large adolescent datasets and found a small negative association between digital technology use and well-being. The result illustrates both the importance of analytical flexibility and the limits of broad digital-use measures for an addiction-specific question. Findings must be presented with their scale, design, uncertainty, and relevance to the actual construct Orben & Przybylski (2019).

Randomized trials should report assignment mechanisms, sample composition, adherence, missing outcomes, effect estimates, confidence intervals, and adverse events. Computational studies should report their model alternatives, out-of-sample predictive tests, parameter recovery where possible, and robustness to platform drift. The 2026 posting-model study offers a preregistered replication but also publicly notes a deviation from preregistration. Such transparency matters because a successful fitted model can be structurally underidentified and still forecast some behavior well. Neural measures add another layer of validation but do not transform a correlation into a clinical marker without prospective validation Turner et al. (2026) Sherman et al. (2016).

Heterogeneous treatment effects are a central problem. Different users have different rewards, social constraints, available alternatives, developmental needs, and baseline harms. Beyens and colleagues (2020) documented person-to-person variability in social-media associations with well-being, offering a reason to avoid assuming one uniform average effect. Nonetheless, subgroup searches can themselves generate false findings, especially when groups are small and hypotheses were not prespecified. A cumulative science of adaptive reward risks requires large and diverse longitudinal cohorts, shared but privacy-sensitive measures, independent replication, and transparent distinction between exploratory and confirmatory interaction tests Beyens et al. (2020) National Academies (2024).

Evidence Appraisal Across Research Traditions

A synthesis of this field must maintain an explicit chain from claim to design. Laboratory feedback experiments show that rewards and social cues can influence immediate behavior. Computational analyses show that certain action histories are compatible with reinforcement-learning and habit components. Platform feed trials show that changing ranking can alter usage under real-world conditions. Surveys link randomized purchases to gambling problems. Clinical classification defines only particular syndromes under specified impairment thresholds. None of these study families alone supplies the full causal sequence from an adaptive algorithm to a new diagnosis; convergence across them may progressively strengthen particular mechanism claims Lindström et al. (2021) Turner et al. (2026) Guess et al. (2023) Zendle & Cairns (2018) World Health Organization (2024).

Positive evidence must be weighed alongside null or contrary results. Dekker and colleagues' preregistered randomized trial with 205 participants found that a one-week notification-disabling intervention did not reduce objectively logged screen time or checking frequency, although perceived checking habit weakened and fear of missing out increased. This finding directly challenges the assumption that removing a cue always reduces use. The trial's sample, period, and manipulation define the scope of inference. It does not imply that every notification setting is inconsequential or that other controls are ineffective Dekker et al. (2025).

For this dossier, confidence is strongest for well-described proximal mechanisms such as social feedback effects on posting and algorithmic feed effects on time spent, and more conditional for direct causal effects on sustained clinical impairment. The distinction is not a declaration that harms are absent. It is a disciplined description of which links in the proposed causal chain have been directly observed. Significant risks can justify precautionary design and better research while their magnitude and prevalence are still uncertain. Future work must test cumulative exposure and clinically relevant endpoints rather than simply extrapolate from immediate engagement National Academies (2024) Guess et al. (2023) Balhara et al. (2026).

Platform Design and Primary Prevention

Prevention in adaptive environments can address the distribution of cues, the intensity of incentives, the accessibility of stopping, and the transparency of choices. Candidate measures include persistent chronological ranking options, meaningful notification controls, session boundaries that remain visible, purchase limits, accurate probability information for randomized rewards, and restrictions on targeting vulnerable users. Each candidate must be tested against realistic alternatives. A design that lowers total engagement might also reduce access to supportive communities or essential communication, while a measure that users value may leave some objective harms unchanged. Evaluation therefore requires predefined endpoints beyond total time spent National Academies (2024) Jahanbakhsh et al. (2026).

The randomized notification study by Dekker and colleagues is instructive. Disabling notifications for a week altered perceived checking habit strength but did not materially alter logged screen time or checking frequency, and increased fear of missing out. An intervention that successfully removes a technical cue may change appraisal without altering total use, because users continue to check for social or informational reasons. This is a reminder to distinguish a mechanism's proposed direction from its actual causal effect and to include unintended consequences when evaluating supposedly protective defaults Dekker et al. (2025).

A risk-sensitive prevention study could randomize users to stable ranking controls, explicit session boundaries, or staged combinations of both, with objective logs and repeated assessments of impairment-related outcomes. Such interventions must respect informed consent and avoid experimenting on high-risk users with intentionally more harmful designs. A population prevention strategy should also evaluate access inequities: people who rely on digital tools for disability support, caregiving, employment, or marginal community connection may bear distinct costs when access is constrained. Meaningful prevention preserves the benefits of connection while reducing modifiable exposure risks National Academies (2024) Beyens et al. (2020).

Digital Interventions, Clinical Evidence, and Treatment Boundaries

Treatment research addressing problematic digital technology use is heterogeneous in definitions, settings, and interventions. Balhara and colleagues' 2026 systematic review and meta-analysis evaluated therapeutic interventions across several domains of problematic technology use, finding evidence favoring structured psychological approaches in some categories while emphasizing high heterogeneity, small-study effects, and limited evidence in several emerging behaviors. A pooled result that combines different scales and populations should not be treated as a universal treatment effect for algorithmically reinforced use. The review informs clinical research priorities and supports the need for more robust controlled trials Balhara et al. (2026).

A small behavioral study published in 2023 used a package intervention involving contingent goals, app-based notifications of use, and alternative activity selection with nine university students screening high on a modified problematic-use scale. Daily use fell to or below targets across participants, while inference was limited by sample size and the intervention package design. This is a promising demonstration of measurable behavioral modification, not evidence of cure for a formally diagnosed disorder. It is also impossible to attribute the combined intervention's effects to one component without additional conditions. Risk research should separate duration change, symptom change, functioning, maintenance, and adverse consequences Stinson & Dallery (2023).

Calvert and colleagues' 2025 cohort study included a one-week social-media detox intervention and reported improvements in several symptom measures after the detox. Because the relevant intervention analysis did not supply the same causal protection as a randomized parallel-group comparison, changes could include regression to the mean, expectancy, contextual events, and other time effects. Such results are relevant to feasibility and hypothesis generation, yet must be interpreted in their actual design. A clinical translation program should use randomized comparisons, appropriate follow-up, validated disorder-specific assessments, and ethically recorded harms. The present monograph provides research analysis rather than individual therapeutic recommendations Calvert et al. (2025) Balhara et al. (2026).

AI-Assisted Risk Prediction and Personalized Support

Machine learning can combine repeated assessments, device logs, context indicators, and self-reports to predict elevated use or future impairment. Prediction and causation are different scientific tasks. A model may predict who is likely to spend additional hours online by identifying a person's previous engagement intensity, yet reveal little about which design change would improve that person's functioning. Prediction models should therefore specify the horizon, target variable, prospective validation population, missing data, calibration, and distribution shift. A high overall discrimination score can conceal poor performance in a vulnerable subgroup. The ethical cost of false positive labels is considerable when a model attributes an addiction risk to everyday social behavior Parry et al. (2021) National Academies (2024).

AI-assisted supports could offer personally chosen limits, explanations of recommendation patterns, or help tracking whether current use serves a person's priorities. Their evaluation must include whether the system increases surveillance, dependency, stigma, or false reassurance. The presence of a conversational interface does not establish therapeutic safety or efficacy. Automated advice may also vary over time as software and model behavior change. Research requires versioned systems, external outcome measures, clear user consent, and escalation pathways for clinically important harm in the jurisdiction where a service operates Balhara et al. (2026) National Academies (2024).

An especially important evaluation compares two individualized policies: one optimized to forecast continued activity and one optimized to support a precommitted user goal such as finishing a learning task or obtaining needed information. The second system might intentionally end an interaction early, a behavior that an engagement-only objective could undervalue. The empirical question is whether alignment with explicit user goals improves sustained functioning, reduces regret and interference, and maintains legitimate benefits. Current work on value-aligned feed ranking provides a practical feasibility foundation, while a reduction in addiction incidence remains a future clinical hypothesis Jahanbakhsh et al. (2026) Milli et al. (2025).

Ethics, Youth Safety, and Public Accountability

Research on addiction risk concerns sensitive behavioral histories, financial transactions, potentially identifiable social relationships, and clinical vulnerabilities. Audit protocols should minimize collected personal data, make consent comprehensible, restrict reuse, and establish independent governance for participant safety. Platform log access can be scientifically valuable but also creates asymmetries between corporate and independent researchers. Replication must balance confidentiality against verification through controlled access, deidentified aggregates, code, preregistration, and published analysis decisions. National Academies recommendations emphasize platform accountability and improved research access as conditions for stronger evidence National Academies (2024).

Youth safety warrants separate evaluation because age changes the practical circumstances of consent, financial access, peer reward, and ability to refuse platform participation. Measuring risks across adolescents requires care with developmental stage, family and school contexts, and the possible benefits of online community. Policies should be evaluated for safety outcomes, educational participation, communication, and unequal exclusion rather than success measured only by lower hours online. A design can change adolescent experience through mechanisms that adults do not encounter, but those differences must be demonstrated and quantified rather than presumed Sherman et al. (2016) Beyens et al. (2020) National Academies (2024).

The most consequential accountability question concerns who bears the cost of optimizing engagement. Operators observe high-frequency behavior and can modify reward exposure continuously; individuals may have less access to these measurements and limited visibility into the rules. This asymmetry supports transparent, privacy-respecting independent audits. Empirical standards should permit comparison of multiple user-benefit objectives, including avoidance of financial loss, preservation of sleep, voluntary stopping, and user-specified values. Such measures tie interface governance to outcomes that matter psychologically, enabling a research conversation about design duties without conflating policy recommendations with experimentally established clinical effects Milli et al. (2025) Jahanbakhsh et al. (2026) National Academies (2024).

Interdisciplinary Implications for the Artificial Era

The Artificial Era places learned human behavior in environments where another adaptive system continuously selects what the human encounters. This adds a reciprocal computational layer to familiar theories of conditioning, social influence, and self-regulation. The person's actions are observations for a model; the model's outputs alter subsequent human opportunities; and the resulting behavioral history becomes data for further updates. The study of digital reward risk therefore connects psychology, reinforcement learning, human-computer interaction, clinical science, and institutional governance. Every layer needs distinct evidence and operational definitions Lindström et al. (2021) Guess et al. (2023) Turner et al. (2026).

Psychology of Human–Artificial Relations is directly relevant when AI systems mediate interpersonal communication, recommend rewarding material, or create responsive conversational encounters. Psychology of Artificial addresses the organization and policies of artificial systems as research objects, while human addiction risk remains a question about human behavior, functioning, and clinical outcomes. A theory of human–artificial interaction can propose how objectives and feedback coevolve, but that model should be evaluated through observable predictions rather than inferred from philosophical terminology alone. The independent system-level description complements individual-level behavioral measurement Milli et al. (2025) Jahanbakhsh et al. (2026).

The synthesis points to a concrete research program: compare engagement-centered objectives with explicit user-purpose objectives; estimate how adaptation changes reward exposure over weeks or months; model which individual differences alter the result; and examine whether short-term behavioral changes accumulate into clinically important impairment. These questions preserve the advantages of artificial mediation, including customized learning and timely assistance, while directing scientific attention to preventable harm. They also demonstrate why a mature psychology of the Artificial Era must investigate the architectures arranging behavior, as well as the person whose behavior is measured National Academies (2024) Brand et al. (2019).

A Causal Research Program for Adaptive Digital Environments

A coherent causal research program begins with an explicit model of the exposure-generating system. At minimum, the model should represent a person's prior vulnerabilities and goals, the platform's recommendation or incentive policy, the sequence of cues and outcomes, subsequent behavioral choices, and clinically relevant changes in functioning. Prior choices influence what is served next, which means conventional one-time exposure measurements can miss the mechanism of accumulation. Investigators should state whether their intervention changes a visible interface, a ranking objective, a reward contingency, a social-feedback display, or an opportunity to disengage. The counterfactual should preserve all other features that are practical to hold constant Guess et al. (2023) Lindström et al. (2021) National Academies (2024).

An exemplary longitudinal study would recruit a sufficiently diverse cohort, collect consented baseline measures of use purposes and functioning, randomize a clearly defined risk-reducing platform setting, and follow objective exposure logs together with brief repeated experience samples. Follow-up should estimate both immediate behavior and sustained consequences for sleep, social obligations, financial losses, mood, and impairment. Randomization establishes an assigned-policy contrast, while mediator analyses require separate assumptions and, where feasible, nested randomization. Because users interact in social networks, design and analysis should explicitly consider interference between groups and spillovers. Attrition and treatment switching must be reported, since the users who are most affected may be least likely to complete follow-up Guess et al. (2023) Parry et al. (2021).

Clinical case ascertainment should be undertaken with instruments and assessment procedures specific to the behavior under study. Gaming outcomes can be connected to established international diagnostic requirements, while emerging questions about algorithmically mediated social media use should be studied with transparent dimensional indicators of impairment until suitable diagnostic validity has been established. A model should avoid both extreme overdiagnosis and dismissal of observable harm. It should state how many participants are affected, how large the effect is, which uncertainty intervals remain, and how results vary across meaningful social and developmental contexts. A research program that treats these as separate endpoints can clarify where platform adaptation is most consequential World Health Organization (2024) Petry et al. (2015) Brand et al. (2019).

Competing Explanations and Falsifiable Predictions

Several mechanisms could yield similar digital engagement histories. A reinforcement-learning account predicts changes in choice following rewarding outcomes and their prediction errors. A habit account predicts repetition that is comparatively insensitive to changed outcome values. An information-seeking account predicts behavior that responds to uncertainty and the instrumental value of new information. A social-needs account predicts sensitivity to affiliation, exclusion, and the accessibility of valued relationships. An emotion-regulation account predicts use tied to affective states and relief. These models are compatible in some circumstances, but each makes distinguishable predictions when the researcher manipulates rewards, devalues an outcome, changes uncertainty, or varies social context Turner et al. (2026) Brand et al. (2019) Oulasvirta et al. (2012).

The claim that personalization itself causes clinically meaningful loss of control can be tested against alternate explanations. If changing the recommendation policy has little effect on persistent engagement or impairment once access costs and content availability are held constant, the platform-adaptation hypothesis loses support for that particular setting. If exposure concentration changes but use-related harms do not, the causal chain must be revised. If a user's own stated goals better predict discontinuation than the algorithm's estimated engagement preferences, alignment may matter more than raw predictive accuracy. Negative results are scientifically informative when the comparison is credible and the outcomes were measured precisely Guess et al. (2023) Milli et al. (2025) Jahanbakhsh et al. (2026).

Another important challenge concerns time scale. A single laboratory session can identify an immediate choice process, whereas a clinical disorder is evaluated through persistent patterns and functional impairment. Ecologically valid micro-randomized trials may test rapid feedback effects, while large prospective studies establish trajectories and symptom persistence. These designs should be linked through explicit intermediate outcomes rather than by loose analogy. A computational model that explains millions of short actions may be valuable even if it predicts little about later clinical outcomes. The next advance will be to estimate the transfer between these levels and to quantify where the transfer fails Lindström et al. (2021) Turner et al. (2026) World Health Organization (2024).

Unresolved Questions, Evidence Gaps, and Priorities

The first unresolved question is how much risk is causally attributable to adaptive personalization after accounting for content, social obligations, user selection, and initial vulnerabilities. The current evidence directly demonstrates effects on exposures and engagement, whereas long-term clinically diagnosed outcomes remain less well established. Second, the field needs better evidence on which design characteristics combine into harmful feedback cycles: outcome uncertainty, reward salience, low-friction continuation, social evaluation, personalized prompts, and monetary stakes may interact in ways that studies of one feature do not capture. Third, clinical outcomes should be measured across diverse platforms, developmental stages, and cultural contexts rather than inferred from a narrow sample of highly connected users Guess et al. (2023) National Academies (2024) Zendle & Cairns (2018).

Fourth, AI-enabled conversational systems, adaptive agents, and generative media introduce new forms of contingent response whose psychological effects remain incompletely characterized. Such systems should be studied through task purposes, observed reciprocity cues, return motivation, user autonomy, social context, and interference with valued life activities. Fifth, operator incentives and public safety objectives may diverge. The empirical challenge is to design ranking evaluations that include both user-valued outcomes and independently meaningful harm indicators, with clear governance over sensitive logs. Sixth, researchers should examine whether personally chosen constraints and alternate reward opportunities produce durable improvements without reducing access to education, support, and needed relationships Jahanbakhsh et al. (2026) Milli et al. (2025) National Academies (2024).

Finally, reproducibility requires more than open statistical code. A study must preserve the version and behavior of the deployed algorithm, sampling and consent criteria, the definition of every observable event, and the path from exposure to reported outcome. Independent collaborations may need secure facilities for replicating analyses on protected platform data. Predeclared analysis plans, complete reporting of null effects, and measurement of unintended outcomes will allow results to accumulate rather than become a series of incomparable screen-time correlations. Reliable knowledge of addiction risk depends on sustained methodological infrastructure as much as on theoretical innovation Parry et al. (2021) Dekker et al. (2025) Balhara et al. (2026).

Scientific Synthesis: Risk Emerges from Dynamic Person–System Relations

The assembled evidence supports a layered account of adaptive digital reward risk. At the first layer, learning and motivation respond to outcomes and their predictive cues. At the second, social networks assign interpersonal value to endorsement, contact, and visibility. At the third, algorithms control the timing and composition of some opportunities and learn from previous human responses. At the fourth, personal resources, distress, developmental needs, and external obligations influence why a person engages and whether stopping is feasible. At the fifth, sustained problems may appear as financial harm, sleep disruption, conflict, diminished control, or other functional impairments. A sound causal explanation connects these layers with independently measured variables rather than assuming a pathway from the technical presence of AI to a clinical syndrome Brand et al. (2019) Guess et al. (2023) National Academies (2024).

The strongest experimental evidence in the present selection concerns proximal mechanisms: social feedback can change posting behavior; ordering a feed can change exposure and time spent; particular peer approval cues influence adolescents' immediate responses; and notification removal does not invariably reduce checking. These are meaningful and in some cases causally well-supported findings. Evidence that adaptive personalization, considered as a broad category, directly causes a defined addiction remains less conclusive because many studies lack valid diagnostic endpoints, adequate follow-up, or isolation of the relevant design feature. This assessment neither dismisses harmful exposures nor converts suggestive correlations into causal certainty. It establishes the work needed to move from a mechanism's plausibility to a reliable estimate of population harm Lindström et al. (2021) Sherman et al. (2016) Guess et al. (2023) Dekker et al. (2025).

A practical consequence is the scientific priority of user-purpose alignment. A system should be evaluated by whether its policy makes it easier for a user to accomplish explicitly chosen goals, maintain control over stopping, and avoid measurable harm while retaining wanted social and informational benefits. The proposed evaluation is testable through randomized objectives and longitudinal outcomes. It joins psychology with the engineering of artificial systems while respecting distinct levels of evidence. The resulting subject belongs simultaneously to addiction science, media psychology, computational learning, human factors, and the psychology of human–artificial relations Milli et al. (2025) Jahanbakhsh et al. (2026) National Academies (2024).

Conclusion

Adaptive digital reward environments are best understood as dynamic arrangements in which consequential feedback, predictive cues, social valuation, monetary incentives, and machine-generated exposure policies shape opportunities for human action. Their psychological effects emerge through the interaction of learning histories, current goals, social circumstances, and the evolving platform policy. The evidence establishes that some social feedback contingencies influence engagement and that changing algorithmic feed arrangements can alter measurable behavior. It also establishes the limits of generalizing from those outcomes to clinically diagnosed addiction. Meaningful risk assessment requires validated functional outcomes, transparent case definitions, and designs capable of identifying the relevant intervention contrast Lindström et al. (2021) Guess et al. (2023) World Health Organization (2024).

The research field has progressed from laboratory reinforcement schedules to computational accounts of social media behavior and randomized experiments within operating platforms. This progression creates new opportunities to compare learning, habit, attention, emotion regulation, and institutional optimization as mechanisms of engagement. The most urgent gaps involve long-term clinical outcomes, algorithmic adaptation across time, vulnerable populations, paid uncertain rewards, user autonomy, and the capacity for independent replication. Prevention and intervention should be judged through their observed consequences for choice, impairment, and valued functioning, including potential unintended harms Turner et al. (2026) Zendle & Cairns (2018) Dekker et al. (2025) Balhara et al. (2026).

The central scientific conclusion is that addiction risk in adaptive digital environments is a relational and longitudinal research problem involving persons, interfaces, algorithms, institutions, and contexts. The Artificial Era gives this problem new technical form, because systems can continually update the opportunities through which behavior is reinforced and regulated. A mature science of these environments must assess the specific policies that shape exposure and must distinguish prediction of engagement from the welfare of the person engaging. This approach offers a coherent foundation for interdisciplinary research, defensible clinical boundaries, transparent platform evaluation, and evidence-grounded design Brand et al. (2019) Milli et al. (2025) National Academies (2024).

References

  • Balhara, Y. P. S., Bhattacharjee, O., Bhatia, R. K., Sanahan, R., Ganesh, R., Sarkar, S., Ranjan, R., & Kattimani, S. (2026). Therapeutic interventions targeted at problematic use of digital technology: Systematic review and meta-analysis of evidence. JMIR Mental Health, 13, e89280. https://doi.org/10.2196/89280
  • Beyens, I., Pouwels, J. L., van Driel, I. I., Keijsers, L., & Valkenburg, P. M. (2020). The effect of social media on well-being differs from adolescent to adolescent. Scientific Reports, 10, Article 10763. https://doi.org/10.1038/s41598-020-67727-7
  • Brand, M., Wegmann, E., Stark, R., Müller, A., Wölfling, K., Robbins, T. W., & Potenza, M. N. (2019). The Interaction of Person-Affect-Cognition-Execution (I-PACE) model for addictive behaviors: Update, generalization to addictive behaviors beyond internet-use disorders, and specification of the process character of addictive behaviors. Neuroscience & Biobehavioral Reviews, 104, 1–10. https://doi.org/10.1016/j.neubiorev.2019.06.032
  • Burton, J. W., Herzog, S. M., & Lorenz-Spreen, P. (2024). Simple changes to content curation algorithms affect the beliefs people form in a collaborative filtering experiment. In Proceedings of the 46th Annual Meeting of the Cognitive Science Society (pp. 811–818). Cognitive Science Society. https://escholarship.org/uc/item/5cj075dp
  • Calvert, E., Cipriani, M., Dwyer, B., Lisowski, V., Mikkelson, J., Chen, K., Flathers, M., Hau, C., Xia, W., Castillo, J., Dhima, A., Ryan, S., & Torous, J. (2025). Social media detox and youth mental health. JAMA Network Open, 8(11), e2545245. https://doi.org/10.1001/jamanetworkopen.2025.45245
  • Dekker, C. A., Baumgartner, S. E., Sumter, S. R., & Ohme, J. (2025). Beyond the buzz: Investigating the effects of a notification-disabling intervention on smartphone behavior and digital well-being. Media Psychology, 28(1), 162–188. https://doi.org/10.1080/15213269.2024.2334025
  • Ferster, C. B., & Skinner, B. F. (1957). Schedules of reinforcement. Appleton-Century-Crofts. https://www.bfskinner.org/publications/books/
  • Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H., Brown, T., Crespo-Tenorio, A., Dimmery, D., Freelon, D., Gentzkow, M., González-Bailón, S., Kennedy, E., Kim, Y. M., Lazer, D., Moehler, D., Nyhan, B., Velasco Rivera, C., Settle, J., Thomas, D. R.,. Tucker, J. A. (2023). How do social media feed algorithms affect attitudes and behavior in an election campaign? Science, 381(6656), 398–404. https://doi.org/10.1126/science.abp9364
  • Jahanbakhsh, F., Zhao, D., Piccardi, T., Robertson, Z., Epstein, Z., Koyejo, S., & Bernstein, M. S. (2026). Value alignment of social media ranking algorithms. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (Article 1401, pp. 1–26). Association for Computing Machinery. https://doi.org/10.1145/3772318.3791281
  • Lindström, B., Bellander, M., Schultner, D. T., Chang, A., Tobler, P. N., & Amodio, D. M. (2021). A computational reward learning account of social media engagement. Nature Communications, 12, Article 1311. https://doi.org/10.1038/s41467-020-19607-x
  • Milli, S., Carroll, M., Wang, Y., Pandey, S., Zhao, S., & Dragan, A. D. (2025). Engagement, user satisfaction, and the amplification of divisive content on social media. PNAS Nexus, 4(3), pgaf062. https://doi.org/10.1093/pnasnexus/pgaf062
  • Montag, C., Wegmann, E., Sariyska, R., Demetrovics, Z., & Brand, M. (2021). How to overcome taxonomical problems in the study of Internet use disorders and what to do with “smartphone addiction”? Journal of Behavioral Addictions, 9(4), 908–914. https://doi.org/10.1556/2006.8.2019.59
  • National Academies of Sciences, Engineering, and Medicine. (2024). Social media and adolescent health. National Academies Press. https://doi.org/10.17226/27396
  • Orben, A., & Przybylski, A. K. (2019). The association between adolescent well-being and digital technology use. Nature Human Behaviour, 3, 173–182. https://doi.org/10.1038/s41562-018-0506-1
  • Oulasvirta, A., Rattenbury, T., Ma, L., & Raita, E. (2012). Habits make smartphone use more pervasive. Personal and Ubiquitous Computing, 16(1), 105–114. https://doi.org/10.1007/s00779-011-0412-2
  • Parry, D. A., Davidson, B. I., Sewall, C. J. R., Fisher, J. T., Mieczkowski, H., & Quintana, D. S. (2021). A systematic review and meta-analysis of discrepancies between logged and self-reported digital media use. Nature Human Behaviour, 5, 1535–1547. https://doi.org/10.1038/s41562-021-01117-5
  • Petry, N. M., Rehbein, F., Ko, C.-H., & O'Brien, C. P. (2015). Internet Gaming Disorder in the DSM-5. Current Psychiatry Reports, 17, Article 72. https://doi.org/10.1007/s11920-015-0610-0
  • Robinson, T. E., & Berridge, K. C. (1993). The neural basis of drug craving: An incentive-sensitization theory of addiction. Brain Research Reviews, 18(3), 247–291. https://doi.org/10.1016/0165-0173(93)90013-P
  • Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. https://doi.org/10.1126/science.275.5306.1593
  • Sherman, L. E., Payton, A. A., Hernandez, L. M., Greenfield, P. M., & Dapretto, M. (2016). The power of the like in adolescence: Effects of peer influence on neural and behavioral responses to social media. Psychological Science, 27(7), 1027–1035. https://doi.org/10.1177/0956797616645673
  • Skinner, B. F. (1938). The behavior of organisms: An experimental analysis. Appleton-Century. https://www.bfskinner.org/publications/books/
  • Stinson, L., & Dallery, J. (2023). Reducing problematic social media use via a package intervention. Journal of Applied Behavior Analysis, 56(2), 323–335. https://doi.org/10.1002/jaba.975
  • Turner, G., Gunschera, L. J., Subrahmanya, S., Salecha, A., Eichstaedt, J. C., Palminteri, S., & Orben, A. (2026). A computational model of reward learning and habits on social media. Nature Communications, 17, Article 7170. https://doi.org/10.1038/s41467-026-73547-6
  • Wadsley, M., & Ihssen, N. (2022). The roles of implicit approach motivation and explicit reward in excessive and problematic use of social networking sites. PLOS ONE, 17(3), e0264738. https://doi.org/10.1371/journal.pone.0264738
  • World Health Organization. (2024). Clinical descriptions and diagnostic requirements for ICD-11 mental, behavioural and neurodevelopmental disorders. World Health Organization. https://www.who.int/publications/i/item/9789240077263
  • Zendle, D., & Cairns, P. (2018). Video game loot boxes are linked to problem gambling: Results of a large-scale survey. PLOS ONE, 13(11), e0206767. https://doi.org/10.1371/journal.pone.0206767