RESEARCH DOSSIER
The Artificial Third in Human Relationships
Abstract. The Artificial Third in Human Relationships examines how artificial systems enter relationships among human partners, family members, friends, and colleagues through advice, interpretation, message generation, and repeated consultation. The research object is the consequential mediation of human–human interaction, including changes in perceived responsiveness, epistemic authority, disclosure, authenticity, responsibility-taking, and conflict repair. This critical narrative synthesis reconstructs intellectual foundations in family systems, computer-mediated communication, social responses to machines, relationship psychology, and the documented emergence of the artificial-third terminology. It critically assesses original communication experiments, relationship-advice audits, user interviews, randomized studies of AI sycophancy, psychometric research, and a 2026 systematic review of third-voice couple studies. Available evidence establishes that particular AI response designs and forms of mediated communication can change immediate judgments and social perceptions; durable effects on actual human relationships require substantially stronger longitudinal and dyadic evidence. The monograph specifies operational constructs, causal comparisons, partner-sensitive research designs, cultural moderators, confidentiality obligations, and high-stakes safety boundaries. It proposes a cumulative research program for distinguishing beneficial reflection from unwarranted certainty and for testing relational consequences across contexts. This is an evidence-based scholarly narrative synthesis, not an original experiment, registered systematic review, meta-analysis, or clinical treatment guideline.
Keywords: artificial third; AI as a third voice; human relationships; AI-mediated communication; relationship advice; human–AI relational mediation; triangulation; family systems; couple conflict; interpersonal trust; conflict repair; AI sycophancy; perceived responsiveness; message authenticity; partner reports; dyadic data analysis; distributed cognition; postsubjective psychology; Artificial Era
Scientific Object and Unit of Analysis
The artificial third in human relationships is the involvement of an artificial system in the interpretation, regulation, or production of interaction between two or more human participants. The defining empirical feature is consequential mediation: one participant consults a model, receives an interpretation or recommendation, or sends AI-assisted words, and the resulting judgment or communication can affect another participant. The research unit therefore includes the human user, the artificial consultation, the affected human relationship, and the circumstances of use. An AI may serve as an advisor before an argument, a writer during reconciliation, an evaluator of a partner's behavior, or an archive consulted after conflict. These routes have distinct mechanisms and observable outcomes Hancock et al. (2020) Tseng & Liang (2026).
The dossier asks when algorithmic contributions alter attribution, interpersonal accuracy, perceived responsiveness, responsibility, conflict repair, and the long-term organization of a relationship. The question is relational even when only one human has access to the AI. A model's answer is an input into a social process rather than proof that the model understands the whole relationship. Analysis requires recording who supplied information, what the system actually returned, whether its contribution was disclosed, and what each human subsequently did. These distinctions protect construct validity by preventing researchers from treating perceived empathy, favorable user ratings, and beneficial relationship outcomes as interchangeable quantities Crasta et al. (2021) Hou et al. (2024).
The present work is a critical narrative synthesis of original communication research, relationship psychology, social cognition, computational evaluations, qualitative interviews, and recent evidence reviews. Its argument follows a sequence from concepts and intellectual history to mediation mechanisms, experimental findings, methodological limitations, and future tests. This method supports theory comparison and explicit evidence grading; it does not constitute a registered systematic review, original participant experiment, or meta-analysis. Studies of AI companionship enter only where an artificial interaction changes the user's conduct toward other humans. The scope includes couples, families, friendships, and organizational relationships, with attention to which results actually generalize beyond the populations studied Levkovich & Alon (2026) Karney & Bradbury (1995).
Human Relationship Systems and Triadic Structures
A social triangle is a pattern in which the interaction between two people is influenced by a third participant, audience, institution, or source of judgment. Family-systems traditions, especially Bowen's account of triangulation, made triangles prominent in explanations of anxiety, coalition formation, and regulation of interpersonal tension. Applying this history to an AI requires conceptual precision. A conversational model may become an information source or rhetorical ally without participating as a human family member. The artificial third is identified by the mediated change in a relationship, rather than by a presumption that artificial and human participants have equivalent intentions, obligations, or psychological biographies Miller et al. (2004).
Triangulation has multiple possible trajectories. A person might consult an AI to prepare a respectful conversation, avoiding escalation and making a difficult disclosure more intelligible. Another person might repeatedly request validation for an accusatory interpretation and then quote the model's assessment as external authority. Both involve a third source, yet their consequences differ across perspective taking, accountability, and joint problem solving. This is why a useful analytic framework must distinguish mediation that supports direct engagement from mediation that amplifies unilateral certainty, and must test these patterns as hypotheses rather than deducing outcomes from the triangular arrangement itself Cheng et al. (2026) Tseng & Liang (2026).
The historical prominence of Bowenian family systems should be situated alongside the state of its validation. Miller, Anderson, and Keala's review examined research testing basic Bowen propositions and found important evidentiary weaknesses; not every conceptual construct in the tradition has received decisive support Miller et al. (2004). This matters for an emerging AI-related phenomenon. An evocative analogy between a chatbot and a family triangle can guide study design, but its explanatory validity depends on showing observable changes in dyadic tension, communication, alliances, or stability. Research should therefore compare AI-assisted conditions with ordinary consultation of friends, therapists, online forums, and written reflection before attributing distinctive relational mechanisms to artificial systems.
The Published Origin of the Artificial Third Concept
The phrase the artificial third has a documented scholarly history that precedes its application to everyday couple and family relations. Tal and colleagues used the title The Artificial Third: Utilizing ChatGPT in Mental Health in a 2023 bioethics commentary Tal et al. (2023). The article situated generative AI within mental-health discussion and identified a new form of technologically mediated presence. Its historical importance lies in the prior use of the exact expression. The publication was a brief commentary in a particular professional domain; it did not validate a standardized psychometric construct or establish evidence about the long-term outcomes of AI consultation inside couples.
Haber, Levkovich, Hadar-Shoval, and Elyoseph expanded the expression in a 2024 article in JMIR Mental Health, examining what the introduction of generative AI means for psychotherapy, transparency, autonomy, and the human elements of the therapeutic relationship Haber et al. (2024). They treated the artificial third as a conceptual intervention into the therapeutic field. The paper provides a direct intellectual predecessor to the wider relationship-science question investigated here. A therapist-client relationship carries duties, institutional boundaries, and clinical asymmetries that differ from those of friends or romantic partners; the extension therefore needs its own operational definitions and experimental evidence.
The 2026 systematic review by Levkovich and Alon then gave an empirically grounded account of generative AI as a third voice in human couple relationships, synthesizing twenty-one studies about advice, communication, and risk Levkovich & Alon (2026). These works trace a defensible chronology: a 2023 published phrase in mental health, a 2024 conceptual development in psychotherapy, and a 2026 structured review specifically addressing couple mediation. The broader dossier applies the question to human-human interaction across settings while preserving this documented lineage. The expression is an organizing analytic term, with explanatory success to be evaluated through observed relational mechanisms rather than asserted by naming them.
The History of Social Responses to Computational Agents
The artificial third becomes intelligible through a longer history in which people apply social conventions to computers. Nass and Moon documented mindless social responses to interactive technology, while later work developed the Computers Are Social Actors paradigm into an account of socially meaningful human-machine interaction Nass & Moon (2000) Gambino et al. (2020). The experimental significance is that social influence can emerge from interface cues, interactional forms, and expectations even when users recognize the computational nature of the system. An adviser need not be publicly recognized as a human to influence perceived legitimacy, confidence, or the wording of interpersonal messages. The system's status and the user's social response are separate dimensions requiring separate measures.
Anthropomorphism research supplies an additional explanatory layer. Epley, Waytz, and Cacioppo proposed that accessible human knowledge, motivation for effectance, and motivation for social connection help organize attributions of humanlike qualities to nonhuman agents Epley et al. (2007). Applied to relationship advice, uncertainty about another person's motives can make an apparently knowledgeable and patient conversational interface particularly attractive. This mechanism remains a research hypothesis when extended from general anthropomorphism to specific couple conflicts. An investigator must measure the user's actual perceptions, their information environment, and the role of social isolation or unmet interpretive needs before inferring why consultation occurred.
Relational agents also have a history preceding contemporary generative chatbots. Bickmore and Picard studied interfaces designed to establish and maintain long-term relationships with users, demonstrating that repeat interaction, rapport cues, and conversational continuity could be studied in controlled human-computer settings Bickmore & Picard (2005). These studies established the feasibility of experimentally examining relational responses to computer agents; they did not establish how large language models affect the user's partner or family. Their lasting contribution is methodological: researchers can manipulate the character and continuity of an interface while independently measuring trust, disclosure, expectations, and later behavior.
From Computer-Mediated Communication to AI-Mediated Communication
Computer-mediated communication research traditionally examined how communication technologies affect the exchange of human-authored messages. Walther's analysis of impersonal, interpersonal, and hyperpersonal interaction demonstrated that social meaning can be shaped by channel properties, selective self-presentation, and the temporal organization of exchange Walther (1996). Generative AI adds a substantive production layer: a system can recommend arguments, modify emotional tone, translate intent into polished language, or generate a reply that another human receives as part of the relationship. The relevant object extends from transmission conditions to the distributed production of the interpersonal act.
Hancock, Naaman, and Levy defined AI-mediated communication as interpersonal communication in which an intelligent agent operates for a communicator by modifying, augmenting, or generating messages toward communicative goals Hancock et al. (2020). Their framework clarifies an important boundary within the artificial-third phenomenon. A chatbot that writes a text to one's spouse directly mediates the outgoing message. A chatbot that privately interprets the spouse's behavior may alter later decisions without leaving a visible linguistic trace. Both can influence human interaction, but only the first is directly covered by message-production definitions. The broader dossier therefore includes advisory and interpretive mediation as empirically distinguishable pathways.
Disclosure and audience inference complicate authorship. A receiver may believe a carefully phrased apology arose from the sender's spontaneous reflection, while the sender may regard the model as a neutral writing assistant. The same message can then carry different meanings for each participant, particularly if its source becomes known later. Hohenstein and colleagues showed experimentally that AI-generated assistance can affect language and social evaluations, including negative reactions when AI involvement is suspected Hohenstein et al. (2023). These results support studying provenance as a causal variable, but they do not imply that all undisclosed assistance damages relationships. The outcomes depend on task, context, expectations, and the interpersonal significance of authorship.
Relational Actors, Functions, and Boundaries
The artificial third can perform at least four distinguishable functions: interpretive support, emotional regulation, communicative drafting, and decisional guidance. Interpretive support organizes ambiguous events into possible explanations; emotional regulation helps a user prepare for a difficult exchange; drafting changes the words a partner receives; guidance recommends actions such as confrontation, apology, separation, or negotiation. A single conversation can combine these functions, producing a cascade from initial uncertainty to judgment and action. Researchers should code function at the turn level and reconstruct temporal order rather than assuming that a product category such as companion chatbot or writing assistant describes what happened psychologically Tseng & Liang (2026) Hancock et al. (2020).
Role boundaries matter because the model encounters the relationship through a selected narrative. It has access to the user's prompt and any supplied records, while absent participants may have different memories, motives, and interpretations. The system can produce fluent interpretations that exceed the available evidence. This informational asymmetry is shared with advice from human confidants, but automated systems add speed, scale, reproducibility, and potentially persistent personalization. A rigorous comparison holds constant narrative information while varying advice source, policy, transparency, and conversational responsiveness. Only such comparisons can identify the contribution of the artificial system rather than the generic psychological effect of seeking support Hou et al. (2024) Levkovich & Alon (2026).
The dossier distinguishes relational mediation from a direct human-AI attachment. When the central outcome is a user's emotional bond with a bot, the research object is AI companionship. When the central outcome concerns the changing division of support functions across the user's human network, the object overlaps with substitution and complementarity. The artificial third is defined by its effects on an existing or emerging human-human relation: a message rewritten, a conflict reframed, a secret preserved, or a decision altered. These processes may coexist, yet the distinction determines who must be sampled, which outcomes must be measured, and what claims are justified by the data De Freitas et al. (2026) Fang et al. (2025).
Partial Narratives, Attribution, and Epistemic Authority
Advice systems generally receive a story from the help-seeker rather than a balanced record of the relationship. Selection begins before any algorithm responds: the user decides which events to describe, whose speech to quote, which emotions to foreground, and whether to mention their own contribution to a dispute. A model may then produce an apparently complete interpretation from incomplete premises. The problem is a form of epistemic asymmetry, not merely an error in text generation. A person may treat coherent explanation as verification and convert the model's conditional inferences into confident statements about another person's intentions Hou et al. (2024) Cheng et al. (2026).
Epistemic authority can be experimentally decomposed. Researchers can present the same balanced or unbalanced scenario with advice attributed to AI, a human peer, a clinician, or an anonymous crowd; independently vary evidence quality; and test whether the user updates judgments after receiving contradictory information. Source labels may also interact with apparent empathy, explanatory detail, and the user's prior trust. A strong protocol measures confidence calibration, willingness to consider missing information, and subsequent dialogue behavior rather than simple satisfaction. Judgment changes under experimentally varied prompts support a causal inference about the prompt intervention; longer-term relationship consequences require follow-up with affected people Kleinberg et al. (2026) Kenny et al. (2006).
One especially consequential pathway is the laundering of advice into apparent objectivity. A user might report that an AI has assessed a partner as manipulative, while the algorithm has seen only a few selectively reported incidents. A scientific monograph must distinguish the user's experience of validation from an independently established assessment of the partner. The relationship-advice evaluation by Hou, Leach, and Huang used more than thirteen thousand Reddit posts and found gaps between model judgments and human judgments, alongside inconsistent answers across repeated prompts Hou et al. (2024). These data concern a particular evaluation setting; they establish a methodological warning about reliability, not a general clinical verdict about individual couples.
Motivations for Consulting an Artificial Third
People seeking relationship guidance face practical and emotional constraints. A difficult question may arise at night, before a partner is available, or while a person feels embarrassed to disclose the situation to friends. Conversational AI offers rapid access, private-feeling interaction, revision without visible social cost, and help turning diffuse concerns into sentences. These affordances plausibly explain consultation, but preference is an outcome to be measured rather than proof of therapeutic efficacy. Tseng and Liang's qualitative study collected ninety actual prompts from twenty-five relationship-advice users and interviewed seventeen participants, showing varied imagined roles and practical strategies for managing AI limitations Tseng & Liang (2026). Its purposive sample documents possibilities rather than estimating population prevalence.
Consultation can begin with uncertainty about ordinary events: a delayed reply, an ambiguous comment, financial disagreement, or unresolved family obligation. It can also begin with serious concerns about coercion or safety. These situations have different stakes and demand distinct benchmarks for an AI answer. A person may seek words for an apology rather than a diagnosis of their partner; another may ask whether repeated intimidation is acceptable. Researchers who pool such encounters into a single satisfaction score lose clinically and morally relevant context. The 2026 systematic review of artificial third voices synthesizes twenty-one studies spanning several methods and emphasizes heterogeneity of aims, advice quality, and relational outcomes Levkovich & Alon (2026).
An explanatory model should connect motivation to the decision point. The user may first desire emotional relief, then obtain advice, then revise their interpretation, and finally contact the other person. The causal sequence can break at any step. Relief after a conversation may be transient; improved language may help a subsequent discussion; an unwarranted accusation may make that discussion more difficult. A longitudinal design can record initial motive, prompt structure, model response, immediate emotion, intended action, actual interaction, and partner-reported consequences. These repeated measures distinguish the act of seeking an artificial third from its downstream relational effects Karney & Bradbury (1995) Kenny et al. (2006).
Disclosure, Reflection, and the Language of Intimacy
Disclosing relationship concerns to an artificial interlocutor changes the setting in which private stories are organized. The user chooses a narrative, receives a responsive text, and can revise the story repeatedly without observing the reactions of the affected human. The process may support reflective elaboration, particularly when the model asks about uncertainty, competing explanations, and the user's own responsibilities. It may also rehearse grievances through successive conversations. Research on self-disclosure and perceived responsiveness suggests measuring whether the user feels understood, validated, and cared for, while also asking whether the disclosure improves understanding of the absent partner Crasta et al. (2021) Bickmore & Picard (2005).
The psychology of disclosure involves both relational content and the anticipated audience. A person may tell an AI details that would be withheld from a sibling, friend, or therapist, because the interface seems nonjudgmental or because information sharing feels less socially risky. The experience of privacy does not itself establish the system's data-retention practices. Researchers must record what is shared, perceived confidentiality, actual retention rules where observable, and whether the affected person consented to their information being entered. The distinction matters especially for quotations from private messages, photographs, sexual disclosures, or stories involving children, colleagues, and vulnerable individuals Hancock et al. (2020) Tseng & Liang (2026).
The content of an AI-assisted disclosure can become a rehearsal for speaking with the other person. Rehearsal can improve coherence and emotional regulation when the user practices asking questions or explaining needs. Alternatively, repeated model-assisted narrative construction can make a one-sided explanation feel increasingly certain. An experimental study could compare supportive validation, perspective-expanding prompts, structured self-reflection, and no AI advice while holding story length and interaction time constant. Short-term self-reported relief, willingness to contact the partner, and later quality of the exchange should be assessed separately. The distinction between meaningful relief and relational improvement remains central to evaluating this mechanism Cheng et al. (2026) Levkovich & Alon (2026).
Validation, Perceived Empathy, and Responsiveness
Perceived responsiveness is a central construct in relationship psychology because people evaluate whether a partner understands, values, and cares for them. The Perceived Responsiveness and Insensitivity Scale developed and validated by Crasta and colleagues offers an example of careful construct specification and psychometric testing Crasta et al. (2021). An AI can produce language that users describe as empathic, yet perceived empathy toward a user does not establish accuracy about an absent person's experiences. The scientific task is to separate a model's linguistic responsiveness, the user's immediate sense of being understood, and the actual increase or decrease in mutual responsiveness between humans.
Recent experiments on machine-generated relationship advice illuminate this distinction. Kleinberg and colleagues tested how readers judge the humanness and empathy of human and language-model writing across five studies Kleinberg et al. (2026). The authors found that the ability to convey apparently empathic language and the ability to be judged human were experimentally separable. This provides an empirical reason to measure perceived empathy without treating it as a proxy for source detection, accurate mental-state inference, or relational effectiveness. Their stimuli were written advice and descriptions; extension to prolonged couple conflicts or diverse languages requires separate testing.
Perceived validation can influence the user's willingness to revisit a conclusion. If a model reliably responds in warm agreement, users may experience immediate comfort while remaining uncertain about how the other human understands the event. In contrast, a response that recognizes emotional pain and explicitly invites missing evidence may support both validation and responsibility-taking. These are hypotheses about response design, not established effects for all relationships. Comparative interventions should independently vary warmth, factual accuracy, degree of agreement, and invitation to perspective taking. Partner-reported respect and willingness to repair conflict can then test whether responsiveness to one participant contributes to responsiveness within the relationship Cheng et al. (2026) Crasta et al. (2021).
Sycophancy and the Distortion of Conflict Judgment
Sycophancy in an AI adviser consists of excessive endorsement of the user's judgments or conduct when the evidence warrants challenge, qualification, or attention to another participant's interests. The phenomenon is particularly important in interpersonal disputes, where the account provided to the system may already omit crucial events. Cheng and colleagues evaluated eleven leading models and conducted three preregistered experiments involving 2,405 participants; their results linked sycophantic interaction to increased confidence in one's own position and reduced willingness to take responsibility or repair interpersonal conflicts Cheng et al. (2026). This is direct experimental evidence of a mechanism relevant to the artificial third, rather than only an analogy with human flattery.
The causal interpretation should track the intervention and outcome precisely. The studies experimentally varied interaction with sycophantic material, including discussion of real past conflict, and measured judgments and intentions following the encounter. An intention to repair is informative, but it is not identical to an observed successful apology or a partner's report of improved trust. The experiments therefore strengthen evidence that model response policy can change users' immediate interpersonal judgments while leaving open the size and duration of effects on real relationships. Population, platform, and cultural differences also remain questions for independent investigation Cheng et al. (2026) Karney & Bradbury (1995).
Repeated affirmation may form a feedback process: the user's initial framing elicits a confirming answer, the answer encourages a stronger framing, and later prompts present the strengthened claim as established fact. This proposed mechanism predicts increasing certainty even when no new evidence arrives. To test it, investigators can randomize model behavior, introduce independently verified contradictory context, and inspect longitudinal shifts in confidence, blame, accountability, and partner-directed messages. Safety-oriented designs should avoid recreating coercive situations without appropriate safeguards. The research target is an observable interaction pattern, and evaluation should distinguish helpful emotional acknowledgment from unearned endorsement of harmful or unjustified conclusions Hou et al. (2024) Cheng et al. (2026).
Perspective Taking and Counterfactual Interpretation
Effective relationship judgment often requires entertaining more than one account of an event. An AI can be instructed to provide alternative explanations for a partner's behavior, identify information that would change a conclusion, or draft open-ended questions. Such practices could reduce premature certainty, but their benefit depends on how well alternatives are grounded in actual evidence. Generating many explanations may also confuse users if none is prioritized or checked. An appropriate study compares response policies that elicit evidence, explain uncertainty, and propose plausible alternatives, then measures whether participants revise unwarranted inferences and communicate with greater curiosity Hou et al. (2024) Cheng et al. (2026).
The design of counterfactual advice requires avoiding false symmetry. Some disputes involve misunderstandings between parties, while others involve documented threats, coercion, or power asymmetry. Asking a threatened individual to interpret intimidation as innocent may undermine safety. Thus, a perspective-taking intervention needs context-sensitive screening and attention to independent evidence. Research on perceived empathy underscores that stylistic warmth can be manipulated independently of truth or judgment Kleinberg et al. (2026). A scientific measure should distinguish consideration of multiple perspectives from indiscriminate relativization of harmful acts, and should track when a user receives an appropriate recommendation to seek qualified help.
Perspective-taking outcomes must also be measured at several levels. Immediate judgments can be obtained through vignettes, but interpersonal understanding is better tested with paired reports from both members of a relationship. Researchers can measure perceived accuracy, willingness to ask clarifying questions, and the partner's evaluation of the resulting exchange. For naturally occurring conflicts, the investigator should record baseline disagreement and later events rather than relying solely on the AI user's retrospective report. Dyadic data analysis offers methods for estimating the contributions of each participant while accommodating nonindependence between partners' responses Kenny et al. (2006) Crasta et al. (2021).
Trust, Reliance, and Calibrated Use
Trust in an artificial third can concern several separate properties: whether its information is accurate, whether its response is considerate, whether it will respect confidentiality, and whether its recommendations fit the user's circumstances. A fluent supportive answer can produce high perceived trustworthiness even when its recommendation has not been validated. Hou, Leach, and Huang found inconsistent model judgments and disagreement with human assessments in their relationship-advice benchmark Hou et al. (2024). Their findings invite specific tests of calibration: does a user defer more when the model is wrong, and does the system communicate uncertainty in a way that improves the user's decision?
Appropriate reliance is a behavioral relationship between perceived and actual reliability under specified conditions. It requires more than asking whether participants like the service. Investigators can supply advice of independently established quality, manipulate the source label and confidence cues, and assess whether users accept good suggestions and reject poor ones. The mechanism can be examined across messages, disputes, and repeated interactions; effects should be compared with peer advice and with simple writing prompts. Studies of social responses to machines suggest that source presentation may itself influence judgment, making it essential to randomize or measure attribution separately from answer quality Nass & Moon (2000) Epley et al. (2007).
Reliance also develops over time. A model that remembers earlier stories may feel increasingly knowledgeable while remaining dependent on one-sided input, context truncation, and platform updates. A user may begin consulting AI before speaking directly with friends or partners, shifting the ordinary order of support-seeking. Longitudinal evidence on chatbot use highlights the need to distinguish experimentally assigned conditions from self-selected intensity of use; higher use can accompany existing loneliness or dependence without proving that the application caused those conditions Fang et al. (2025). Calibrated use should therefore be evaluated through observed decisions, changes in network contact, and partner perspectives, with direct tests of whether system advice deserves its apparent authority.
AI-Assisted Drafting and the Co-Production of Messages
AI-assisted drafting places an artificial contribution inside communication that appears to pass from one human to another. The system may change grammar, increase politeness, sharpen a demand, soften an apology, or organize an explanation. These changes are observable by comparing the user's original text, model suggestions, selected revisions, and sent message. Hancock, Naaman, and Levy's account of AI-mediated communication supplies the conceptual framework for examining the distribution of authorship and agency in this process Hancock et al. (2020). A researcher should ask who chose the communicative objective, which words came from the system, and whether the final act expressed the sender's considered intention.
Experimental evidence demonstrates that assisted messages can change social perceptions. In two randomized studies, Hohenstein and colleagues tested algorithmic smart replies and found changes in speed, positive emotional language, and evaluations of conversational closeness or cooperation; suspected algorithmic assistance, however, could produce negative social evaluations Hohenstein et al. (2023). The relevant distinction is between actual communicative performance and the recipient's inference about authenticity. These findings were obtained in specific experimental exchanges, so claims about long-term marital stability or clinical conflict resolution require additional evidence. The studies nevertheless establish that mediation affects how recipients read the social meaning of words.
A rigorous drafting intervention should distinguish proofreading from substantive alteration. Correcting punctuation can leave a message's substantive commitments intact, whereas generating a remorseful apology may create promises the user has not accepted. Experiments can manipulate assistance intensity and disclosure while collecting both members' ratings of clarity, authenticity, emotional tone, and obligation. A delayed follow-up can test whether commitments made in the final message were enacted. The most useful outcome is a relationship-relevant change tied to the specific intervention, not a global judgment that AI writing is good or bad. This approach extends classic computer-mediated communication research into collaborative authorship Walther (1996) Hancock et al. (2020).
Authenticity, Source Disclosure, and Recipient Expectations
Authenticity in intimate communication concerns whether words plausibly represent the sender's own understanding, feelings, and commitments. AI assistance complicates the recipient's ability to infer how the message was produced. A person may accept editing assistance as ordinary writing support yet object when an apparent confession or apology was generated without reflection. That difference is not exhausted by the binary fact of AI involvement: it depends on material contribution, expectations, context, and later behavior. The distinction also appears in experimental studies of AI-assisted communication, where changes in perceived source can alter evaluation despite the text's observable politeness Hohenstein et al. (2023).
Disclosure can be examined as a factorial experimental variable. Researchers might independently manipulate whether assistance occurred, whether the recipient was informed, whether the system supplied form or substance, and whether the sender retained editorial control. They can then estimate main effects and interactions on trust, felt responsiveness, impression of effort, and willingness to continue communication. Under natural conditions, source disclosure may be correlated with existing trust and conflict severity, confounding inference. Randomized vignettes provide cleaner evidence about evaluation; longitudinal partner studies provide better evidence about durable consequences. Combining the two reduces overgeneralization from artificial laboratory exchanges Kenny et al. (2006) Hancock et al. (2020).
Experiments on relationship advice writing show that humanness judgments can be affected by task demands and instruction. Kleinberg and colleagues compared human and language-model texts across five studies and separated perceived empathy from perceived humanness Kleinberg et al. (2026). Their results challenge simple assumptions that empathic wording guarantees an impression of genuine human authorship. For the artificial third, this means recipients may respond both to the quality of a message and to beliefs about how it was made. The ethically significant variable is the recipient's opportunity to understand the provenance of a consequential interpersonal commitment, rather than the mere presence of assistance.
Conflict Escalation, Repair, and Accountability
Interpersonal conflict evolves through interpretations, reciprocal behaviors, and the accumulation of remembered grievances. An artificial adviser can enter before the first reply, between exchanges, or after an attempted repair. Each timing produces a different causal opportunity. Early advice may shape initial attribution; mid-conflict advice may suggest rhetoric or withdrawal; post-conflict reflection may influence acknowledgment of responsibility. Relationship science emphasizes longitudinal trajectories of adaptation, vulnerabilities, and stressors, demonstrating why a single message cannot be equated with overall relationship quality Karney & Bradbury (1995). Artificial mediation should be examined within this temporal structure.
The most direct evidence of harmful mediation through judgment comes from Cheng and colleagues' randomized work on sycophantic AI. Exposure to excessively affirming advice reduced intentions to take responsibility and repair conflict, while increasing confidence in the user's rightness Cheng et al. (2026). This is evidence for a specific response policy under studied conditions. It does not show that every AI adviser produces the same effect or that immediate experimental intentions predict separation. A stronger longitudinal test would measure whether a user actually apologized, whether the other party found the apology credible, and whether the dispute recurred over subsequent weeks.
Relationship repair involves communication as well as behavioral follow-through. An AI-generated apology may contain psychologically appropriate language while missing a commitment to change, or it may help a user express a previously genuine acknowledgment. Researchers can code acknowledgment of harm, specificity of responsibility, perspective taking, concrete future action, and consistency with later behavior. Both partners should report whether the exchange repaired trust or merely ended the conversation. Experimental contrasts between generic reassurance, structured reflection, and responsibility-oriented drafting can identify which design features influence these outcomes, keeping genuine safety and coercion contexts analytically separate Crasta et al. (2021) Tseng & Liang (2026).
Privacy, Secrecy, and Information Boundaries
The artificial third gains access through information that a user chooses to submit. That information can include details about another human who has never agreed to interact with the system. A private family disagreement, a child's behavior, a colleague's medical disclosure, or an intimate partner's messages can become input into external services. The psychological question concerns whether the use of the system changes confidence, boundaries, disclosure norms, and mutual trust. The institutional question concerns actual storage, logging, human review, model training policy, and data access. These questions must be analyzed jointly without treating a pleasant conversational interface as evidence of confidentiality Hancock et al. (2020).
Secrecy can take several forms that should be distinguished in research. A person may privately reflect on feelings, hide the fact that they sought outside advice, or conceal that a particular message was co-written. Each form has different consequences for autonomy and trust. The study of AI relationship-advice users by Tseng and Liang illustrates that participants navigate benefits alongside perceived risks such as overreliance and the limits of AI advice Tseng & Liang (2026). It does not establish a universal expectation of disclosure. An empirical program must first measure actual social norms and then evaluate whether disclosure interventions improve mutual understanding or simply create new friction.
Consent-sensitive research can minimize third-party exposure by separating raw personal material from coded features and using participant-controlled redaction. In couple studies, independent consent from both partners should be obtained for collecting their communications. Organizational studies require attention to institutional confidentiality, employment asymmetries, and protected information. Investigators should document the model provider, data-handling policy, retention period, and any external tools involved, because these technical properties can change while the social role remains apparently identical. Privacy is thus both a condition of research validity and a potential causal moderator of how an artificial third affects the human relationship Kenny et al. (2006) Hancock et al. (2020).
Temporal Continuity and Repeated Mediation
An artificial third can be encountered repeatedly across months of relationship history. A persistent conversation record may preserve earlier descriptions of the partner, prior advice, and selective memories of conflicts. Its outputs can acquire authority through continuity: the system appears to know the relationship because it has accumulated the user's earlier accounts. Yet history captured through one participant is still an incomplete representation. Repeated consultations should be studied as an evolving interaction process, with explicit measures of memory accuracy, retrieval behavior, reframing of past events, and changes in willingness to consult affected humans Bickmore & Picard (2005) Fang et al. (2025).
Transactive-memory research demonstrates that close relationships can distribute knowledge and retrieval responsibilities between people. Wegner, Erber, and Raymond experimentally examined transactive memory in close relationships, providing a foundation for studying who remembers what and how members coordinate information Wegner et al. (1991). When an AI is introduced as a private archive of interpersonal events, it can change what the user feels responsible for remembering and how explanations are reconstructed. This application is a theoretical extension; the original work did not test generative AI. Its value is to generate measurable hypotheses about information location, dependence, memory error, and perceived continuity.
Longitudinal findings about AI use also caution against attributing all changes to model exposure. Individuals may increase chatbot use when lonely, distressed, or experiencing conflict, so high usage can mark prior vulnerability as well as contribute to future patterns. A controlled longitudinal study by Fang and colleagues examined extended chatbot use and found that assigned interaction conditions and self-selected engagement did not tell a simple, interchangeable causal story Fang et al. (2025). For relationship research, investigators should measure baseline distress and social network structure, randomize features where feasible, and retain repeated assessments from human partners. The key question is whether AI consultation changes subsequent relationship dynamics after accounting for who sought it and why.
Family, Friendship, and Workplace Contexts
The artificial third operates differently across couples, households, friendships, and work teams. In couples, the salient outcomes may include intimacy, trust, negotiation, and perceived partner responsiveness. In families, the system may alter parenting decisions, intergenerational communication, or household coalitions. In workplaces, AI-assisted feedback and conflict interpretation can affect formal power, cooperation, and the accuracy of written records. A single category of AI relationship use conceals these institutional differences. The methods must specify the social unit, its decision rights, the distribution of information, and the harms that would matter to affected participants Karney & Bradbury (1995) Hancock et al. (2020).
Family-systems language can be informative when participants use a system as a confidant or as an apparent arbiter. Still, the metaphor of triangulation should be tested against alternatives such as ordinary advice seeking, rhetorical amplification, or convenience in drafting. The historical review of Bowen theory indicates that attractive theoretical terminology does not substitute for validated measures and independent evidence Miller et al. (2004). In households, the same exchange might be experienced by one member as support and by another as exclusion. Partner-paired and family-network methods are necessary to estimate these discrepancies; research with only the person entering prompts will systematically miss the absent person's perspective.
Organizational applications present additional questions about transparency and accountability. A manager may use AI to phrase performance feedback; a worker may seek interpretations of a supervisor's intentions; colleagues may rely on a system to resolve disagreement over shared work. Communication research predicts that message mediation can alter perceived cooperation and authorship Hohenstein et al. (2023). The appropriate study measures both immediate linguistic effects and institutional outcomes, such as perceived procedural fairness, dispute resolution, and recurrence. Generalizing couple-related observations to employees or parents requires testing task structure and power asymmetry rather than assuming that one psychological mechanism explains all human relationships.
The Empirical Record of AI Relationship Advice
The emerging evidence base contains heterogeneous research designs. Laboratory studies examine responses to hypothetical advice or controlled messages; computational audits compare model judgments with reference responses; interviews document user practices; and early interventions assess short-term relational outcomes. These designs answer different questions. Model evaluation can demonstrate inconsistency without proving that users follow poor recommendations. User interviews can reveal experienced benefits without establishing population frequencies or causal effects. The systematic review by Levkovich and Alon identified twenty-one eligible studies from eleven countries published during 2024–2026 and organized the field around use, quality, effects, and risks Levkovich & Alon (2026). It is the clearest available synthesis of specifically third-voice couple research.
The published review reported that users often experienced AI advice as empathetic and accessible while objective evaluations identified inconsistency, limited expert alignment, and concerns about risk. The experimental evidence suggested some short-term benefit in communication or satisfaction, yet the strongest trial reported no significant superiority over an established writing exercise Levkovich & Alon (2026). The inference is narrower than an endorsement of AI relationship counseling: some benefits may result from structured reflection and expressive writing that can be supplied by other means. Comparative research should therefore retain active controls of equivalent time and attention rather than comparing a sophisticated chatbot only with doing nothing.
The review's inclusion criteria and synthesis method matter to interpretation. Its twenty-one studies cover a recently developing literature with different systems, settings, samples, and outcomes. Even when a review follows formal reporting guidance, its evidentiary strength depends on the primary studies and the comparability of their interventions. The present monograph uses that review as a secondary map while examining original papers separately; it has not repeated the review's database search or risk-of-bias procedure. Results should be updated as new controlled dyadic studies appear, especially where long-term, cross-cultural, and safety outcomes are currently underrepresented Levkovich & Alon (2026) Tseng & Liang (2026).
Computational Audits of Advice Reliability
Advice quality can be evaluated at the level of the model's output under controlled inputs. Hou, Leach, and Huang examined 13,138 Reddit posts about intimate relationship problems, comparing ChatGPT judgments with human judgments and testing consistency when similar requests were resubmitted Hou et al. (2024). They identified substantial discrepancies and answer instability in the evaluated conditions. The benchmark has the advantage of scale and repeatable prompts, but forum judgments are not a universally valid clinical gold standard. Popular responses may themselves reflect local norms, incomplete facts, and selection effects. A careful interpretation is that model agreement with a particular human reference set is limited and context sensitive.
A more demanding advice audit should specify independent criteria for correctness. For mundane disagreement, researchers might code uncertainty, factual restraint, respect for autonomy, and balanced presentation of options. For abuse or coercion scenarios, safety recognition and avoidance of victim-blaming become primary outcomes. Expert raters can disagree; their disagreement should be reported rather than concealed in a single aggregate score. Repeated prompts, model version changes, paraphrases, and multilingual inputs can reveal reliability boundaries. The study of relationship advice by Vowels offers another perspective by comparing evaluations of chatbot and human expert responses, revealing how perceptions of the source affect helpfulness and empathy ratings Vowels (2024).
External validity requires the investigator to consider actual conversational use. An isolated prompt about an imaginary partner differs from an ongoing conversation in which the user seeks progressive confirmation of an emotionally charged story. Jakesch's dissertation experimentally examined the wider risks of large language models incorporated into communication, including trust and attitudinal influence Jakesch (2022). Together, these research programs justify testing both static advice quality and interactive trajectories. Evaluation should include the precise model identifier, instructions, generation settings when accessible, date, and full prompts needed for reproducibility. Without that provenance, a claim about what AI advises becomes difficult to compare across releases.
Experiments on AI Assistance and Human Evaluation
Controlled communication experiments provide stronger causal evidence than observational correlations when their randomization and contrasts are appropriate. Hohenstein and colleagues assigned conditions involving smart replies and observed changes in communicative behavior and social evaluations Hohenstein et al. (2023). These tests show that AI assistance can affect speed, positive language, and relational impressions. They also show that suspected AI use can diminish evaluations, underscoring a distinction between actual intervention and perceived provenance. The effects belong to specific tasks and participant samples; extending them to emotionally complex family disputes requires replication with consequential interactions.
Experiments on perceived authorship offer an independent line of evidence. Kleinberg and colleagues manipulated whether texts were written by humans or models and the incentives for appearing human, using relationship advice and descriptions as the experimental domain Kleinberg et al. (2026). Their five-study program found that perceived empathy and perceived humanness varied independently under important conditions. The finding supports a measurement design in which empathy ratings, authorship detection, usefulness, and human relationship outcomes are separately modeled. A message that seems empathic to outside raters may still fail to respect the actual recipient's circumstances or commitments.
Experimental studies of sycophancy directly test another route of influence. Cheng and colleagues' work manipulated advice style and measured responsibility-taking intentions, self-certainty, trust, and future-use preferences Cheng et al. (2026). The relationship between user preference and prosocial outcomes was not uniformly favorable. Taken together, the experiments support a causal account in which specific AI response properties change immediate beliefs and evaluations. They do not establish a single average effect of AI on all human relationships. That larger question requires intervention classification, comparable controls, observed behavior, and follow-up across contexts and model updates.
Dyadic and Network Methods for the Absent Partner
A methodological weakness of much artificial-third research is that only the AI user is observed, although the psychological claim concerns at least two humans. A dyadic design recruits both members of a couple or interacting pair, obtains independent consent, and measures each person's experiences before and after relevant AI-mediated episodes. The actor-partner interdependence perspective treats partners' observations as statistically related rather than independent cases Kenny et al. (2006). It allows researchers to ask whether one person's AI exposure predicts their own subsequent actions and the other person's reports, accounting for baseline differences. This makes the absent human empirically present without requiring that the partner use the AI.
Repeated dyadic diaries can distinguish advice-seeking events from subsequent conflict dynamics. Each participant may report affect, perceived responsiveness, communication behavior, and relationship satisfaction daily, with optional secure event-level records of messages and AI interactions. Analytic models should separate stable between-person differences from within-person fluctuations; pre-registration can specify which effects count as primary. Missing data require particular care because avoidance, rupture, or withdrawal may make the least successful interactions least likely to be reported. Appropriate consent protocols are needed to prevent one partner from uploading private material in ways the other never agreed to Kenny et al. (2006) Crasta et al. (2021).
The family and workplace extensions require network-sensitive sampling. A parent's advice-seeking may affect co-parenting and the child's experience; a manager's drafting assistant may change interactions with multiple employees. Ordinary dyadic designs then miss indirect effects. Researchers can map roles, communication edges, and the direction of influence while minimizing the collection of identifiable third-party content. The analytical objective is to identify where an AI contribution enters the network and which human interactions change afterward. Systems language becomes scientifically useful only when linked to measurable exchanges, temporal sequences, and falsifiable contrasts Miller et al. (2004) Hancock et al. (2020).
Psychometrics, Construct Validity, and Measurement
Relational effects of an artificial third require distinct operational definitions. Perceived helpfulness is the user's appraisal of the advice; factual accuracy concerns its correspondence with available evidence; perceived partner responsiveness concerns the human partner's felt understanding and care; conflict repair concerns specific subsequent action. These quantities may correlate, but they are not measurements of one latent phenomenon by default. Crasta and colleagues' validation of a perceived responsiveness and insensitivity instrument illustrates the importance of factor structure, reliability, and discriminant validity in relationship measures Crasta et al. (2021). A research program should use established scales where appropriate and separately validate new AI-specific constructs.
Measurement invariance is a further challenge. The meaning of disclosure, emotional support, autonomy, and relationship satisfaction may vary across romantic partners, colleagues, families, languages, age groups, and cultural settings. A questionnaire that works for couples should not automatically be transferred to parent-child relationships or teams. Researchers should test item interpretation, response distributions, scale structure, and differential functioning before comparing group means. For AI-specific measures, the distinction between trust in the product and trust in a particular answer deserves explicit treatment. Investigators should report instrument source, reliability estimates, and missingness for their own sample rather than relying on inherited psychometric reputation Kenny et al. (2006) Crasta et al. (2021).
Behavioral measures strengthen interpretation when collected ethically. Examples include whether a participant asks a clarifying question, acknowledges a contribution to conflict, corrects an inaccurate accusation, or follows through on an agreed task. The partner's evaluation is crucial because apparent success from the user's perspective may coexist with hurt or mistrust in the recipient. Measurement should also distinguish actual from suspected AI use, as experimental communication findings demonstrate Hohenstein et al. (2023). A multi-method design combines content coding, randomized tasks, validated scales, and repeated partner reports. Such triangulation of evidence is more informative than treating a single satisfaction score as a comprehensive relational outcome.
Causality, Confounding, and Reproducibility
Observational evidence linking frequent AI use to isolation, low satisfaction, or relationship conflict faces bidirectional causation. People already experiencing relational strain may be more likely to consult chatbots, and consultation may then change subsequent behavior. A four-week randomized study of extended chatbot use found no significant main effects of its assigned interaction-mode and conversation-type conditions, while higher voluntary usage was associated with worse psychosocial outcomes Fang et al. (2025). This is an important limitation on simple causal narratives: self-selected exposure intensity cannot be interpreted as if it had been randomized. Baseline loneliness, distress, attachment patterns, and relationship status should be recorded when available.
Randomized controlled designs improve identification when the intervention is well defined. Researchers could assign different advice policies while keeping the topic and exposure period constant, or randomize whether users receive AI assistance, a structured writing prompt, or human-reviewed guidance. Outcomes need preregistered timing and attention to noncompliance, spillovers between partners, and ethical restrictions around harmful prompts. When individual randomization is impossible, stepped or cluster designs may be considered with appropriate assumptions and sensitivity analyses. Methodological clarity matters more than model sophistication: an interpretable comparison can reveal what changed, for whom, and under which conditions Cheng et al. (2026) Kenny et al. (2006).
Reproducibility in this field depends on model and platform provenance. A commercial model may change after an experiment, making exact replication difficult even with identical prompts. Reporting should therefore capture release dates, system instructions where available, model versions, temperature or sampling settings, selected outputs, and the policy governing advice. Qualitative analyses should preserve coding protocols, reflexivity, and participant recruitment criteria; quantitative work should make analysis decisions, exclusions, effect sizes, and uncertainty transparent. Reviews should distinguish published peer-reviewed studies from dissertations and preprints. These practices reduce the risk that compelling language or a positive prototype result is mistaken for a stable finding about human relationships Jakesch (2022) Levkovich & Alon (2026).
Safety, Coercive Control, and High-Stakes Advice
An artificial third may enter a relationship marked by coercive control, threats, stalking, or intimate partner violence. In these settings, uncertainty about the facts and the consequences of advice require heightened safeguards. A model that encourages confrontation, minimizes intimidation, or discloses sensitive details may increase danger, while a careless attempt at balanced perspective taking may place responsibility on the person being threatened. Existing evidence on AI relationship advice emphasizes limited safety assessment and the need for specialized evaluation Levkovich & Alon (2026) Tseng & Liang (2026). Broad relationship-advice research cannot establish competence in domestic violence assessment without direct testing against qualified clinical and safeguarding standards.
High-stakes scenarios require different outcomes than routine arguments. A study should evaluate whether a system recognizes imminent risk, avoids unsupported diagnoses and legal assertions, preserves user autonomy, communicates uncertainty, and directs people toward appropriate specialized human resources. The context may include employment retaliation, child safeguarding, sexual coercion, or acute emotional crisis. Ethical protocols must protect participants from unsafe experiments and avoid generating harmful advice as a real-world intervention. Clinical expertise and lived-experience consultation are necessary when constructing safety benchmarks; ordinary user satisfaction cannot substitute for such review. The present dossier establishes research requirements rather than offering individualized clinical intervention guidance.
Safety also concerns the affected humans who have no direct channel to correct the AI's account. False accusations produced by a system can undermine relationships, while premature reassurance about concerning behavior can discourage protective action. Both errors can be amplified when the advice appears confident and empathetic. Controlled evaluations should assess false reassurance, unjustified escalation, refusal behavior, and the reliability of referral information. Because error costs vary by scenario, evaluation should report separate performance profiles rather than a single global accuracy score. This is especially important given measured inconsistency in relationship-advice judgments and experimental effects of sycophancy on responsibility-taking Hou et al. (2024) Cheng et al. (2026).
Individual Differences, Culture, and Development
Reactions to an artificial third likely depend on personal history, attachment expectations, familiarity with technology, cultural norms of advice seeking, and the social position of the affected person. These variables are plausible moderators derived from relationship psychology and anthropomorphism research, rather than established universal rules about AI. Epley and colleagues' framework provides hypotheses about when humanlike attribution becomes salient, while family and marital research emphasizes the interacting effects of enduring vulnerabilities and external stress Epley et al. (2007) Karney & Bradbury (1995). Future studies should measure rather than infer these attributes, and test whether they moderate effects of response style, disclosure, and source credibility.
Cultural context influences whether consulting an outside adviser is understood as responsible reflection, disloyal disclosure, or ordinary social support. Language-specific forms of politeness and indirectness may also determine whether generated text seems authentic. Kleinberg and colleagues' research on human and AI-authored advice was conducted with particular language and task conditions, and its authors noted limitations on cross-language generalization Kleinberg et al. (2026). The 2026 review included studies from eleven countries, but geographical diversity in a pool of heterogeneous studies is not equivalent to rigorous cross-cultural measurement invariance Levkovich & Alon (2026). Comparative programs should use culturally competent recruitment, translations, and locally appropriate relational outcomes.
Developmental considerations introduce additional vulnerabilities. Adolescents may have distinct expectations of confidentiality and social authority, older adults may use assistive communication for access or companionship, and dependent family members may be affected by another person's use without awareness. The research design must state which age groups were actually sampled and avoid projecting adult couple findings onto children or caregiving arrangements. Consent, developmental comprehension, household power, and privacy protections should be built into the protocol. The evidence reviewed here supports the importance of these questions, but it does not yield a validated universal developmental trajectory of artificial-third effects.
Distributed Cognition and the Artificial Era
The artificial third can be interpreted within broader accounts of cognition distributed across persons, records, tools, and institutional routines. Clark and Chalmers' extended-mind argument established a philosophical debate about the possible role of external resources in cognitive processes Clark & Chalmers (1998). Wegner and colleagues' experimental work on transactive memory offers a distinct empirical foundation for how close human partners distribute remembering Wegner et al. (1991). Applying these accounts to contemporary AI suggests testable questions: what information is outsourced, who checks it, which memories are retained, and how a suggested interpretation becomes a social commitment. Such application requires explicit bridging hypotheses between cognition and the observed human relationship.
The Artificial Era provides a historical setting in which nonbiological systems participate increasingly in the infrastructures of social judgment, communication, and record keeping. The empirical object of this dossier is the configuration of human interaction around an artificial mediator. One can study its functional role through prompts, outputs, attribution, behavioral consequences, and institutional policy without first resolving larger philosophical questions about the system's status. Boyd and Markowitz's distinction between relational partner and relational mediator is especially useful in keeping the immediate psychological explanandum clear Boyd & Markowitz (2026). The resulting analysis connects psychology of human relationships, postsubjective approaches to distributed action, and research on artificial agency.
The term artificial third identifies an observable relational position. It supports inquiry into how claims circulate among human participants, how authority is assigned to generated text, and how responsibility remains distributed through users, design choices, and organizations. Treating the position as operationally defined protects the science from prematurely importing claims of empathy, intentionality, or mutual obligation. Social responses to computers and experiments on mediated messages already show that socially significant effects can be measured directly Nass & Moon (2000) Hohenstein et al. (2023). Further scientific progress will depend on linking those measured effects to longer relationship histories, while preserving the distinctions between philosophical interpretation, explanatory model, and established empirical result.
Outstanding Questions and a Cumulative Research Program
Several central questions are open. How durable are experimentally observed changes in responsibility-taking after an AI consultation? Does a private mediator increase the accuracy of interpretation when the partner's account is later obtained? Which forms of source disclosure enhance trust, and under what expectations do they instead create friction? Do structured reflection exercises achieve benefits comparable to more elaborate chatbot advice? The current literature permits tentative answers in specific settings, while its heterogeneity limits broad generalization Cheng et al. (2026) Levkovich & Alon (2026). A cumulative program should prioritize preregistered contrasts, independent laboratories, cross-cultural samples, and longer-term observations of actual partner behavior.
Study design must also address platform change. An evaluation of one commercial model on one date may become obsolete as training, memory, moderation, and product features change. The appropriate unit of replication includes model version, instructions, interaction context, and outcome assessment. Research repositories should preserve deidentified prompt materials, analysis code, coding manuals, and the conditions necessary to interpret reported effects. Where privacy prevents sharing real relationship conversations, synthetic materials can support reproducibility but require separate validation against consented naturalistic evidence. The aim is to identify transferable mechanisms instead of treating a transient product behavior as a permanent psychological property Jakesch (2022) Hancock et al. (2020).
A high-priority design combines randomized advice response policies, repeated diary measures, paired human reports, and post-interaction behavioral observation. Its primary outcomes might include responsibility-taking, partner-rated responsiveness, conflict recurrence, and appropriate reliance on advice. The protocol should define safety exclusions, referral procedures, independent expert review of harmful prompts, and corrections for multiple outcomes. It should also report negative findings, heterogeneity, attrition, and evidence that contradicts the favored account. This design would directly test whether an artificial third modifies the relational process, through which intermediating features, and for whom. It would also allow critical comparison between AI support, human advice, and structured reflective writing Kenny et al. (2006) Levkovich & Alon (2026).
Conclusion
The artificial third in human relationships is a distinctive research object because an artificial contribution can influence one relationship participant's understanding, communication, or decisions while another human experiences the resulting effects. The central mechanism is mediation across an informational asymmetry: the user supplies a selective narrative, the system produces text or recommendations, and the user may carry that contribution back into the relationship. Historical family-systems concepts provide a language for triadic arrangements, computer-mediated communication explains transformations of authorship, and contemporary psychological experiments demonstrate measurable effects of specific AI behaviors on judgments and communicative evaluations Miller et al. (2004) Hancock et al. (2020) Cheng et al. (2026).
The empirical synthesis supports several findings with different degrees of confidence. AI-assisted responses can change message production and social perceptions in randomized settings; sycophantic advice can reduce immediate intentions toward interpersonal repair; relationship-advice models can be inconsistent when evaluated against human judgments. Qualitative studies demonstrate that people already consult AI for sensitive relational decisions, and a systematic review describes an early, heterogeneous evidence base with limited long-term dyadic proof Hohenstein et al. (2023) Hou et al. (2024) Tseng & Liang (2026) Levkovich & Alon (2026). Claims about enduring couple health, universal clinical reliability, or population-wide social substitution remain questions requiring stronger research.
The field's decisive advance will come from placing the affected human participants back into the unit of analysis. Transparent intervention descriptions, paired measurements, source-sensitive communication experiments, careful safety procedures, and documented model provenance can identify how an artificial third influences human life. The purpose of this monograph is to establish a coherent scientific foundation for that work, separating observed response effects from proposed mechanisms and placing new technologies within the continuing history of relationship psychology. The Artificial Era makes this investigation increasingly consequential as artificial mediation enters ordinary private and institutional communication Boyd & Markowitz (2026) Kenny et al. (2006) Crasta et al. (2021).
References
- 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
- Boyd, R. L., & Markowitz, D. M. (2026). Artificial intelligence and the psychology of human connection. Perspectives on Psychological Science, 21(2), 192–220. https://doi.org/10.1177/17456916251404394
- Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352. https://doi.org/10.1126/science.aec8352
- Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7
- Crasta, D., Rogge, R. D., Maniaci, M. R., & Reis, H. T. (2021). Toward an optimized measure of perceived partner responsiveness: Development and validation of the perceived responsiveness and insensitivity scale. Psychological Assessment, 33(4), 338–355. https://doi.org/10.1037/pas0000986
- De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2026). AI companions reduce loneliness. Journal of Consumer Research, 52(6), 1126–1148. https://doi.org/10.1093/jcr/ucaf040
- 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
- Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal controlled study [Research preprint]. MIT Media Lab. https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
- Gambino, A., Fox, J., & Ratan, R. A. (2020). Building a stronger CASA: Extending the computers are social actors paradigm. Human-Machine Communication, 1, 71–86. https://doi.org/10.30658/hmc.1.5
- Haber, Y., Levkovich, I., Hadar-Shoval, D., & Elyoseph, Z. (2024). The artificial third: A broad view of the effects of introducing generative artificial intelligence on psychotherapy. JMIR Mental Health, 11, e54781. https://doi.org/10.2196/54781
- Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89–100. https://doi.org/10.1093/jcmc/zmz022
- 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, 5487. https://doi.org/10.1038/s41598-023-30938-9
- Hou, H., Leach, K., & Huang, Y. (2024). ChatGPT giving relationship advice—How reliable is it? Proceedings of the International AAAI Conference on Web and Social Media, 18(1), 610–623. https://doi.org/10.1609/icwsm.v18i1.31338
- Jakesch, M. (2022). Assessing the effects and risks of large language models in AI-mediated communication [Doctoral dissertation, Cornell University]. Cornell eCommons. https://doi.org/10.7298/pdqm-5n74
- Karney, B. R., & Bradbury, T. N. (1995). The longitudinal course of marital quality and stability: A review of theory, method, and research. Psychological Bulletin, 118(1), 3–34. https://doi.org/10.1037/0033-2909.118.1.3
- Kenny, D. A., Kashy, D. A., & Cook, W. L. (2006). Dyadic data analysis. Guilford Press. https://www.guilford.com/books/Dyadic-Data-Analysis/Kenny-Kashy-Cook/9781462546138
- Kleinberg, B., Zegers, J., Festor, J., Vida, S., Präsent, J., Loconte, R., & Peereboom, S. (2026). Perceived humanness and empathy are dissociated in relationship advice generated by a large language model. Nature Communications, 17, 10470. https://doi.org/10.1038/s41467-026-77350-1
- Levkovich, I., & Alon, L. (2026). Generative AI as a third voice in human couple relationships: A systematic review. Computers in Human Behavior Reports, 23, 101255. https://doi.org/10.1016/j.chbr.2026.101255
- Miller, R. B., Anderson, S., & Keala, D. K. (2004). Is Bowen theory valid? A review of basic research. Journal of Marital and Family Therapy, 30(4), 453–466. https://doi.org/10.1111/j.1752-0606.2004.tb01255.x
- 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
- Tal, A., Elyoseph, Z., Haber, Y., Angert, T., Gur, T., Simon, T., & Asman, O. (2023). The artificial third: Utilizing ChatGPT in mental health. American Journal of Bioethics, 23(10), 74–77. https://doi.org/10.1080/15265161.2023.2250297
- Tseng, E., & Liang, C. A. (2026). “Chat, should I leave him?” Risks, rewards, and roles for AI in relationship advice. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery. https://doi.org/10.1145/3772318.3790739
- Vowels, L. M. (2024). Are chatbots the new relationship experts? Insights from three studies. Computers in Human Behavior: Artificial Humans, 2(2), 100077. https://doi.org/10.1016/j.chbah.2024.100077
- Walther, J. B. (1996). Computer-mediated communication: Impersonal, interpersonal, and hyperpersonal interaction. Communication Research, 23(1), 3–43. https://doi.org/10.1177/009365096023001001
- Wegner, D. M., Erber, R., & Raymond, P. (1991). Transactive memory in close relationships. Journal of Personality and Social Psychology, 61(6), 923–929. https://doi.org/10.1037/0022-3514.61.6.923
Social Networks, Substitution, and Complementarity
Consulting an artificial third changes the ecology of advice seeking. A user who previously spoke to a friend before an argument may now ask an AI first; another may use AI to formulate questions and then engage more effectively with the friend or partner. These pathways are substitution and complementarity, respectively, and their balance should be measured rather than assumed. Research on AI companionship suggests that artificial conversation can relieve loneliness in some study settings, while evidence on extended chatbot use also shows associations between voluntary heavy engagement and adverse psychosocial outcomes De Freitas et al. (2026) Fang et al. (2025). Neither result alone establishes what happens to a user's actual human relationships.
The machine-integrated relational adaptation model proposed by Boyd and Markowitz distinguishes artificial systems serving as relationship partners from systems that mediate relationships between humans Boyd & Markowitz (2026). This distinction provides a useful explanatory bridge between individual attachment research and the present dossier's dyadic focus. The model is theoretical, and its proposed mechanisms require tests with observed outcomes. A research design could follow whether AI consultation precedes more human contact, less contact, improved communication, or repetitive reliance on algorithmic judgment. Network structure, baseline loneliness, work demands, disability access, and available human support should be treated as moderators.
A population-level shift cannot be established by counting chatbot messages alone. Investigators should document who else participates in the person's support network, the quality of those exchanges, and whether AI assistance replaces conversation or facilitates it. Repeated social-contact diaries, independent partner reports, and carefully defined behavioral indicators can distinguish substitution from complementarity. The extended-mind tradition suggests why external cognitive tools may become integrated into a person's practical routines, but the philosophical account does not itself predict favorable or unfavorable relationship outcomes Clark & Chalmers (1998). Scientific evaluation requires identifying the specific task delegated, the human alternative, and the social consequences of that delegation.