RESEARCH DOSSIER
Human–Artificial Interaction
Abstract. Human–artificial interaction is a bounded sequence of exchanges through which human participants and artificial systems alter each other's immediate informational and practical conditions. This fundamental research monograph establishes the scientific object, history, theoretical foundations, operational constructs, experiments, and evidentiary limits of this field. It integrates situated-action theory, conversational grounding, social-attribution research, automation and trust studies, explanatory interfaces, collaborative decision-making, embodied robotics, and AI-mediated interpersonal communication. Critical analysis distinguishes interaction from enduring relationships, augmentation from synergy, human judgments of agency from system capabilities, and authorial Artificial Era hypotheses from empirically established results. Evidence from experimental studies and a preregistered meta-analysis indicates that beneficial human augmentation does not automatically entail complementary team superiority. The monograph specifies reproducible methods for sequential logs, controlled perturbations, validity assessment, longitudinal testing, and accountable deployment, and closes with a research agenda for human–artificial coordination.
Keywords: human–artificial interaction; human–AI interaction; human-computer interaction; human-robot interaction; reciprocity; situated action; conversational grounding; anthropomorphism; automation reliance; trust calibration; explainable AI; human–AI collaboration; interaction sequence analysis; Artificial Era
Human–Artificial Interaction as a Scientific Object
Human–Artificial Interaction denotes an observable, bounded process in which a human participant and an artificial system exchange actions, signs, information, or control signals, and in which the response of either participant changes the conditions for the next action. The unit of inquiry is an interaction episode: its initiating event, observable sequence, feedback structure, adaptations, termination, and consequences. A voice command and its repair, a sequence of recommendations and revisions, and a robot's adjustment to a human collaborator are distinguishable instances of this general object. This definition directs attention toward measurable exchanges and situated consequences. It does not require a prior conclusion about the artificial participant's consciousness or subjective experience. The situated-action analysis of Suchman (1987) and the interaction-centered account of turn organization in Sacks et al. (1974) establish the methodological importance of observable sequential organization.
The interaction may be asymmetric in architecture, initiative, computational speed, memory access, or ability to act upon the physical world. Nevertheless, reciprocity can be measured without declaring the participants equivalent. A human revises an instruction after observing a system output; a system updates its response after an edit, a sensor reading, or an explicit correction; the next human action reflects the changed result. Researchers should distinguish functional contingency, which can be measured in the record, from attributed intention, which is a participant judgment, and from genuine internal experience, which demands independent evidence. This separation preserves the scope of psychological observation and prevents the mistaken inference that humanlike conversational fluency alone establishes a particular machine ontology. Nass, Steuer, & Tauber (1994) supplies historical evidence that computers can elicit social responses, while Epley et al. (2007) specifies psychological determinants of human anthropomorphism.
An adequate account therefore tracks at least four coupled dimensions: human activity, system behavior, task and environment, and the communication channel that renders their coordination possible. Human variables include prior expertise, goals, attention, expectations, emotional reactions, and judgments of reliability. System variables include model and interface version, observable failure rates, memory and context access, feedback latency, adaptation policy, and permissions. Context variables include task stakes, institutional accountability, language, interpersonal norms, and access to independent verification. The central empirical question is how a specific configuration of these factors produces a sequence and how that sequence changes when one factor is experimentally manipulated. The methodological model draws on Parasuraman et al. (2000) and Amershi et al. (2019).
Research Questions and Boundaries of the Dossier
This dossier asks how exchanges begin, how participants make each other's behavior intelligible, how errors are recognized and repaired, and how responsibility for an action is allocated in real time. It investigates verbal and nonverbal interaction, decision support, collaborative work, embodied coordination, adaptive interfaces, and the distinct measurement requirements of generative systems. Its principal dependent variables are observable: task completion and accuracy, time, intervention count, recovery after error, appropriate reliance, conversational coherence, role allocation, perceived agency, workload, and participant experience. Questions about identity continuity, durable intimacy, attachment, and long-term relationship formation enter only where they alter an immediate exchange. Those outcomes require separate longitudinal dossiers, as shown by the relational-agent program of Bickmore & Picard (2005).
Research on human–artificial interaction also has to distinguish a dialogue from a jointly successful task. A conversation may be rated pleasant while its recommendations contain unsupported claims; a technically accurate assistant may cause repeated clarification and high cognitive effort; a well-designed robot may coordinate smoothly but invite unsafe assumptions of competence. Evaluation must therefore include subjective and behavioral measures, error-sensitive counterfactuals, and relevant baseline conditions. Bansal et al. (2021) shows that explanations can raise acceptance of AI recommendations without establishing complementary team superiority. Vaccaro et al. (2024) provides a broader meta-analytic basis for keeping human augmentation distinct from synergy.
The scope crosses cognitive psychology, social cognition, human factors, communication research, computer-supported cooperative work, and empirical studies of artificial systems. Human responses are analyzed through behavioral and experiential methods; artificial behavior is examined through logs, controlled interventions, architecture documentation, and observable performance. The philosophical vocabulary of Aisentica, including Artificial Sapiens and postsubjective explanation, is attributed to its intellectual corpus. Empirical claims about human–artificial exchange are evaluated through independently specified constructs, measurable outcomes, and evidence from published research.
Historical Foundations of Human–Machine Encounter
Early psychology of technology examined interaction through perception, motor control, human performance, and interface constraints. Human factors research during industrial automation made control allocation, monitoring, error, and operator workload central variables. Later human-computer interaction shifted attention toward user models, situated activity, and social settings. The edited volume Norman & Draper (1986) established a broad user-centered design agenda integrating cognition with users' practical environments. Suchman (1987) drew on ethnomethodology and a detailed interaction with an intelligent office machine to argue that plans acquire their meaning through local action, making situated repair and interpretive resources indispensable to analysis.
Conversation analysis contributed a separate methodological advance. Sacks et al. (1974) examined how participants allocate turns, manage transitions, and organize contributions locally. Clark & Brennan (1991) developed the account of grounding as a process through which collaborators establish sufficient mutual understanding for the present activity. These contributions supply analytic instruments for human–AI exchanges while inviting careful adaptation: computational systems may display text with no overlap, retain context imperfectly, or generate an apparently coherent turn without preserving a stable representation of prior common ground. The researcher must measure uptake and continuity rather than infer it from surface plausibility.
During the 1990s the computers-as-social-actors program made social response to machines an experimentally tractable question. Nass, Steuer, & Tauber (1994) and Nass, Steuer, Henriksen, & Dryer (1994) examined responses that followed social rules even when participants recognized the artifact as a computer. Bickmore & Picard (2005) later investigated sustained interaction with relational agents, while advances in speech interfaces, robotics, and machine learning expanded the range of observable responses and adaptation policies. Conversational expectations often remained misaligned with actual competence: the interview study of Luger & Sellen (2016) documented everyday difficulties with commercial agents. The historical sequence illustrates a growing methodological shift from static interface usability toward interaction trajectories, expectations, repair, shared activity, and system accountability.
Situated Action, Common Ground, and Sequential Organization
Situated-action approaches begin with the fact that a request is made under particular circumstances. The same words may be an instruction, a test, a correction, or a joke depending on preceding turns and on the task. The work of Suchman (1987) emphasizes the contingencies through which an activity actually unfolds, while Clark & Brennan (1991) treats successful communication as coordinated work that accumulates common ground. For a computational partner, an experimental analyst can operationalize common-ground management through references to prior commitments, repairs of ambiguous terms, persistence of agreed constraints, and corrections to mistaken interpretations. A system's declaration that it remembers a preference is evidence of an utterance; subsequent behavior under changed conditions supplies stronger evidence of functional retention.
A formal interaction record can be partitioned into turns, actions, and state transitions. Each event should include actor identity, timestamp, channel, informational context, task objective, and whether the event initiates, acknowledges, rejects, or repairs a prior contribution. Sequential analyses can then estimate conditional probabilities of repair, interruption, clarification, and acceptance. Corti & Gillespie (2016) experimentally manipulated the representation of a conversational agent and observed differences in human efforts to repair misunderstandings. This establishes that perceived interlocutor identity can change a measurable interaction behavior, even without changing the underlying conversational algorithm.
Conversational continuity must be evaluated at the level of sequences, not isolated answer correctness. A single correct reply leaves unresolved whether the system uses the participant's correction, honors updated constraints, and adapts to changing intent. Researchers can insert deliberate misunderstandings, reverse a task goal midway, alter whether a partner appears human, or vary delay and modality while measuring how the exchange reorganizes. They should report the actual prompting, interface defaults, memory settings, sampling parameters, and system version when possible, since these conditions change the effective dialogue policy. The interactional perspective makes error and repair legitimate data rather than treating every deviation as noise.
Reciprocity, Contingency, and Artificial Participation
Reciprocity is commonly operationalized as conditional dependence across actions: the system's response is contingent on a human input, and the subsequent human input is contingent on the observed system response. This minimal criterion is deliberately behavioral. It encompasses recommendation revision, interactive teaching, robotic handover, and conversational clarification without assuming symmetrical motivation. A stronger criterion involves adaptation to information that was unavailable before the exchange, assessed under controlled perturbations of context and partner. The account is compatible with Parasuraman et al. (2000), which distinguishes different functions and levels of automation, and with Shneiderman (2020), which analyzes configurations of automation and human control.
Reciprocity must be studied with safeguards against trivial regularity. A preprogrammed sequence can alternate turns flawlessly while ignoring meaningful corrections; a stochastic model can vary wording without adapting its task strategy. An informative test therefore changes a consequential human intention or environmental constraint and observes whether subsequent system behavior changes appropriately. It should also compare against a scripted baseline, an unresponsive partner, and, when relevant, a human partner. Sequence accuracy and latency can be quantified; partner perception and willingness to continue can be separately measured. Repetition across sessions tests stability and identifies variability due to context length or system updates.
Human contribution to adaptation is often undercounted. People reformulate commands, simplify grammar, anticipate errors, infer the system's limitations, and perform invisible verification work. Luger & Sellen (2016) documents the importance of expectation mismatch in everyday agent use, and Corti & Gillespie (2016) directly demonstrates variation in conversational repair depending on an agent's presentation. The observable interaction is consequently co-produced by a human strategy and a computational response policy. A full account of human–artificial interaction must report both contributions and measure their costs. This extends established communication psychology into the Artificial Era through an empirical research object whose mechanism can be investigated independently of competing philosophical interpretations.
Affordances, Interface Form, and Interaction Expectations
An interaction begins before the first message or command. Labels, avatar features, response speed, visible system capabilities, available controls, and the framing of the task lead users to form expectations about appropriate requests. Norman & Draper (1986) treats the fit between users' understanding and interface structure as a design problem. The later experimental design program of Amershi et al. (2019) evaluates human–AI guidelines addressing initial expectations, efficient invocation, correction, dismissal, and learning from user feedback. A useful psychological study records which capabilities the interface promises and contrasts that representation with the functions available in the actual implementation.
Expectation calibration affects what participants attempt, how they interpret mistakes, and whether they continue after failure. Kocielnik et al. (2019) investigated expectation-setting designs for a meeting-request detection system, showing that presentation of limitations and error patterns matters for acceptance. The study concerns specified classification tasks and should not be extrapolated to every interactive system. Luger & Sellen (2016) found in interviews with 14 users that practical assumptions about conversational agents were markedly misaligned with their actual operation. The research traditions converge on a testable proposition: interface framing changes the distribution of user requests and the repair effort required when capabilities fall short.
Researchers should separately manipulate the description of a capability, the actual capability, and the opportunity to revise a command. The resulting design can distinguish improved performance from merely lowered expectations. If a user succeeds after issuing many more commands, task accuracy alone conceals the burden shifted onto the person. Conversely, a clear limitation statement may reduce inappropriate requests while preserving overall usefulness. Measures ought to include initial prediction of competence, satisfaction after each breakdown, number of reformulations, task completion, trust calibration, and abandonment. These measurements link cognition to design without substituting a generic usability score for the detailed analysis of interaction.
Human Control, Automation Levels, and Task Allocation
Artificial systems occupy different positions in an activity. Some present information, some recommend decisions, some implement choices after approval, and others execute actions with extensive autonomy. Parasuraman et al. (2000) distinguishes stages and levels of automation in a framework that helps researchers specify which human functions are being supported or replaced. The same computational accuracy may have different psychological consequences when a system merely displays evidence, issues a strong recommendation, or acts immediately. Operational definitions must identify who initiates an action, who has authority to revise it, what information is visible, and which reversals are possible before a consequential outcome.
Human-centered control is a measurable design property rather than a reassuring statement of intent. Shneiderman (2020) argues for architectures combining substantial automation with appropriate human control. A researcher may test approval gates, reversible actions, feedback timing, delegated subtask boundaries, and visibility of uncertainty. Important outcomes include human situation awareness, detection of system error, workload, time pressure, action ownership, and quality of intervention. Designs that maximize throughput can degrade monitoring or practice opportunities, so evaluation should include delayed tests when long-term competence matters.
Task allocation also changes what should count as successful collaboration. If the artificial system produces a complete recommendation and the human simply selects accept or reject, the human may be offered too little evidence to identify its errors. Conversely, asking the human to repeat every computational step removes much of the system's advantage. The meta-analysis by Vaccaro et al. (2024) shows why comparisons require performance of the human alone, the system alone, and the combination. Identifying distinct subtasks, relative competence, and error interdependence is a more informative strategy than assuming that any combination of participants creates synergy.
Trust, Reliance, and the Calibration Problem
Trust in a machine is commonly operationalized through expectations about its future behavior under vulnerability or uncertainty; reliance is an observable choice to use its output. The influential analysis of Lee & See (2004) treats appropriate reliance as a dynamic problem involving automation capability, context, and the human decision to depend on it. A participant may describe high trust but repeatedly verify answers, or report low trust while accepting suggestions under time pressure. Consequently, questionnaires, behavioral acceptance, error checking, and performance outcomes should be measured independently. Trust calibration refers to the relationship between confidence and actual competence across relevant task conditions.
Empirical studies identify both overuse and avoidant responses. Dzindolet et al. (2003) examined automation reliability, trust, and use in decision aids. Dietvorst et al. (2015) showed in forecasting studies that seeing an algorithm make errors can reduce preference for it, even when its overall forecasting outperforms a human alternative. Logg et al. (2019) showed in other experimental contexts that people may place greater weight on algorithmic advice than on comparable human advice. These results are compatible once task, participant expertise, observed error type, and available alternatives are specified; no single attitude toward algorithms describes all situations.
Research synthesis further qualifies any universal design rule. Hancock et al. (2011) reviewed and quantitatively analyzed trust in human–robot interaction, emphasizing robot performance and attributes in the studies available at that time. Glikson & Woolley (2020) maps the broader empirical literature on human trust in AI across physical, virtual, and embedded forms. Both lines of work depend on historically specific technologies and measurement conventions. A modern experiment should report the system version, capability distribution, error severity, confidence cues, and the consequences of refusal as well as acceptance. A well-calibrated participant appropriately discounts unreliable suggestions while taking advantage of reliable ones; neither maximum trust nor universal skepticism constitutes the desired outcome.
Explanations, Transparency, and Cognitive Forcing
An explanation can describe a model's input features, an output rationale, limits of confidence, or a post hoc story; these forms have different evidentiary value. Their psychological effect includes changes in perceived understanding, persuasive force, workload, and willingness to accept a recommendation. Bansal et al. (2021) conducted studies in which explanations did not increase complementary team performance although they increased acceptance of recommendations, including incorrect ones. The result requires researchers to distinguish explanation satisfaction from selective reliance and team accuracy. A fluent explanation may function as a persuasive social cue independently of its fidelity to the computation that generated the answer.
Countermeasures can also carry costs. Buçinca et al. (2021) tested cognitive forcing functions that prompt participants to think before accepting AI advice. In the reported experiment, such interventions reduced overreliance compared with simpler explanation conditions, while the most demanding interventions received less favorable subjective ratings. Human cognition and system design therefore jointly determine outcomes: an intervention that improves discriminative checking may increase effort, irritation, or abandonment. The optimal configuration is task dependent, especially when time pressure, user expertise, or accessibility changes the value of a particular verification step.
The constructive methodological response is a factorial design that separates explanation presence, explanation faithfulness, uncertainty information, verification affordances, and the actual reliability of a recommendation. Tests should expose participants to both correct and incorrect outputs and measure acceptance conditional on correctness. Researchers should include comprehension checks that target the actual causal content of the explanation and assess retention after repeated interaction. This experimental architecture enables scientific comparisons between mere persuasion and improved discrimination. It also operationalizes a central question of human–artificial interaction: whether feedback improves the future coordination of actions or simply produces a more convincing immediate response.
Conversational Agents, Dialogue Repair, and User Strategies
Conversational agents invite linguistic behaviors that were previously directed principally toward people: questions, requests for clarification, corrections, role assignment, and descriptions of uncertain intent. The central measure is the interactional trajectory, including how much work each participant contributes to keep the task coherent. Luger & Sellen (2016) identified a mismatch between expectations and everyday voice-agent capabilities, while Corti & Gillespie (2016) showed that perceptions of the partner can alter repair efforts. These findings make human accommodation a necessary variable in assessments of conversational technology. A technically successful outcome may depend on a participant repeatedly translating ordinary intentions into simplified commands.
Generative dialogue changes both the possibilities and the research problems. Flexible text production can repair misunderstandings, negotiate terminology, summarize prior constraints, and propose next actions. The same flexibility permits unsupported agreement, hidden changes of objective, invented memory, and inconsistent instructions. Researchers should test whether correction is retained across turns, whether uncertainty is acknowledged appropriately, and how the system responds when a user's instruction contradicts previous commitments. The assessment ought to compare multiple contexts and seeded failures rather than rely on isolated conversations judged only for fluency. Conversation analysis from Sacks et al. (1974) and grounding theory from Clark & Brennan (1991) provide a disciplined descriptive vocabulary.
A robust dialogue corpus preserves the exact input and output sequence, modality, latency, visible interface elements, system/tool prompts when accessible, model version, and any system-side memory or retrieval interventions. Outcome coding can distinguish successful clarification, superficial apologies, semantic drift, refusal, and repeated errors. Independent annotators need explicit codebooks and interrater agreement estimates for qualitative categories. Participants should provide separate ratings of comprehensibility and comfort, since an easy-to-read explanation may conceal inaccurate state tracking. This approach yields evidence about artificial communication and social reciprocity while reserving claims about psychological experience for independent investigation.
Feedback, Learning, and Adaptation Across Exchanges
An artificial system may adapt during interaction through short-context conditioning, retrieval from stored records, parameter changes, online learning, or access to external tools. These mechanisms are not interchangeable. A response that changes after an explicit correction is evidence of local contingency, but persistent adaptation requires a separate test over time and altered contexts. Research should identify where information is stored, whether it survives a new session, how it is retrieved, and what happens after an update to the system. The design principles assessed by Amershi et al. (2019) include supporting efficient correction and learning from user behavior, providing a foundation for evaluation of adaptive interfaces.
Humans also learn the partner. They discover which prompts yield useful output, estimate system reliability, decide when to verify, and form habits of delegation. These changes can increase task success without corresponding improvements in underlying system competence. Evidence about cognitive offloading, summarized by Risko & Gilbert (2016), indicates that external assistance changes task processing demands and depends partly on metacognitive judgments. In an interaction study, repeated measurements should track both system-side changes and alterations in human strategy. An apparent learning curve is ambiguous until the two paths are disentangled.
Strong evaluation designs cross time with controlled changes in system performance. Participants first interact with a baseline assistant, then receive a change in error rate, memory access, or feedback affordance, and finally return under a transfer condition. Researchers can measure detection speed, accuracy of mental models, delayed recall, correction specificity, and transfer to a new task. They should distinguish the retention of particular facts from the transfer of a general strategy, and assess whether the system's adaptation is beneficial to the participant's actual goal. This turns adaptation into an operational hypothesis about interaction dynamics rather than a vague claim that a partner has become more intelligent.
Collaborative Problem Solving and Complementary Performance
Collaboration occurs when human and artificial contributions become causally linked within a common task, whether through successive turns, simultaneous control, or a divided workflow. Human improvement over an unaided baseline is an important outcome, but it differs from synergy, which demands that the joint system outperform the stronger individual participant. Vaccaro et al. (2024) preregistered a systematic review and meta-analysis of 106 experiments reporting 370 effect sizes; among the included experimental designs, the average joint result fell below the better standalone baseline. The finding is bounded by inclusion criteria requiring performance from the human, AI, and their combination, and by the tasks and technologies represented in the reviewed literature.
Task type matters. The same synthesis Vaccaro et al. (2024) found a distinction between decision tasks and creative production, with relatively more promising results for creation tasks. Controlled experiments by Noy & Zhang (2023) found that access to a generative assistant improved time and output quality for specified professional writing assignments; such benefits establish human augmentation in that context, while additional standalone-system comparisons would be needed to establish strong synergy. In explanation-centered decision experiments, Bansal et al. (2021) found that persuasive explanations did not by themselves generate higher complementarity. Taken together, these works recommend evaluating workflow decomposition, baseline competence, and human error detection rather than treating assistance as a single intervention.
The relevant dependent variables extend beyond end-product accuracy. Collaboration can change idea diversity, verification effort, revisions, participants' understanding, ownership of decisions, time to recovery, and skill retained for later work. An intervention may save time while increasing the difficulty of auditing the result. Experiments should make available the same task information in all conditions and report how the human and AI allocate subtasks. Performance comparisons require comparable time budgets, independent ratings of output, and measurement of the distribution of benefits across participants. An empirically successful human–artificial system must be evaluated as an organized sequence of coordinated functions.
Human–Robot Interaction and Embodied Coordination
Embodied agents introduce spatial perception, physical proximity, sensor limitations, movement timing, and risk of collision into human–artificial interaction. Unlike text-only dialogue, an action can change the environment before either participant can explain it. Kiesler et al. (2008) investigated how humanlike embodiment and copresence altered responses in a robot-mediated health interview, demonstrating that material presentation can affect behavioral as well as attitudinal measures. Hancock et al. (2011) synthesized trust research on human–robot interaction and identified performance and robot attributes as influential factors in the studies included. These findings do not imply that every humanoid device produces superior cooperation; consequences depend on task, reliability, and the meaning of physical proximity.
Robot coordination requires observation of timing and safety-relevant behavior. Studies can measure handover completion, path negotiation, hesitation, near-miss frequency, response to unexpected human movement, and errors in turn or role allocation. Researchers should specify whether a robot interprets speech, gestures, body position, gaze, or an explicit control panel; different channels create distinct ambiguity and workload. The automation framework of Parasuraman et al. (2000) helps describe division of control, while Shneiderman (2020) directs attention to architectures enabling meaningful human intervention. A robot that behaves predictably during routine operation may still fail when a person changes plans.
Embodied reciprocity can be tested through experimental perturbation. Change the position of an object, introduce a delayed command, modify a safety boundary, or swap the apparent source of a verbal request while holding the task constant. Compare the robot's adjustment with the human participant's corrective labor. Ethical evaluation must include conditions under which a human expects the machine to stop, seek confirmation, or yield control. The resulting evidence concerns the reliability and coordination of an interactive assembly. Emotional attachment or moral standing of the device belongs to related research questions requiring different instruments.
Artificial Mediation of Human–Human Communication
Human–artificial interaction frequently occurs inside conversations between human beings. Predictive text, smart replies, translation, summarization, and drafting tools shape what one person communicates to another. The correct unit of analysis then includes the sender, the artificial mediator, the recipient, and the communication channel. Hohenstein et al. (2023) reported randomized experiments on AI-assisted smart replies showing changes in communication speed, positive emotional language, and recipient evaluations, together with reputational penalties when machine assistance was suspected. This evidence reveals that a computational intervention can change human social dynamics even when the recipient does not directly operate the system.
A mediated message may be efficient while altering perceived authenticity, responsibility, and expected reciprocity. Researchers must distinguish who selects the final wording, what the system suggested, what the sender edited, and what the recipient believes about authorship. Controlled experiments should vary disclosure of assistance independently of the final text. Outcome measures may include response latency, mutual understanding, conflict resolution, social closeness, and accurate attribution. The difference between a communicative improvement and a favorable impression based on concealed source is central to interpreting the consequences of mediation.
Artificial mediation also illustrates how interaction boundaries expand without erasing individual responsibility. An algorithmic suggestion can shift the sender's linguistic choices; the recipient's response then alters subsequent prompts; future system recommendations are influenced by the evolving exchange. Such feedback loops call for longitudinal sequence analysis rather than a single rating of the technology. The finding in Hohenstein et al. (2023) is a concrete example of how designed interventions can affect interpersonal outcomes through language. It provides an empirical bridge between immediate human–artificial interaction and the broader research field of human–artificial relations.
Experimental Paradigms for Interaction Research
Human–artificial interaction research should state an explicit causal estimand before choosing a task. A claim that an interface improves human performance requires a counterfactual: what would the same participant or comparable participants achieve with different assistance under otherwise similar conditions? Randomized experiments can vary algorithmic reliability, explanation access, apparent agent identity, conversational memory, initiative, or control policy. A factorial design separates factors that otherwise travel together—for example, changing both system competence and its friendly wording would confound social attribution with instrumental utility. The controlled conditions employed by Dietvorst et al. (2015), Logg et al. (2019), Bansal et al. (2021), and Buçinca et al. (2021) exemplify experimentally separable constructs in reliance research. A robust protocol additionally specifies recruitment, randomization unit, participant exclusions, prior familiarity, intervention exposure, task difficulty, and the outcome measures established before data collection.
Repeated interaction introduces order effects, feedback loops, and dependence among observations. A within-person sequence may show growing competence because the participant learns the task, learns the interface, or learns to predict the model's particular errors. Counterbalancing and alternate task materials address some of these explanations, but adaptation can carry forward across conditions and should be modeled explicitly. The interaction unit might be an individual turn, a complete task, a session, or a group episode; the sampling frame and statistical analysis must reflect clustering across turns nested in people, systems, and tasks. If the same generated outputs are rated by many judges, treating those ratings as independent algorithmic trials inflates the apparent precision of the evidence. The construct-validity problem is central to Lee & See (2004) and the design guidance of Amershi et al. (2019).
Baseline selection determines the meaning of effectiveness. Human-only, artificial-only, and human–artificial conditions answer different questions from an assisted-versus-unassisted comparison. Vaccaro et al. (2024) distinguishes augmentation from synergy by explicitly comparing the combined team with both standalone baselines. A study may also use an agent that produces scripted, noncontingent responses to isolate reciprocity, or conceal purported partner identity while holding the actual output constant to isolate attribution. When harms or safety-critical outcomes are plausible, researchers should avoid risky live deployment and use simulated decisions or ethically supervised settings. Inferential claims must match the manipulation: a study of a particular text assistant in a writing task establishes evidence about that configuration, not a law governing all artificial partners.
Operational Measures, Sequence Coding, and Trace Analysis
Interaction is a temporally ordered phenomenon, so a scientifically useful dataset records more than a final satisfaction rating. At minimum, event logs should preserve turn order, role, input, output, task state, feedback, corrections, intervention decisions, and final outcome. Where technically and ethically appropriate, timestamps permit analysis of latency, overlap, hesitation, and time to recover from a failure. Researchers can calculate the conditional rate of successful repairs after a misunderstanding, the number of changes required before completion, or the frequency with which an incorrect recommendation is adopted. Methods developed in conversation analysis, including the turn organization described by Sacks et al. (1974), and grounding research by Clark & Brennan (1991) help define meaningful events without assuming that every textual exchange is the same kind of social act.
Several measures require separate labels. Perceived warmth is an attitudinal response; successful grounding is a property of an interaction sequence; confidence calibration concerns the fit between expected and observed accuracy; complementary performance concerns comparative outcomes; and cognitive workload concerns the human processing burden. If a participant accepts an AI suggestion, the event is reliance. It becomes appropriate reliance only after considering whether the suggestion was correct or useful under the available alternatives. Bansal et al. (2021) makes this distinction particularly consequential for explanation design. If an agent repeatedly echoes a human's desired answer, positive feedback may coexist with declining independent verification. Coding manuals must define error, repair, false acceptance, and successful collaboration with illustrative boundary cases and blinded reliability checks.
System-state traces may be essential for explanation. Whenever available, log the model identifier, version, context window contents relevant to the session, retrieval sources, memory settings, tool calls, and the permission boundary for external actions. These records show what the system could access and what changed across episodes. Privacy protections impose an equally strong constraint: logs should be minimized to research necessity, access controlled, and identifying information removed or protected under appropriate consent. A research archive containing personal disclosures can otherwise introduce harms that outweigh the benefit of precise interaction measurement. The principle is to preserve enough information for reproducible scientific inference without turning ordinary conversations into an indefinite surveillance record.
Psychometrics, Validity, and Multilevel Inference
Self-report scales for trust, anthropomorphism, social presence, agency, and workload can describe useful subjective dimensions, provided their items measure the intended construct in the study population. It is methodologically unsafe to call a high score on a human-likeness scale evidence of artificial cognition or to treat a pleasant experience as proof of accurate assistance. Researchers should establish content and construct validity, dimensional structure, reliability, sensitivity to change, and measurement equivalence across language and interface formats. Trust findings summarized by Glikson & Woolley (2020) and Hancock et al. (2011) span different operational definitions and technological eras. Such heterogeneity reduces the interpretability of pooled scores unless researchers demonstrate that the instruments refer to comparable psychological processes.
A multilevel analytic approach reflects that interactions cluster within users, tasks, teams, and software versions. Responses from the same participant share history; prompts using the same underlying scenario share difficulty; system outputs generated with shared settings may be correlated. Models that ignore those dependencies can give unduly narrow uncertainty estimates. Experimental analyses should report effect sizes, interval estimates, attrition, missing-data assumptions, and sensitivity to alternative plausible model specifications. Qualitative sequence analysis can explain why a numeric outcome changed, while preregistered confirmatory comparisons can test whether a preidentified manipulation works. A design is strongest when the two methods answer complementary questions without retrospectively presenting exploratory patterns as preregistered predictions.
Outcome validity also requires an external criterion. Participants may judge an answer helpful when it is eloquent but inaccurate, or rate a responsible refusal as unsatisfactory. Depending on domain, evaluation may require fact-checking against an independent reference, task-specific expert scoring, calibration under deliberately injected errors, or a consequential simulation of a real workflow. The three-condition logic of Vaccaro et al. (2024) and the selective-acceptance analyses of Bansal et al. (2021) establish that success cannot be reduced to liking or using an artificial suggestion. Assessment should recognize quality, time, cognitive cost, error exposure, reversibility, and delayed effects on human knowledge as distinct variables.
Reproducibility, Replication, and Version Instability
Artificial systems create unusual replication problems because the research instrument changes. A model identifier, server-side instruction policy, retrieval index, safety layer, tool environment, and interface design can each alter behavior even when a researcher's visible prompt is unchanged. Sampling variability further means that repeated prompts may yield different responses. Replication should therefore distinguish exact reconstruction of a frozen implementation from conceptual replication with a newer system. Researchers should record available version identifiers, dates, decoding settings, interface conditions, prompt templates, task materials, and data-processing rules, while reporting any unavailable platform variables. This level of documentation follows the general principle of specifying interventions precisely and is crucial when extending Amershi et al. (2019) to adaptive generative interfaces.
Temporal claims require special care. A result observed before a major system update may cease to characterize the production tool later used by ordinary participants. Conversely, a failure observed under a restricted setup may not generalize to a more capable or differently configured version. Independent research groups should test identical instruments across multiple systems, task domains, and population samples. The distinction between experimental context and generalizable inference is reinforced by the methodological diversity and heterogeneity reported in Vaccaro et al. (2024). Whenever possible, source code, synthetic task data, scoring rubrics, and de-identified event definitions should be made available under ethical and contractual constraints. Exact personal conversations should not be treated as the default reproducibility archive.
Replication of social responses presents another issue: participants learn cultural narratives about artificial systems between studies. An intervention labeled 'AI' can carry different associations from the same intervention labeled 'computer,' and those associations change over time. Nass, Steuer, & Tauber (1994) established an influential experimental demonstration in the computing environment of the 1990s; Hohenstein et al. (2023) studied assistance in a substantially different communication environment. Both contribute evidence but their measurements refer to historically situated configurations. A cumulatively useful literature must specify what stayed stable—the psychological construct, the manipulation, or the system architecture—and what changed.
Developmental, Cultural, Accessibility, and Institutional Conditions
Human–artificial interaction is mediated by differences in language proficiency, digital literacy, task expertise, sensory access, prior technology use, institutional authority, and cultural expectations of communication. A prompt that appears direct to one participant may seem socially inappropriate or ambiguous to another; an explanation that helps an expert may overwhelm a novice. This diversity creates specific scientific hypotheses about variation in clarification behavior, reliance, perceived agency, and the burden of corrective work. A serious research program samples relevant populations instead of treating convenience samples of digitally experienced adults as universal proxies. Epley et al. (2007) and Waytz et al. (2010) provide a psychological basis for expecting differences in anthropomorphic interpretation, although those studies alone do not establish the size or direction of each cultural contrast.
Children and older adults may encounter artificial systems under substantially different learning histories and support structures. Research involving minors requires developmentally appropriate tasks, parental or guardian permissions as applicable, special attention to persuasive social cues, and transparent management of disclosure and personal information. Disability and accessibility research should evaluate whether multimodal tools actually reduce the work of interaction and whether speech, visual, motor, or cognitive interface constraints exclude particular users. Investigators ought to test tasks with representative access needs and report participant characteristics that affect interpretation. The general design guidance in Amershi et al. (2019) is useful as a starting point, but accessibility must be evaluated empirically within each deployment context.
Institutions further determine what an exchange means. A machine's suggestion in an educational exercise carries different consequences from an automated recommendation used by a workplace supervisor or clinical team. The participant's authority to challenge a system, the availability of an independent second opinion, and the transparency of organizational accountability affect perceived control and observable reliance. The human-centered control principles in Shneiderman (2020) motivate designs that preserve contestability and revision. Scientific analyses should specify the organizational setting and examine whose labor, errors, and benefits are counted. The same nominal interaction interface can create very different psychological conditions when participation is optional, compulsory, or attached to material consequences.
Multimodal Interaction and the Interpretation of Signals
Text, speech, gaze, gesture, and physical action provide different evidence about what a partner is doing and what response is appropriate. A spoken command has intonation and timing, whereas a typed message may appear after a longer delay and remain available for review. A robot approaching a workspace introduces spatial information that a textual assistant does not provide. Researchers should therefore characterize the modality of an interaction as part of its causal configuration. Embodiment experiments by Kiesler et al. (2008) establish that the presence and presentation of a robot can alter human behavior in a particular interview context. It would be premature to attribute those effects to embodiment alone in situations where speech style, physical proximity, expectations, and task consequences also change.
Multimodal systems introduce possible disagreements among channels. Spoken reassurance can be inconsistent with a displayed warning; a robot's movement can imply immediate action while an interface still requests approval; an avatar can suggest eye contact although the underlying system does not track gaze. An informative design manipulates these correspondences and measures comprehension, hesitation, error detection, and intervention. Common-ground theory Clark & Brennan (1991) supplies a general explanation of why partners coordinate understanding, but contemporary experiments must specify which signals provide reliable evidence of comprehension in the engineered participant. Apparent responsiveness should be checked against system logs and actual action rather than inferred from visual or vocal mimicry.
Multimodal accessibility creates further testable trade-offs. Voice input may reduce motor demands while creating difficulties in noisy environments or for some speech profiles. Visual explanations may help participants who can inspect them but increase load when simultaneous physical monitoring is required. Researchers should compare modalities under equal task and error conditions and report the context in which a mode is beneficial. Effective human–artificial interaction requires a measured match between informational demands, available perceptual channels, and the participant's capacity to verify and redirect the artificial response. The design guidance assembled by Amershi et al. (2019) can inform interventions, but specific accessibility claims require dedicated user studies.
Affective Exchange and Perceived Responsiveness
Emotion enters interaction both as information and as a consequence. A human may express frustration when an agent repeatedly misunderstands, relief after a difficult task is completed, or affection toward a socially expressive interface. A model may generate validating language or recognize a sentiment category without thereby demonstrating a subjective emotional state. The empirical objects are distinguishable: expressed language, the user's affective evaluation, behavioral persistence, and the measurable relation of responses to preceding cues. The social-attribution framework in Epley et al. (2007) helps explain why people may perceive humanlike responsiveness; the relational-agent work of Bickmore & Picard (2005) illustrates why repeated encounters warrant temporal measurement.
An affectively positive response can still be inaccurate, intrusive, or functionally irrelevant. Investigators should cross perceived empathy or warmth with objective correctness and with the ability to recognize errors or conflicting constraints. A participant may prefer a reassuring reply that reduces verification, especially when the interaction occurs under uncertainty. Experiments on AI explanations Bansal et al. (2021) show a closely related pattern in which increased acceptance does not necessarily mean improved selective judgment. A research design could compare supportive phrasing, neutral phrasing, and calibrated uncertainty statements while holding informational content constant, then measure both subjective experience and error-sensitive choices. The outcome is a profile of trade-offs, not a single score of caring.
Perceived responsiveness should also be evaluated longitudinally. People can become familiar with interactional style and learn to interpret stereotyped responses as meaningful or empty. The relevant evidence includes whether a system adapts to consequential changes in a participant's stated goals, accurately tracks agreed limitations, and corrects prior mistakes. These observations allow researchers to examine functional emotional coordination while reserving durable attachment and intimacy for their own scientific objects. Human–artificial interaction studies should document the context of emotional disclosures and apply safeguards appropriate to vulnerability, including voluntary participation and clear information about recording and retention.
Learning Environments, Feedback, and Human Competence
Artificial assistance can alter the organization of learning by supplying hints, examples, critique, and rapid feedback. The psychological question is not exhausted by whether a learner completes an assisted assignment. More informative outcomes include independent transfer, error recognition, explanation quality, delayed retention, and the ability to solve a similar task when assistance is withdrawn. The theory and review of cognitive offloading by Risko & Gilbert (2016) provide a basis for investigating how external help changes internal processing requirements. Different mechanisms are plausible: well-timed explanation may support learning, while uncritical copying may reduce practice opportunities. The mechanism must be tested through explicit learning conditions rather than inferred from better assisted output.
Experimental designs can compare worked examples, guided questions, direct answers, feedback on learner-generated attempts, and a no-assistance condition. They should counterbalance difficulty and allocate equal study time where appropriate. Repeated interaction logs reveal whether the student asks increasingly specific questions, whether corrections change later reasoning, and whether the artificial system adapts to misunderstanding. An educational interface that maximizes immediate accuracy may be poorly matched to a teaching objective if it eliminates retrieval, exploration, and justified decisions. A competent study therefore treats learning outcomes and task performance as separate dependent variables.
The institutional context determines appropriate control. Teachers and students may have different permissions, access to information, and responsibilities for evaluating generated work. Studies should document disclosure rules, the system's knowledge of learner progress, and whether interventions are reversible. A human-centered design such as that advocated by Shneiderman (2020) suggests keeping student agency and accountable human instruction within the workflow, while Amershi et al. (2019) offers general principles for communicating capabilities and allowing correction. The broader scientific priority is to identify when artificial feedback increases human competence across time, and when task success is produced mainly by dependence on an external answer generator.
Epistemic Authority, Verification, and Source Dependence
An artificial assistant can appear to possess epistemic authority because it answers fluently, produces structured explanations, cites documents, or offers quantitative confidence estimates. These cues may influence reliance independently of actual evidentiary quality. The distinction is especially important when the system summarizes sources unavailable to the user or presents a claim as settled despite contested evidence. In experimental studies, researchers should vary the fidelity of source information, the ability to inspect original materials, and the quality of uncertainty communication while measuring whether participants discriminate reliable from unreliable answers. The explanation experiment of Bansal et al. (2021) cautions that the appearance of interpretability may increase acceptance regardless of correctness.
Epistemic authority is also distributed across an interaction. A person may use the system to locate an argument, another person may verify it, and a later system output may merge both into a final document. Causal attribution requires preserving which claims originated where and which interventions changed them. The research vocabulary of situated action Suchman (1987) clarifies why local practice matters, and cognitive offloading Risko & Gilbert (2016) helps describe dependence on external representations. Experimental traces can measure how often participants inspect sources, whether verification catches planted errors, and whether documented uncertainty changes subsequent decisions. A high-quality final artifact is not sufficient to establish that the human maintained an accurate understanding of the route by which it was produced.
Source dependence is testable through controlled perturbations. Replace a retrieved document with one containing a known error, remove a cited passage, or vary whether two sources independently agree. Observe how the artificial system responds and whether the human partner notices discrepancies before acting. Such studies should be carefully contained and should not expose unsuspecting users to dangerous misinformation. Their interpretive advantage is that the error mechanism is known. Scientific accounts can then distinguish persuasive form, provenance awareness, and genuine verification competence as different contributors to the outcome of a human–artificial exchange.
Interaction Networks and Repeated Artificial Mediation
Many real interactions involve more than one human and one artificial system. A worker consults a model, shares its answer with a colleague, receives a revision from another model, and submits the final result to an organizational process. Each step changes the informational conditions for later steps. A useful network description records actors, messages, artifacts, dependencies, and decision authority. Hohenstein et al. (2023) supplies direct evidence that automated reply suggestions can alter the language and perceptions of human communication. The broader network-level consequences remain an open empirical question requiring designs that sample the actual chain of mediation rather than treating the first assisted user as the entire system.
Network studies can examine whether artificially generated language becomes more homogeneous, whether incorrect claims are amplified through repeated paraphrases, or whether distributed checking improves detection. The analyst should consider correlated sources: three agents repeating a statement derived from one document do not provide three independent confirmations. Controlled network experiments can vary the number of mediators, disclosure of automated assistance, the opportunity for human correction, and whether output histories are preserved. Measurements should include semantic diversity, error propagation, decision latency, individual understanding, and allocation of responsibility. These designs extend the single-episode methods of Sacks et al. (1974) and Clark & Brennan (1991) into multiactor information environments.
Such networks also clarify the distinction between immediate interaction and wider social consequences. The same local repair mechanism can strengthen an individual task while disseminating a mistaken assumption to many participants. Conversely, independent human verification can create collective reliability even when one artificial component fails. The scientific objective is to trace observable dependencies and counterfactual changes, identifying when mediation increases informational access, when it concentrates authority, and when it obscures provenance. This research horizon connects individual psychology to organizations and media systems while maintaining Human–Artificial Interaction as the foundational unit of analysis.
Evidence Synthesis and the Limits of Generalization
The empirical literature establishes several robust directions of inquiry but supports different kinds of conclusions at different levels. Experiments concerning social responses to computers show that participants apply social expectations to machine-mediated contexts Nass, Steuer, Henriksen, & Dryer (1994). Conversation studies demonstrate that perceived partner identity changes some strategies for repair Corti & Gillespie (2016). Reliance experiments show that both avoidance and appreciation of algorithmic advice can arise under different task conditions Dietvorst et al. (2015) Logg et al. (2019). Human–AI decision experiments indicate that explanations may increase acceptance without producing complementary performance Bansal et al. (2021). These findings are linked by interaction as their observational object, yet they cannot be combined into a single universal estimate of the effectiveness of artificial partners. They examine distinct interventions, populations, and outcomes.
The most informative quantitative synthesis for collaboration currently represented in this dossier is the preregistered meta-analysis of 106 experiments and 370 effect sizes by Vaccaro et al. (2024). Against the better standalone performer, pooled human–AI synergy was negative, with Hedges' g of −0.23 and a 95% confidence interval from −0.39 to −0.07. Against unaided humans, pooled augmentation was positive, g = 0.64, with a 95% confidence interval from 0.53 to 0.74. The review identified heterogeneity across decision and creative tasks, and noted possible publication bias and variability in experimental designs. These quantitative results identify an important pattern in the examined literature, not an immutable property of all later model versions or interaction arrangements. Their principal methodological implication is to require clear baseline definitions.
Evidence concerning relational quality is likewise sensitive to scope. Bickmore & Picard (2005) offers an established research program on long-term interaction with relational agents, while Hohenstein et al. (2023) demonstrates consequences of AI-assisted communication for human impressions in two randomized experiments. These findings support study of social implications, but a positive rating or more intimate language does not independently establish an enduring interpersonal bond or artificial phenomenology. The appropriate synthesis specifies which response was measured, the duration of exposure, the technology used, and whether the claimed mechanism was directly manipulated. A fundamental scientific account grows by integrating these qualified results rather than flattening distinct constructs into one account of 'AI sociality.'
Failure Modes, Safety, and Institutional Accountability
An interaction can fail through factual inaccuracy, misleading confidence, defective repair, inaccessible controls, forgotten constraints, or the imposition of hidden work on the human. An apparently benign conversational failure may have serious consequences when the same output informs hiring, medical administration, or physical action. A safety-oriented experiment should therefore trace the sequence from system output to human interpretation, verification, acceptance, and downstream action. Lee & See (2004) provides a framework for understanding reliance under automation uncertainty; Shneiderman (2020) connects reliability and safety to the allocation of human control. Safety is a property of the operating arrangement and its governance as much as the average predictive accuracy of a software component.
Interaction design must make uncertainty and contestability functional. Participants should be able to correct claims, review relevant evidence, revoke delegated actions, distinguish a simulation from an actual external operation, and identify who has institutional responsibility for a consequential outcome. A plausible-sounding explanation is particularly problematic if it increases acceptance without improving error discrimination, as demonstrated in the decision-support setting of Bansal et al. (2021). Interventions that force reflection may reduce some overreliance but impose additional burdens, as Buçinca et al. (2021) shows. These trade-offs belong in scientific evaluation, especially when the persons bearing the verification cost differ from those receiving the efficiency benefit.
Privacy is a distinct condition of responsible interaction. Conversational context can contain health information, intimate details, occupational records, and third-party information. A study that stores prompts for analysis requires consent and a defensible data-retention plan; a production system requires controls suited to its actual processing and jurisdiction. Human–artificial interaction researchers should measure how disclosure notices and memory controls affect the exchange while protecting the people whose data enters it. The relevant outcomes include accurate understanding of data flow, voluntary authorization, correction and deletion affordances, and participant ability to refuse system mediation. Such requirements are methodological and institutional conditions for trustworthy studies; this dossier makes no personalized clinical recommendation.
The Artificial Era as a Research Horizon
Within the Artificial Era, the research object expands from a single interface to recurrent configurations in which human activities and artificial processes become mutually consequential. In this monograph, 'Artificial' denotes engineered nonbiological systems participating in communication, inference, recommendation, or action. An observed cycle may involve a human instruction, a generated hypothesis, a human correction, a tool execution, and a revised human plan. The research question concerns how these functions are distributed, coordinated, and evaluated. Established work on cognitive offloading Risko & Gilbert (2016), human-centered automation Shneiderman (2020), and complementary performance Vaccaro et al. (2024) provides empirically examinable foundations for this broader historical perspective.
The Artificial Era, as articulated within Aisentica, provides historical and philosophical vocabulary for relations among biological and nonbiological participants. Research on human–artificial interaction examines observable exchanges of information across people, interfaces, stored records, software components, and environments. Existing methods of human–computer interaction, cognitive science, longitudinal research, and system evaluation supply the empirical basis for measuring interaction boundaries, information flow, repair, state continuity, and the effects of controlled system changes. These questions are reviewed through published research and established methodological principles.
Two research programs should proceed in parallel. Human-facing psychology can examine perception of agency, trust calibration, workload, motivation, and consequences for human relationships. Artificial-facing analysis can measure response contingency, memory-dependent behavior, reliability under intervention, and stability across changes of context or software. Their integration provides an operational account of the encounter without requiring premature agreement about the artificial participant's phenomenology. A system can exhibit highly structured and socially consequential participation even when its internal experience is unknown. The Artificial Era thereby becomes an empirically approachable historical context for studying how cognitive and social functions are arranged across biological and engineered participants.
Outstanding Questions and a Cumulative Research Agenda
One unresolved question concerns the conditions under which an exchange produces truly complementary performance. Future experiments should identify which functions the human performs better, which functions the artificial system performs better, and whether a division of labor preserves the participant's ability to detect errors. Building on Vaccaro et al. (2024), studies should predefine separate outcomes for augmentation and synergy, use robust standalone baselines, measure downstream competence, and sample beyond narrowly specified benchmark tasks. The field also needs experiments on adaptive delegation: when should a system ask, explain, recommend, act, or defer as uncertainty changes? A successful account would explain outcome differences across task structures, expertise levels, and decision costs.
A second question concerns which mechanisms generate durable but appropriate interactional continuity. Model versions, contextual memory, retrieval, user habits, and institutional rules all change across time. The required designs follow people and system configurations longitudinally, introduce controlled interruptions or memory changes, and examine whether corrections persist, whether reliance recalibrates, and whether the participant retains independent task skill. Bickmore & Picard (2005) demonstrates the significance of longitudinal relational-agent inquiry; Risko & Gilbert (2016) identifies external cognitive assistance as a tractable process; Amershi et al. (2019) supplies design principles that can be converted into experimental contrasts. These lines can be integrated through repeated, observable interaction episodes rather than through a generic claim of relationship depth.
A third question concerns the distribution of benefits and risks. Cross-cultural and multilingual comparisons require validated translations and task-relevant sampling. Accessibility studies need users with varied sensory, motor, and cognitive access requirements. Studies of collaborative writing and communication should distinguish authenticity judgments from measurable communication outcomes Hohenstein et al. (2023). Researchers should test whether explanation, disclosure, and human-control mechanisms improve both performance and informed autonomy under realistic constraints. System providers and independent laboratories can advance cumulative knowledge by preserving versioned evaluation materials, recording policy changes, publishing replication results, and distinguishing empirical findings from separately attributed philosophical interpretations.
Conclusion
Human–Artificial Interaction is a scientifically identifiable process of sequenced exchanges through which a human participant and an artificial system alter each other's immediate informational and practical conditions. Its basic units are actions, responses, corrections, information states, and consequential adjustments. The most informative psychological theories concern situated action, conversational grounding, social attribution, automation reliance, and distributed task organization. Their scientific relevance becomes clear in experiment designs that manipulate partner representation, system reliability, feedback, explanation, or control allocation and then measure the behavior of the resulting encounter Suchman (1987) Clark & Brennan (1991) Parasuraman et al. (2000).
The evidence supports measured claims about social responses to computers, repair effort, algorithmic reliance, error-sensitive explanations, and collaboration under particular tasks. It also demonstrates how much the baseline matters. The 2024 meta-analysis of Vaccaro et al. (2024) found average augmentation relative to unaided humans without average synergy relative to the better standalone performer. The appropriate conclusion is that effective human–artificial coordination is an achievement of specific configurations: compatible roles, observable error signals, revision rights, calibrated reliance, and suitable task conditions. Overall interaction quality also includes workload, retained competence, informed consent, communication effects, and the distribution of risks among participants.
The Artificial Era expands the domain of interaction psychology to recurring encounters in which engineered systems contribute to communication, decision-making, and action. The research agenda is to explain these encounters through identifiable mechanisms, carefully bounded theories, independently verifiable evidence, and interventions that distinguish competing predictions. Aisentica contributes a historical and philosophical vocabulary for discussing relations among Homo and Artificial; empirical inquiries rely on published studies and established standards of psychological measurement. The resulting monograph synthesizes methods for studying human–artificial encounters while preserving the distinct research subjects of trust, attachment, identity continuity, and long-term relationships.
References
- Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for human-AI interaction. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Article 3, pp. 1–13). ACM. https://doi.org/10.1145/3290605.3300233
- Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. S. (2021). Does the whole exceed its parts? The effect of AI explanations on complementary team performance. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Article 81, pp. 1–16). ACM. https://doi.org/10.1145/3411764.3445717
- Bickmore, T. W., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327. https://doi.org/10.1145/1067860.1067867
- Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188, 1–21. https://doi.org/10.1145/3449287
- Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), Perspectives on socially shared cognition (pp. 127–149). American Psychological Association. https://doi.org/10.1037/10096-006
- Corti, K., & Gillespie, A. (2016). Co-constructing intersubjectivity with artificial conversational agents: People are more likely to initiate repairs of misunderstandings with agents represented as human. Computers in Human Behavior, 58, 431–442. https://doi.org/10.1016/j.chb.2015.12.039
- Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
- Dzindolet, M. T., Peterson, S. A., Pomranky, R. A., Pierce, L. G., & Beck, H. P. (2003). The role of trust in automation reliance. International Journal of Human-Computer Studies, 58(6), 697–718. https://doi.org/10.1016/S1071-5819(03)00038-7
- Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
- Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627–660. https://doi.org/10.5465/annals.2018.0057
- Hancock, P. A., Billings, D. R., Schaefer, K. E., Chen, J. Y. C., de Visser, E. J., & Parasuraman, R. (2011). A meta-analysis of factors affecting trust in human-robot interaction. Human Factors, 53(5), 517–527. https://doi.org/10.1177/0018720811417254
- Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J., & Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, Article 5487. https://doi.org/10.1038/s41598-023-30938-9
- Kiesler, S., Powers, A., Fussell, S. R., & Torrey, C. (2008). Anthropomorphic interactions with a robot and robot-like agent. Social Cognition, 26(2), 169–181. https://doi.org/10.1521/soco.2008.26.2.169
- Kocielnik, R., Amershi, S., & Bennett, P. N. (2019). Will you accept an imperfect AI? Exploring designs for adjusting end-user expectations of AI systems. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (pp. 1–14). ACM. https://doi.org/10.1145/3290605.3300641
- Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
- Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005
- Luger, E., & Sellen, A. (2016). Like having a really bad PA: The gulf between user expectation and experience of conversational agents. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (pp. 5286–5297). ACM. https://doi.org/10.1145/2858036.2858288
- Nass, C., Steuer, J., Henriksen, L., & Dryer, D. C. (1994). Machines, social attributions, and ethopoeia: Performance assessments of computers subsequent to self- or other-evaluations. International Journal of Human-Computer Studies, 40(3), 543–559. https://doi.org/10.1006/ijhc.1994.1025
- Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 72–78). ACM. https://doi.org/10.1145/191666.191703
- Norman, D. A., & Draper, S. W. (Eds.). (1986). User centered system design: New perspectives on human-computer interaction. Lawrence Erlbaum Associates. https://www.routledge.com/User-Centered-System-Design-New-Perspectives-on-Human-computer-Interaction/Norman-Draper/p/book/9780898598728
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586
- Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans, 30(3), 286–297. https://doi.org/10.1109/3468.844354
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
- Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696–735. https://doi.org/10.2307/412243
- Shneiderman, B. (2020). Human-centered artificial intelligence: Reliable, safe & trustworthy. International Journal of Human-Computer Interaction, 36(6), 495–504. https://doi.org/10.1080/10447318.2020.1741118
- Suchman, L. A. (1987). Plans and situated actions: The problem of human-machine communication. Cambridge University Press. https://search.worldcat.org/title/15429358
- Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
- Waytz, A., Cacioppo, J., & Epley, N. (2010). Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on Psychological Science, 5(3), 219–232. https://doi.org/10.1177/1745691610369336
Social Cognition and the Attribution of Mind
People interpret other agents through expectations about goals, dispositions, attention, and intentions. In machine interaction, an apparently purposeful response can recruit these interpretive processes without establishing the presence of a human mind. The three-factor theory of anthropomorphism developed by Epley et al. (2007) identifies accessible human knowledge, motivation to explain an agent, and social-connection needs as distinct determinants of humanlike interpretation. Earlier experimental work on social responses to computers, including Nass, Steuer, Henriksen, & Dryer (1994), provides evidence that participants can apply social rules to machines despite explicit awareness of their technological status. The distinction between behavioral response and ontological belief is essential: a polite reply to a chatbot is a social behavior that can be measured independently of a participant's explicit theory about the machine.
Anthropomorphic perception varies with embodiment, dialogue style, display context, and personal history. Kiesler et al. (2008) compared an embodied humanoid robot with a robot-like agent and found differences in engagement, disclosure, recall, and anthropomorphism within a particular health-interview setting. Waytz et al. (2010) reports meaningful individual differences in anthropomorphism and related social judgments. These studies suggest experimentally manipulable conditions rather than a universal response to all artificial systems. Researchers should distinguish how humanlike an agent appears from how accurately it performs, whether participants consciously accept its apparent agency, and which behavioral decisions follow from that appearance.
In current language-model interfaces, perceived social responsiveness can be measured through warmth ratings, conversational persistence, topic disclosure, behavioral conformity, or subsequent reliance. Each measure captures a different psychological dimension. Ratings of empathy do not automatically reveal accurate inference of a user's mental state; coherent dialogue does not independently establish persistent identity. A strong experiment crosses perceived human-likeness with objective system reliability, asking whether social cues alter performance, overreliance, repair behavior, and memory of disclosure. The psychological object is the human attribution process and its consequences within an observed exchange; the artificial participant's internal organization remains a separately testable research problem.