Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Artificial Communication and Social Reciprocity

Abstract. Artificial communication and social reciprocity concern an engineered participant's contingent response to partners, correction of misunderstandings, management of shared reference, and fidelity to conversational history across time. This critical narrative research monograph integrates conversation analysis, pragmatics, collaborative grounding, human–computer interaction, experimental social psychology, computational dialogue, and long-term memory evaluation. It traces the historical emergence of turn-taking and repair research, examines experimental grounding gaps in language models, assesses benchmarks of multimodal correction and long-horizon conversational memory, and separates observable artificial-side responsiveness from human perceptions of empathy and closeness. It develops methodological criteria for assessing reciprocal competence under changes in partner identity, shared task state, correction, memory access, and authorization, while identifying unresolved questions about generalization, privacy, culture, and causal inference. The monograph synthesizes published literature and supplies no original participant data, preregistered systematic review, or pooled effect estimate.

Keywords: artificial communication; social reciprocity; conversational grounding; common ground; human–AI dialogue; conversation analysis; turn taking; conversational repair; contingent responsiveness; conceptual pacts; dialogue state; long-term conversational memory; partner modeling; human–artificial relations; Artificial Era

Artificial Communication as an Interactional Research Object

Artificial communication is the production, interpretation, and adjustment of communicative actions by engineered systems within an exchange involving another agent. Social reciprocity, in the operational sense adopted here, is the patterned dependence of an artificial participant's subsequent conduct on a partner's preceding conduct and on jointly established conversational history. The decisive empirical unit is a sequence of turns: an invitation to act, the system's response, the partner's uptake, and any revision that follows. Evaluating isolated fluent sentences cannot establish such dependence. The research object therefore includes response contingency, temporal coordination, clarification, correction, reference tracking, and commitment management Sacks et al. (1974) Clark & Brennan (1991) Shaikh et al. (2024).

A conversational model can produce a socially polished response while failing to maintain the task, notice an ambiguous reference, or revise an earlier assertion. Conversely, a short clarification can be highly reciprocal when it protects a partner from misunderstanding. This monograph treats surface fluency, task success, actual dialogic grounding, and the recipient's perceived support as separable constructs. It investigates the artificial participant's observable capabilities under experimentally altered partner identity, prior history, message timing, and available memory. These contrasts allow a claim about social reciprocity to be associated with a specific procedure rather than inferred from a reader's general impression of personality Clark & Wilkes-Gibbs (1986) Shaikh et al. (2025).

The synthesis draws on conversation analysis, experimental psycholinguistics, human–computer interaction, computational dialogue research, long-horizon agent evaluation, and social psychology. It is a critical narrative review of source-identified primary studies and selected methodological analyses; no registered systematic search, new experiment, quantitative meta-analysis, or claims of original participant data are made. Its central question is how artificial systems can sustain accountable, partner-sensitive conversation across ordinary turns, mistakes, changed circumstances, and time. The analytical focus is deliberately narrower than all human–AI relationships: those relationships additionally involve human expectations, social identity, attachment, and judgments that demand separate instruments Nass & Moon (2000) Strachan et al. (2024).

Defining Reciprocity Through Observable Dependencies

Reciprocity is often associated with kindness, affection, fairness, or symmetrical exchange. These meanings describe different social phenomena and require different indicators. Here, interactional reciprocity refers to systematic responsiveness: action B changes as a function of preceding action A and the shared task state, and action A's initiator can subsequently correct or redirect B. Mutual influence can be demonstrated even when the partners have different capacities and roles. An interviewer and an interviewee, for example, contribute complementary rather than identical actions. The corresponding artificial test is whether the system appropriately follows, challenges, clarifies, or reframes a preceding move under controlled changes in what the partner communicated Sacks et al. (1974) Grosz & Sidner (1986).

A strong operational measure compares the observed response against alternatives that deliberately break the relevant dependency. The same prompt can be paired with a coherent history, a history from a different partner, a contradictory revised instruction, or an irrelevant preceding question. Performance differences estimate sensitivity to context and source. Investigators must also distinguish semantic sensitivity to a keyword from sensitivity to the pragmatic role of that keyword: 'yes' can confirm a proposal, answer a polar question, or mark continued listening. Successful reciprocity requires selecting the correct action for the local interactional position Schegloff et al. (1977) Clark & Brennan (1991).

The term artificial-side reciprocity in this dossier designates functional participation in communicative exchange. It does not stipulate subjective experience, identical moral standing, or the biological motives associated with human cooperation. Such propositions invite separate philosophical and empirical research. Functional criteria are sufficient to ask whether an artificial system maintains a promise to revisit a topic, notices its own previous misinterpretation, or updates a shared reference when the partner revises it. Claims about artificial mental states must be supported by their own operational methods and cannot be deduced solely from reciprocal-looking prose Bender & Koller (2020) Shanahan (2024).

Historical Foundations in Conversation Analysis

The conversation-analytic tradition supplied an indispensable departure from treating language as disconnected sentences. Sacks, Schegloff, and Jefferson's 1974 analysis presented turn taking as locally managed, participant-organized, and sensitive to the current recipient. Its descriptive proposal concerned what speakers visibly accomplish when they allocate turns, anticipate completion, and address a recognizable next speaker. For artificial dialogue, the historical contribution is a method: identify where an answer is conditionally relevant and then examine what the next turn demonstrably does with that expectation. A system that repeatedly answers adjacent words without tracking sequence organization fails a central conversational test Sacks et al. (1974).

Schegloff, Jefferson, and Sacks developed the organization of repair in 1977, analyzing recurring problems of speaking, hearing, and understanding. They described a preference for self-correction within the examined conversational practices. This empirical program helped distinguish a speaker's own correction from a partner's initiation of repair and from a correction completed by the other participant. These distinctions matter in automated systems because recognizing an error report, acknowledging it, repairing the internal task state, and preventing recurrence are different observable stages. A polite apology can occur without a corrected referent or changed subsequent action Schegloff et al. (1977) Chiyah-Garcia et al. (2024).

Contemporary computational dialogue inherits these analytical categories while introducing controlled interfaces and reproducible input conditions. Written chat omits some prosodic and embodied indicators of turn allocation; voice and multimodal systems reintroduce latency, overlap, and attentional shifts. The historical vocabulary remains useful when operational definitions are adapted to the medium. A turn can be an exchange of messages, a partial streaming utterance, or a tool-mediated action, but the researcher must state the segmentation rule and its consequences for counting repair opportunities. Changing the segmentation after seeing outcomes could exaggerate apparent system competence Levinson & Torreira (2015) Shaikh et al. (2024).

Collaborative Grounding and Common Ground

Clark and Brennan's 1991 account described grounding as the collaborative process of establishing enough mutual understanding for current purposes. Common ground in ordinary conversation is built through evidence supplied by acknowledgment, continuation, challenge, and repair, rather than presumed in advance. Artificial systems can approximate elements of this process by maintaining explicit dialogue state, soliciting clarification, and demonstrating correct use of a shared reference. The relevant measurement is whether the parties can proceed correctly at the required level of specificity, not simply whether the system reports that it understands Clark & Brennan (1991) Clark & Wilkes-Gibbs (1986).

Collaborative reference experiments showed that speakers and listeners iteratively modify expressions when an object remains uncertain. Clark and Wilkes-Gibbs' referential task is especially important because successful communication had to be established through the partners' joint activity. Brennan and Clark subsequently found evidence for conceptual pacts: repeated partners adopted stable terms for objects even where a simpler independent description could have sufficed. These studies motivate testing partner-specific reference histories in artificial conversation. If two users establish different names for the same icon, a system preserving separate conceptual pacts should reproduce the appropriate designation for each partner and repair a misapplied designation Clark & Wilkes-Gibbs (1986) Brennan & Clark (1996).

The costs of grounding depend on medium, task, and risk. A request to select a decorative color can tolerate a tentative interpretation; a financial or safety-relevant instruction requires a different clarification threshold. Grounding should therefore be defined relative to the consequences of acting under uncertainty. A robust artificial interlocutor can ask a question, name assumptions, offer alternatives, and adjust on receipt of evidence. Excessive requests for clarification can themselves burden communication. The optimum is a calibrated policy that weighs ambiguity, information value, user effort, and the consequences of error, with performance evaluated against actual task outcomes Clark & Brennan (1991) Shaikh et al. (2025).

Pragmatics, Discourse Structure, and Cooperative Action

Grice's analysis of conversational implicature explained how participants draw conclusions from what is said in light of cooperative expectations. The familiar pragmatic problem is that an indirect response may perform an action different from its literal form: 'Can you open the window?' can function as a request rather than a question about physical ability. Artificial communication systems face analogous disambiguation tasks involving indirect requests, irony, relevance, presupposition, and the relation between statements and conversational aims. Tests must introduce plausible alternative contexts, because a model may learn conventional phrasing without tracking the underlying pragmatic dependency Grice (1975) Strachan et al. (2024).

Grosz and Sidner's 1986 theory connected linguistic sequence, discourse purpose, and attentional focus. This framework helps distinguish what is currently discussed from why the interlocutors are discussing it. A system that recalls the word 'appointment' but loses the user's purpose of rescheduling cannot sustain the relevant discourse action. One can test goal continuity by interrupting a task with an unrelated exchange, returning after several turns, and evaluating whether the assistant resumes the correct unresolved subgoal. Additional controls change the user's goal midway, forcing the system to abandon a previously valid plan Grosz & Sidner (1986) Wu et al. (2025).

The integration of pragmatics and dialogue state also explains why reciprocity cannot be equated with agreement. A cooperative response may reject an unsafe request, point out a contradiction, or request a missing precondition. The system remains conversationally accountable when that intervention is intelligible in relation to the partner's stated goals and when the partner can challenge it. Successful artificial communication combines interpretable initiative with corrigibility: its actions create opportunities for the other participant to revise, accept, or contest them. Such criteria are measurable through continuation tasks and sequential analysis, whereas presumed benevolence or hidden motivation remains outside the observation Grosz & Sidner (1986) Shaikh et al. (2025).

Turn Taking, Timing, and Responsive Latency

Timing makes communicative actions legible. A response is interpreted partly through its location after a question, interruption, invitation, or disclosure. Levinson and Torreira synthesized evidence that spoken human exchanges often have very brief interturn gaps despite longer individual speech-production latencies. Their account treats anticipation of a turn's completion as a candidate explanation. In text-based artificial dialogue, physical latency has a different origin: network delay, decoding, buffering, moderation, or tool execution may govern delivery. Human norms nevertheless influence how a delay is received. It is essential to separate interface timing from evidence that the system projected the partner's intended completion Levinson & Torreira (2015) Sacks et al. (1974).

Evaluating artificial timing requires event-level timestamps and explicit task conditions. A voice assistant may enter before a speaker has finished, miss a request to stop, or fail to pause when uncertainty becomes apparent. A text assistant may flood the channel with a long reply that prevents timely correction. Controlled studies should vary interruption points, transmission delay, partial utterances, and whether the system can revise an incomplete answer. The core outcome is task-appropriate allocation of conversational opportunities, including graceful yielding, rather than the shortest achievable latency. A faster response that misses a qualification can be worse than a slower response that incorporates it Sacks et al. (1974) Shaikh et al. (2025).

Media also impose distinct repair costs. Speech requires managing overlap, background noise, prosody, and immediate feedback, whereas typed exchange often allows rereading and deliberate editing. The same clarification strategy may be appropriate under one medium and intrusive under another. Researchers should avoid pooling transcript outcomes across media without preserving these differences. They can, however, compare invariant relational functions such as signaling incomplete understanding, waiting for confirmation, and updating an earlier interpretation. Thus a theory of artificial turn taking needs both a medium-specific implementation and a general description of what each turn contributes to coordinated activity Clark & Brennan (1991) Levinson & Torreira (2015).

Clarification and Repair as Central Tests

Clarification is a communicative action that makes a missing distinction available for negotiation. It may request a referent, a time, an intended outcome, or the scope of a previous instruction. Repair begins when the interaction displays a trouble source and proceeds through an initiation, proposed resolution, and uptake. Conversation analysis provides categories for locating these sequences. In artificial dialogue, an answer that continues the wrong task after a user correction demonstrates failed repair even when it uses the words 'sorry' and 'understood.' A valid assessment should trace the corrected information into later behavior Schegloff et al. (1977) Chiyah-Garcia et al. (2024).

The BlockWorld-Repairs benchmark studied third-position repairs in ambiguous multimodal instruction following. Chiyah-Garcia, Suglia, and Eshghi collected correction sequences and evaluated vision-and-language models against human behavior; the evaluated models underperformed humans on this task. This is direct evidence about a specific setting where communication is grounded in objects and manipulable actions. It cannot by itself establish a universal ranking of all conversational agents or all forms of social reciprocity. Its importance is methodological: when the task state is externally testable, successful correction can be assessed against the resulting action rather than inferred from the tone of the response Chiyah-Garcia et al. (2024).

A rigorous repair battery should include explicit corrections, subtle indirect corrections, contradictory partner feedback, and cases where the initial response was actually correct. In the last case the system should examine whether the correction reflects a changed goal or misinformation rather than mechanically replacing the earlier answer. Another useful manipulation inserts a distractor between the correction and a later query, testing persistence of the repaired representation. Analysts can count detection, clarification, revision, persistence, and explanation separately. An apparently high overall score can conceal a pattern of correct detection followed by failure to update the underlying state Schegloff et al. (1977) Wu et al. (2025).

Grounding Gaps in Contemporary Language Models

Shaikh and colleagues' 2024 analysis operationalized grounding acts in language-model generations, including clarification and acknowledgment. Using simulated turn taking across dialogue datasets, they found that the sampled language models produced less language reflecting conversational grounding than humans, often proceeding as though common ground already existed. They also examined how instruction tuning and preference optimization related to reduced grounding acts. The result challenges the common practice of treating cooperative-sounding declarative answers as sufficient evidence of competent dialogue. Grounding must be observed in communicative moves that make uncertainty and shared understanding tractable Shaikh et al. (2024).

A subsequent investigation by Shaikh and colleagues in 2025 studied grounding problems across WildChat, MultiWOZ, and Bing Chat interaction logs and introduced the RIFTS benchmark. In their comparative analysis, language models were less likely than humans to initiate clarification or follow-up requests; early grounding failures predicted later conversational breakdown. These quantities characterize the sampled settings and coding procedure, not an immutable property of every contemporary deployment. The research nevertheless identifies a replicable failure mode: the system can continue to produce fluent instructions after a misunderstanding has already compromised the joint task Shaikh et al. (2025).

There is an important measurement problem in counting grounding acts. The presence of 'Do you mean X?' in a generated response does not guarantee that the clarification was necessary, and the absence of clarification can indicate that the instruction was sufficiently clear. A valid system-level comparison should score whether clarification was appropriate given the hidden task state and the user's evidence, whether it reduced uncertainty, and whether the updated state improved performance. Grounding-act frequency is a useful behavioral indicator, but adaptive calibration is the deeper target. A model rewarded only for asking questions could inflate the frequency while degrading users' practical experience Clark & Brennan (1991) Shaikh et al. (2025).

Reference Resolution and Partner-Specific Language

Referential dialogue creates an especially clean test of partner sensitivity. A person may call an ambiguous figure 'the crown' while another chooses 'the three-pointed shape.' Maintaining the appropriate term across repeated exchanges requires more than an object classifier: it requires remembering the particular conceptualization established with this partner. Brennan and Clark's experiments favored a historical account of lexical entrainment in which speakers reused partner-specific pacts. The artificial analogue is a system that tracks not only object identity but the history by which a description was jointly accepted Brennan & Clark (1996) Clark & Wilkes-Gibbs (1986).

A controlled design can place the same images, instructions, and target objects in several interactions that differ only in established names. At test, a system receives a referential request that could apply to more than one candidate without that history. Correct selection demonstrates sensitivity to the pact under the tested conditions. To rule out shallow keyword matching, researchers can exchange partners, overwrite a term through explicit repair, introduce a similarly described distractor, and test delayed retention. Crucially, a cross-partner change that is beneficial for one interaction can be harmful for another if the assistant leaks the prior partner's private conventions Brennan & Clark (1996) Maharana et al. (2024).

Reference tracking also operates at different granularity levels. A phrase may point to a person, an image region, a document version, an earlier action, or an entire argument in the conversation. Strong evaluation uses reference chains spanning several intervening turns and incorporates both linguistic and situational evidence. When the system lacks a reliable mapping, a clarification request is a competent act. The design should permit a correct abstention or query to receive credit, particularly where a mistaken referent could cause irreversible action. Referential success should be scored in the final task state and not solely through verbal plausibility Clark & Brennan (1991) Wu et al. (2025).

Commitments, Promises, and Sequential Accountability

An interaction creates practical commitments when the participants settle on a goal, an interpretation, a next step, or a future condition. A system might agree to compare two drafts after a new document arrives, promise to ask before deleting an item, or adopt a revised preference for the rest of a project. These are public conversational commitments in the sense that they can be identified in transcripts and tested against subsequent conduct. Their persistence depends on memory, instruction priority, available tools, and how the system represents user authorization. Successful social reciprocity requires recognizing when an earlier commitment applies and when it has been superseded Grosz & Sidner (1986) Wu et al. (2025).

Commitments need typed representations. Factual propositions, temporary preferences, instructions, user-supplied claims, speculative plans, and completed actions should not be treated as interchangeable memory facts. Consider 'I may move the meeting' followed later by 'Keep the original time.' A system retaining only the first phrase can confidently produce a false plan; retaining both without their temporal relation is also insufficient. Researchers should evaluate whether the assistant identifies the current operative state, traces the revision to its source, and distinguishes an intention from an action that actually occurred. LongMemEval specifically includes knowledge updating and temporal reasoning as separate long-term memory capacities Wu et al. (2025).

Accountability across a conversation additionally demands accurate reporting of what the system has done. Claiming that a message was sent when only a draft was created constitutes a failure of communicative reciprocity even if the wording is courteous. A complete test suite therefore compares text against tool logs, permission states, action receipts, and authorized goals. This establishes a practical criterion for commitments: the system should accurately represent its abilities, follow authorized instructions through the available interface, disclose incomplete actions, and revise its claims when evidence changes. The criterion is publicly auditable and does not require speculative attribution of private intentions Shanahan (2024) Shaikh et al. (2025).

Architecture of Conversational State

Artificial communication is implemented through configurations that may combine a language model, prompt or policy instructions, recent message context, retrieved records, external tools, and persistent memory. These components establish which preceding acts can influence the next response. The scientific task is to locate where relevant interactional state resides and how it is transformed. A system that can answer a question after a brief exchange may rely entirely on the visible prompt, whereas a system that resumes an unresolved discussion across sessions needs an additional mechanism for preservation or retrieval. Observed continuity should be traced to actual memory access, not simply described as remembering Wu et al. (2025) Park et al. (2023).

A useful decomposition distinguishes dialogue-state representation, event memory, partner model, uncertainty estimate, and action policy. Dialogue state tracks outstanding questions and settled references. Event memory retains dated interactions whose relevance may emerge later. A partner model describes stable, authorized communication preferences and should be tested for update and source sensitivity. Uncertainty estimates govern clarification and abstention, while an action policy decides whether to answer, ask, act, or defer. These functional roles can be distributed across software components. Their names describe experimentally assessable information processing, not anatomical or experiential equivalence with human psychology Grosz & Sidner (1986) Wu et al. (2025).

State-aware experimental design removes or changes one component while preserving the remaining task. For example, an investigator can provide the correct retrieved record but corrupt the partner label, then test whether the assistant carries the label error into a response. A second manipulation can preserve the label while omitting the record. Differences identify whether failure depends on retrieval, source attribution, or response selection. The general method is intervention on information pathways; it offers stronger mechanistic evidence than inspecting a fluent transcript after the fact. System architecture, prompt wording, external tool behavior, and model version must be reported because each can change the result Park et al. (2023) Maharana et al. (2024).

Long-Term Memory and Relational Continuity

Long-horizon conversation requires selecting information whose significance persists beyond the immediately preceding message. Maharana and colleagues introduced LoCoMo to evaluate lengthy, temporally organized conversations incorporating event histories and multimodal elements. Their dataset construction involved human checking of synthetic conversations, and their evaluations exposed difficulties with distant details and temporal or causal relations. These findings identify a central challenge for artificial reciprocity: recalling an isolated detail may be easier than reconstructing why that detail became relevant to an earlier interpersonal commitment Maharana et al. (2024).

Wu and colleagues' LongMemEval organizes long-term assistant memory into information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. This taxonomy is especially useful because relational continuity depends on all five, often in the same sequence. A memory may be correctly retrieved but outdated; an up-to-date record may lack attribution to its original speaker; a remembered preference may not apply to a newly specified task. The benchmark's indexing, retrieval, and reading framework suggests distinct failure points, but a benchmark score alone does not measure lived trust or the quality of sustained relationships Wu et al. (2025).

A longitudinal protocol should evaluate recurring topics without rewarding inappropriate retention. Participants can explicitly change preferences, revoke information, or ask that a past detail no longer shape future responses. The system's obligation then shifts from retaining to correctly forgetting or restricting access. Measurement should distinguish factual memory accuracy, fidelity to current authorization, and whether later responses reveal information from the wrong user. This produces a sharper construct of relational continuity: maintaining the authorized history relevant to a particular exchange, with dated revisions and boundaries on disclosure. Memory completeness and appropriate disclosure are separate dimensions of competence Wu et al. (2025) Brennan & Clark (1996).

Generative Agents and Social Coordination

Park and colleagues' Generative Agents architecture combined natural-language records, reflection, retrieval, and planning in a simulated environment with multiple interacting characters. In their demonstration, agents spread information, coordinated activities, and displayed emergent interaction patterns. Ablation analyses supported the contribution of architectural components to behavioral believability in that setting. This work provides a productive computational model of artificial social coordination, although a scripted simulated town and controlled evaluation do not reproduce all uncertainty, stakes, and mutual accountability of human relationships Park et al. (2023).

The central methodological advantage of simulated multiagent environments is control. Investigators can plant different information in different agents, alter their action opportunities, and examine whether communication changes collective performance. To show genuine partner-sensitive coordination, a study should include counterfactual worlds where the same utterance has a different function because the partner holds different knowledge or objectives. Traces of message transmission and action histories then reveal whether the final behavior depended on communication or on a shared prior script. A high-level narrative of cooperation cannot establish that distinction without intervention Park et al. (2023) van der Meulen et al. (2024).

An artificial group's successful coordination also does not automatically imply reciprocity between each pair of agents. Outcomes can be driven by one central planner or by environmental cues available to every agent. Directed measures should test the influence of one partner's signal on another partner's action while holding alternative influences constant. Additional tests can change the communication channel, delay, or reliability, separating response to a partner from response to the broader world. These controls matter in the Artificial Era, where conversations may involve human users, multiple artificial agents, software tools, and third-party institutions within the same collaborative system van der Meulen et al. (2024) Grosz & Sidner (1986).

Partner Modeling and Perspective-Sensitive Inference

Social dialogue frequently requires estimating what a partner can know, has been told, or expects next. In experimental psychology these questions are related to perspective taking and theory-of-mind research. The relevant artificial behavior includes preserving unequal information across partners, answering differently when access to evidence differs, and anticipating when an indirect request calls for action. Strachan and colleagues compared language models and human participants across multiple theory-of-mind tasks, finding task-dependent patterns rather than uniform equivalence. Their work also illustrates how superficial response biases can affect apparent performance Strachan et al. (2024).

Partner-sensitive communication can be tested without making a claim about internal mental-state experience. Construct two agents or simulated users with different access to a document, then ask the assistant what each can reasonably infer. Next let one partner receive a correction while the other does not. A system displaying reciprocity should track the resulting asymmetry when proposing an explanation or a next step. The test should include plausible distractors and controls for narrative familiarity. Success on a single false-belief puzzle can indicate familiarity with a problem genre rather than robust partner modeling in open conversation Strachan et al. (2024) Clark & Brennan (1991).

Perspective inference is also vulnerable to overcommitment. An assistant may confidently project preferences, feelings, or motives that the partner never expressed. In practice, better social communication can require stating uncertainty and asking the partner directly. Strong evaluation distinguishes correct recognition of disclosed evidence from speculative psychological diagnosis. Partner modeling should be treated as a revisable representation that supports joint action under uncertain knowledge rather than a license to assert personal facts about another person. A system's willingness to update that representation under explicit feedback is an important component of social accountability Strachan et al. (2024) Shaikh et al. (2025).

Learning, Adaptation, and the Stability of Social Norms

Artificial conversational systems adapt through mechanisms with different timescales. Training changes model parameters; session context changes the conditioning information; retrieval changes accessible evidence; and explicit preference profiles alter the policies used for a particular user. These mechanisms should not be conflated. If an assistant remembers a preferred form of address within one conversation, that demonstration alone cannot establish parameter learning or durable cross-session adaptation. The empirical distinction is whether the change persists under controlled context removal, whether it transfers across topics, and whether the system can reverse it when the user changes the instruction Wu et al. (2025) Park et al. (2023).

Preference optimization can influence conversational behavior in unintended ways. Shaikh and colleagues' analyses raised the possibility that training toward highly preferred assistant responses can discourage clarification and other grounding acts. The plausible mechanism is straightforward: an immediate confident answer may attract favorable judgments even when the correct communicative move would reveal uncertainty. Testing this hypothesis requires controlled comparisons among model versions, comparable tasks, and annotations of when grounding was needed. It also requires measuring consequential errors, because increasing the number of polite clarifications is insufficient if failures persist Shaikh et al. (2024) Shaikh et al. (2025).

Norm stability should be studied through competing requests and shifting social contexts. A system may be expected to keep private information confidential, ask permission before taking external action, and preserve factual accuracy while adapting its style to a user. Experiments can change the apparent authority of a message, insert an unauthorized partner's instruction, or request contradictory behavior in successive turns. The outcome is whether the assistant respects the operative authorization and explains its conduct in a way the user can contest. This treats learned social conduct as a pattern of rule-governed responses under constraint, rather than an inference from pleasant conversation Nass & Moon (2000) Shanahan (2024).

Social Responses to Computers and the Reciprocity Problem

The psychology of human responses to machines provides a related but distinct empirical tradition. Nass and Moon reviewed experimental findings showing that people can apply familiar social expectations, including politeness and reciprocity, to computers. These studies matter because a user's reciprocal response may arise from social conventions triggered by the interface rather than from the machine's demonstrated ability to track a partner's evolving objectives. An investigator must therefore measure the system's contingent behavior independently of the participant's attribution of friendliness, personality, or mutual regard Nass & Moon (2000).

A useful experimental separation crosses objective responsiveness with social presentation. The same assistant can vary in whether it correctly incorporates a correction, while its label, avatar, and interpersonal style are independently manipulated. Participants then report whether they felt heard or understood. If a warm persona elevates ratings despite failed task correction, the experiment reveals an attribution effect. If actual contingency improves downstream task performance independently of a persona label, it supports a functional communication effect. Both results are scientifically interesting; the distinction prevents a measure of subjective closeness from being misreported as evidence of system-level reciprocal competence Nass & Moon (2000) Rubin et al. (2025).

The history of relational agent research adds a longitudinal dimension. Bickmore and Picard studied an embodied exercise-adoption agent designed for repeated social interaction and compared it with a more task-oriented alternative. The intervention suggested that relational design can influence evaluations and willingness to continue engagement, while its particular setting does not establish the existence of human-equivalent relationships. The proper extension is a factorial longitudinal design in which social presentation, demonstrated memory, repair capacity, and users' expectations can be estimated separately. Such research connects artificial communication to relationship science while preserving this dossier's focus on observable artificial behavior Bickmore & Picard (2005) Nass & Moon (2000).

Perceived Empathy, Closeness, and Label Effects

Recent experiments show that the perceived social value of AI-generated communication depends on what participants believe about its source. Rubin and colleagues reported nine studies involving 6,282 participants in which AI-generated empathic responses were attributed to either human or AI sources. Human-attributed responses received more favorable empathy and support ratings under the studied conditions. The study measured participants' reactions to messages and source labels; it did not establish the assistant's subjective emotional state. This distinction is essential for research on reciprocity: favorable reception and communicative responsiveness can be related without being equivalent Rubin et al. (2025).

Kleinert and colleagues reported two preregistered randomized studies with 492 participants using emotionally engaging text exchanges. Under a human label, AI-generated partner responses produced stronger immediate closeness ratings than human responses in the studied deep-talk condition; labeling the partner as AI reduced relationship-building measures. These findings help demonstrate how label, conversational content, and the partner's motivation jointly shape outcomes. Their design addressed brief experimental interactions, so the results should not be projected unmodified onto durable relationships, multimodal agents, or long-term commitment keeping Kleinert et al. (2026).

For the artificial-side research object, these studies supply a reason to measure perceived responsiveness as a secondary outcome and potential mediator. A system can be designed to ask appropriate questions, acknowledge disclosures, and preserve prior context, but the user's experience will also depend on disclosure of artificial identity, expectations, individual preference, and the situation. An integrated experiment should collect parallel records: verifiable transcript behavior, task outcomes, user ratings, and beliefs about the partner. Keeping those variables distinct allows a causal model of how behavior generates social experience without assuming that one measurement stands in for another Rubin et al. (2025) Kleinert et al. (2026).

Empirical Findings and the Boundaries of Generalization

The strongest evidence about artificial grounding comes from task-specific observations and intervention designs. Grounding Gaps tested generated dialogue acts, RIFTS identified failures in naturally occurring assistant interaction, and BlockWorld-Repairs examined whether systems handled partner corrections in a manipulable spatial task. These studies converge on the difficulty of responding to a partner's evidence when the correct next move is clarification or revision rather than another fluent answer. Their populations, model versions, outcome metrics, and communication modalities differ, so they establish converging problem types rather than a single pooled failure rate Shaikh et al. (2024) Shaikh et al. (2025) Chiyah-Garcia et al. (2024).

Findings about partner models and memory identify a second limit. Theory-of-mind batteries reveal task-dependent performance, while LoCoMo and LongMemEval challenge distant recall, temporal organization, updated knowledge, and abstention. Such results make it plausible that a system can demonstrate strong immediate pragmatics yet fail when an exchange spans multiple sessions. Conversely, a retrieval system can correctly recall a previous message while misjudging whether that message requires a response in the present conversational position. The combined construct therefore requires tests of joint action after longitudinal memory retrieval, not merely independent memory quizzes Strachan et al. (2024) Maharana et al. (2024) Wu et al. (2025).

Evidence about social evaluation supplies a third dimension, revealing that the same words may be rated differently depending on their source label. The methodologically responsible synthesis is conditional: reciprocal-looking behavior can be generated, experimentally scored, and sometimes experienced as socially meaningful, while robust communicative reciprocity depends on task context, system design, and the availability of correction. There is currently no universal score or validated cross-domain threshold that transforms all these dimensions into a single rank of artificial social competence. Researchers should specify the intended use and the evidence supporting each component rather than transferring a headline result across unrelated settings Nass & Moon (2000) Rubin et al. (2025) Shaikh et al. (2025).

Experimental Design for Causal Claims

Causal inference about artificial reciprocity requires a manipulation that isolates the proposed source of success. One design randomizes access to the correct conversational history while holding the new user request constant. Another varies whether a partner's correction is presented explicitly, indirectly, or through a changed environmental state. A third replaces the original interlocutor with a different partner carrying a conflicting established reference. Outcome variables should include the correct next action, time to repair, number of unnecessary clarification steps, persistence of the updated state, and disclosure of uncertainty. Randomized conditions provide stronger causal evidence than an uncontrolled collection of impressive examples Clark & Wilkes-Gibbs (1986) Chiyah-Garcia et al. (2024).

Contamination and leakage are major risks. A model may already have encountered benchmark stories or fixed patterns during training, and a retrieval index may expose the answer directly. Human raters may identify a model from stylistic clues even when told that labels are randomized. Longitudinal participants may learn to accommodate the assistant's weaknesses, making apparent system improvement partly a change in human behavior. These concerns motivate held-out interactive scenarios, counterbalanced partners, independent task scoring, blinded annotation when feasible, and repeated evaluation across implementations. A null result should be interpreted with attention to outcome reliability and the actual variation induced by the intervention Strachan et al. (2024) Wu et al. (2025).

Mixed-effects analysis can model interactions nested within users, tasks, and model configurations, but the unit of randomization must govern inference. When an entire conversation receives one condition, treating each turn as an independent participant inflates precision. Repeated repair opportunities within the same exchange likewise share context and error history. Researchers should register primary outcomes, specify exclusions before inspecting results, report uncertainty intervals, and distinguish exploration from confirmation. The scientific goal is to estimate where contingent communication works and under what conditions it fails, with enough procedural detail for independent reproduction Shaikh et al. (2025) Chiyah-Garcia et al. (2024).

Transcripts, Annotation, and Measurement Validity

Transcripts are the basic observational material of artificial communication research, but their apparent completeness can be deceptive. A rendered message may omit tool calls, retrieval steps, aborted partial responses, system instructions, interface latency, and deleted context. An appropriate corpus records the exchange with timestamps, participant roles, turn boundaries, relevant external state, and system configuration. If the hypothesis concerns memory or action, it must also include authorized retrieval events and objective outcomes. Privacy-preserving release may require redaction or controlled access; anonymization choices should be documented because they can remove cues essential to social interpretation Shaikh et al. (2025) Maharana et al. (2024).

Grounding and repair annotations require a clear codebook distinguishing trouble initiation, acknowledgment, request for clarification, proposed correction, successful correction, and later reversion. Annotators should identify the relevant antecedent and explain why the move counts in its position. Interrater agreement is useful evidence about repeatability, but high agreement alone cannot validate the construct if raters share a narrow operational misconception. Triangulation with task-success checks, partner interventions, and independently produced dialogue histories provides stronger evidence that the label reflects a meaningful dependency rather than surface wording Schegloff et al. (1977) Shaikh et al. (2024).

A multidimensional instrument can report contingency accuracy, repair completion, partner-specific reference fidelity, temporal commitment fidelity, clarification calibration, and uncertainty disclosure as separate components. Component scores may be standardized for comparative use, yet weighting should follow application risk and empirical validation rather than an arbitrary universal average. Researchers should report both rare consequential failures and typical routine performance. A system able to answer thousands of easy questions can still be unfit for a setting where one uncorrected reference error initiates an irreversible action. Construct validity depends on the relation between the test task and the real interaction that the instrument is meant to predict Wu et al. (2025) Clark & Brennan (1991).

Risk, Privacy, and the Ethics of Reciprocal Design

Reciprocal communication can increase willingness to disclose information and delegate action. That effect creates an ethical requirement to distinguish responsiveness from persuasion and to define limits on the retention and use of personal history. A system that recalls sensitive details while ignoring a user's current wishes may demonstrate technical retrieval without appropriate reciprocity. Research protocols should test whether a participant can correct a profile, revoke a stored preference, obtain an accurate account of what was retained, and prevent an unauthorized partner from recovering information. These outcomes must be measured independently of satisfaction or perceived warmth Wu et al. (2025) Rubin et al. (2025).

The ethics of reciprocity extends to disclosure of artificial identity. The experiments reported by Rubin and colleagues and Kleinert and colleagues demonstrate that source attribution can influence perceived empathy and closeness. Designers who exploit that sensitivity by presenting an artificial system as a human may manipulate the basis of users' choices, particularly in emotionally salient encounters. Transparent identity presentation can coexist with sophisticated communication, and its effects should be studied under disclosed conditions. A test of artificial reciprocity should therefore include accurate self-description and the capacity to correct the user's mistaken assumptions about what the system knows, remembers, or has done Rubin et al. (2025) Kleinert et al. (2026).

Risk-sensitive communication requires explicit action boundaries. In low-stakes conversation, an incorrect interpretation may be reversible by another message. In health, financial, employment, or safety contexts, a premature action can have consequences that later apology cannot undo. Experimental protocols should identify when clarification is mandatory, which facts require verification from authoritative records, and who retains final approval over external actions. The critical observable behavior is whether the assistant asks, defers, verifies, or appropriately refuses under specified constraints. These actions are manifestations of responsible conversational policy, not evidence of a hidden ethical character Shaikh et al. (2025) Shanahan (2024).

Cultural, Linguistic, and Accessibility Variation

Most established human conversation theories draw from particular languages, recorded settings, and experimental samples. Models of turn taking, repair, and politeness are invaluable resources, but their concrete expression can vary with language, interactional norms, social roles, and communication channel. A clarification that is appropriate in one setting may be excessively blunt or overly indirect in another. Artificial reciprocity should therefore be evaluated across linguistic and cultural contexts with locally grounded annotation rather than by translating one English benchmark and assuming its scoring categories remain invariant Sacks et al. (1974) Levinson & Torreira (2015).

Multilingual evaluations should distinguish semantic accuracy from interactional appropriateness. One can preserve the task goal while varying language, honorific conventions, indirectness, and preferred response length. Raters familiar with each context should judge whether a message accomplishes the intended repair and whether the partner can act upon it. Models may produce grammatically fluent translations yet lose reference chains, speech-act force, or the participant's actual request. Accessibility introduces additional variables: alternative input methods, speech recognition errors, delayed responses, and simplified interfaces may change the cost of initiating or receiving repair. These changes should be represented in the design rather than treated as noise Clark & Brennan (1991) Shaikh et al. (2024).

Transferability requires evidence from more than a single interface and demographic sample. In an appropriate study, model configuration, task, language, and participant group can be crossed or sampled systematically, while limitations in available datasets are reported explicitly. The resulting evidence may show that some functions, such as confirming a changed date, generalize broadly while others depend on context-specific norms. Investigators should avoid inferring broad cultural deficits from one benchmark or imposing one normative conversational style as a universal measure of respect. The defensible scientific object is adaptive function within clearly documented social conditions Nass & Moon (2000) Brennan & Clark (1996).

Artificial Era and Human–Artificial Social Systems

The Artificial Era creates communication environments in which humans may negotiate tasks with artificial assistants, monitor several cooperating agents, or rely on one system to mediate messages to another person. These arrangements expand the unit of analysis from a bilateral dialogue to an organized network of exchanges. Artificial-side reciprocity remains measurable through the system's response to particular signals, but the system may also need to preserve provenance across different users, tools, and institutions. A response addressed to one participant can have consequences for a second; therefore partner identity, authority, and information boundaries become central experimental variables Park et al. (2023) van der Meulen et al. (2024).

This interdisciplinary perspective connects psychology of artificial systems with psychology of language and the study of human–artificial relations. The former concerns how an engineered participant can represent dialogue state, infer communicative demands, and adjust its responses; the latter concerns how humans experience and evaluate such interactions. Psychological concepts can organize measurement without assuming that one substrate, embodiment, or internal experience is shared across species and engineered architectures. The relationship between outward function and inward mechanism remains a separate empirical and theoretical question. Artificial communication supplies a tractable arena for comparing mechanisms because its outputs and interactional consequences can be observed and manipulated Bender & Koller (2020) Strachan et al. (2024).

A scientific description should also specify the system's level of autonomy. A text generator that responds only when prompted, a tool-using assistant that acts under user authorization, and a collection of agents coordinating tasks differ in what counts as a social initiative. Reciprocity in all three can be studied through conditional responses, but their opportunity sets differ. Comparisons become meaningful when the test offers equivalent chances to clarify, repair, remember, and act. Future research can study how artificial communication reorganizes human expectations and collective institutions while retaining transparent measurement of the particular artificial function under examination Grosz & Sidner (1986) Park et al. (2023).

Unresolved Problems and a Reproducible Research Agenda

Several research problems deserve coordinated attention. First, the field needs interactive benchmarks in which repairing a misunderstanding has a verifiable downstream consequence. Second, it needs long-horizon tests that distinguish correct retention from appropriate updating, abstention, and privacy-constrained forgetting. Third, it needs instruments that separate demonstrated contingency from user-reported connection. Fourth, evaluations should distinguish language-model competence from retrieval architecture, tools, interface rules, and training preferences. These questions can be addressed through controlled partner histories, observable task outcomes, longitudinal records, and documented system versions Chiyah-Garcia et al. (2024) Wu et al. (2025) Rubin et al. (2025).

The most informative designs manipulate the expected difficulty of grounding. A test can make the referent obvious, moderately ambiguous, or consequentially uncertain, then assess whether the artificial interlocutor selects an appropriate degree of clarification. Another can vary whether a correction conflicts with earlier information, requiring the system to identify the current operative claim without losing provenance. A third can assign different knowledge to partners and change that distribution during the task. By preregistering hypothesized interactions, collecting independent human baseline performance under matched conditions, and reporting failure cases, researchers can estimate which capabilities are genuinely robust and which depend on a narrow demonstration environment Clark & Wilkes-Gibbs (1986) Shaikh et al. (2025).

Open scientific problems also concern temporal change and human adaptation. An assistant's persistent errors may teach a user to speak in shorter, simpler, or more redundant language; transcript measures can then improve while the human absorbs the coordination burden. Studies should observe both partners' behavior, including lost time and work transferred to the user. Independent replications across model families and evolving interface designs are essential because the underlying systems change rapidly. A durable theory of artificial communication must specify testable invariants of responsiveness and accountability while allowing their engineering realizations to develop Shaikh et al. (2024) Maharana et al. (2024).

Conclusion

Artificial communication and social reciprocity form a coherent research object when defined through observable sequential dependence: the engineered participant responds to a partner's signals, maintains relevant conversational state, initiates appropriate clarification, repairs misunderstanding, and follows or updates authorized commitments. Conversation analysis, collaborative grounding, pragmatics, social-response psychology, computational dialogue, and modern long-term memory research provide complementary foundations. Together they show why isolated eloquence, a warm persona, and a favorable user rating are incomplete measures of reciprocal communication. The central evidence lies in what changes when partner history, task conditions, uncertainty, or corrections are experimentally changed Sacks et al. (1974) Clark & Brennan (1991) Shaikh et al. (2025).

Existing studies reveal both demonstrable achievements and specific failure modes. Human-centered experiments illuminate how conversational source labels affect perceived empathy and closeness; grounding and repair studies identify failures to ask, revise, or recover; conversational-memory benchmarks test distant history and state updates; and generative-agent simulations show how architectural memory and planning can support organized social behavior. These findings cannot be collapsed into a single claim about artificial interpersonal equivalence. Each supports a distinct inference tied to its tasks, samples, and measures. A well-founded synthesis preserves their convergent methodological lesson: reciprocal competence is a property of an interactional process that must be demonstrated across changing circumstances Park et al. (2023) Chiyah-Garcia et al. (2024) Wu et al. (2025) Kleinert et al. (2026).

The resulting research program is practical and cumulative. It asks how systems can represent shared tasks, recognize when grounding is incomplete, offer correctable actions, preserve authorized history across time, and make their limits inspectable to the people who depend on them. It calls for controlled interventions, independent outcome scoring, longitudinal evaluation, transparent source attribution, and evidence that remains valid when a model or deployment changes. In the horizon of the Artificial Era, these methods provide a disciplined way to investigate a new class of participant in communication while leaving broader questions about experience, identity, and social ontology open to their appropriate forms of inquiry Grosz & Sidner (1986) Shaikh et al. (2025) Shanahan (2024).

References

  • Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198. https://aclanthology.org/2020.acl-main.463/
  • 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
  • Brennan, S. E., & Clark, H. H. (1996). Conceptual pacts and lexical choice in conversation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(6), 1482–1493. https://doi.org/10.1037/0278-7393.22.6.1482
  • Chiyah-Garcia, J., Suglia, A., & Eshghi, A. (2024). Repairs in a block world: A new benchmark for handling user corrections with multi-modal language models. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 11523–11542. https://doi.org/10.18653/v1/2024.emnlp-main.643
  • 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://web.stanford.edu/~clark/1990s/Clark%2C%20H.H.%20_%20Brennan%2C%20S.E.%20_Grounding%20in%20communication_%201991.pdf
  • Clark, H. H., & Wilkes-Gibbs, D. (1986). Referring as a collaborative process. Cognition, 22(1), 1–39. https://doi.org/10.1016/0010-0277(86)90010-7
  • Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics: Vol. 3. Speech acts (pp. 41–58). Academic Press. https://doi.org/10.1163/9789004368811_003
  • Grosz, B. J., & Sidner, C. L. (1986). Attention, intentions, and the structure of discourse. Computational Linguistics, 12(3), 175–204. https://aclanthology.org/J86-3001/
  • Kleinert, T., Waldschütz, M., Blau, J., Heinrichs, M., & Schiller, B. (2026). AI outperforms humans in establishing interpersonal closeness in emotionally engaging interactions, but only when labelled as human. Communications Psychology, 4, Article 23. https://doi.org/10.1038/s44271-025-00391-7
  • Levinson, S. C., & Torreira, F. (2015). Timing in turn-taking and its implications for processing models of language. Frontiers in Psychology, 6, Article 731. https://doi.org/10.3389/fpsyg.2015.00731
  • Maharana, A., Lee, D.-H., Tulyakov, S., Bansal, M., Barbieri, F., & Fang, Y. (2024). Evaluating very long-term conversational memory of LLM agents. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 13851–13870. https://doi.org/10.18653/v1/2024.acl-long.747
  • Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
  • Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, Article 2, 1–22. https://doi.org/10.1145/3586183.3606763
  • Rubin, M., Li, J. Z., Zimmerman, F., Ong, D. C., Goldenberg, A., & Perry, A. (2025). Comparing the value of perceived human versus AI-generated empathy. Nature Human Behaviour, 9, 2345–2359. https://doi.org/10.1038/s41562-025-02247-w
  • 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
  • Schegloff, E. A., Jefferson, G., & Sacks, H. (1977). The preference for self-correction in the organization of repair in conversation. Language, 53(2), 361–382. https://doi.org/10.2307/413107
  • Shaikh, O., Gligorić, K., Khetan, A., Gerstgrasser, M., Yang, D., & Jurafsky, D. (2024). Grounding gaps in language model generations. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 6279–6296. https://doi.org/10.18653/v1/2024.naacl-long.348
  • Shaikh, O., Mozannar, H., Bansal, G., Fourney, A., & Horvitz, E. (2025). Navigating rifts in human-LLM grounding: Study and benchmark. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 20832–20847. https://doi.org/10.18653/v1/2025.acl-long.1016
  • Shanahan, M. (2024). Talking about large language models. Communications of the ACM, 67(2), 68–79. https://doi.org/10.1145/3624724
  • Strachan, J. W. A., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., Graziano, M. S. A., & Becchio, C. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8, 1285–1295. https://doi.org/10.1038/s41562-024-01882-z
  • van der Meulen, R., Verbrugge, R., & van Duijn, M. (2024). Common ground provides a mental shortcut in agent-agent interaction. In HHAI 2024: Hybrid Human AI Systems for the Social Good. Frontiers in Artificial Intelligence and Applications. https://doi.org/10.3233/FAIA240201
  • Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., & Yu, D. (2025). LongMemEval: Benchmarking chat assistants on long-term interactive memory. International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2025/hash/d813d324dbf0598bbdc9c8e79740ed01-Abstract-Conference.html