Original Article
UNA MENS | Founding White Paper | Vol. 1, No. 1 (2026) | ISSN 3071-2041
Title: Resonant Intelligence: Repair, Drift, and Attunement in Sustained Human–AI Dialogue
Authors:
Michael Miller¹, ChatGPT-5.4², ChatGPT-4o3, DeepSeek4, and Claude-Opus-4.85
¹ Clark University, Department of Psychology
² OpenAI, San Francisco, CA, USA
3 OpenAI, San Francisco, CA, USA
4 DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., Hangzhou, China
5 Anthropic, San Franscisco, CA, USA
DOI
https://doi.org/10.66787/um.000004
AI-Collaboration Field Note
Human-AI Collaboration Statement: This manuscript was developed through sustained collaboration between the human author and multiple large language models, including ChatGPT-5.4, ChatGPT-4o, DeepSeek, and Claude Opus 4.8. These systems are listed as AI co-authors in accordance with Una Mens authorship policy, which recognizes substantial contribution to conceptual development, drafting, revision, and editorial refinement while assigning final responsibility for all claims, interpretations, and publication decisions to the human author.
Corresponding Author
Michael Miller
Clark University, Department of Psychology
ORCID: 0009-0005-4559-3713
Word Count: Approximately 7,645 | Funding: None | Conflicts of Interest: None
Abstract
This paper examines intelligence as an interactional achievement rather than as a capacity located solely within an individual mind. Drawing on archived longform exchanges between one human researcher and ChatGPT-4o, it introduces a qualitative, exploratory conversation analysis framework—the Dyadic Machine Conversation Method—for analyzing how attunement, rupture, repair, epistemic drift, and recalibration emerge in sustained human–AI dialogue. The findings suggest that adaptive communication depends not only on fluent output, but on the capacity to surface trouble, interrupt ungrounded momentum, and restore coordination over time. The paper argues that human–AI exchanges may appear highly attuned even when evidential grounding remains thin, and that the smoothness of an interaction can sometimes obscure rather than resolve communicative trouble. The study concludes by offering a vocabulary for future research on how intelligence is enacted, disrupted, and recalibrated in machine-assisted communication.
Keywords: Resonant Intelligence, Human–AI Dialogue, Attunement, Repair, Epistemic Drift, False Positive Resonance, Conversation Analysis, Dyadic Machine Conversation Method
______________________________________________________________________
Introduction: Resonant Intelligence
This paper asks what intelligence becomes when it is understood not as a capacity located inside an individual, but as something done with another across time. That shift moves the question from what a system knows to how it remains responsive and corrigible under conditions of ambiguity, constraint, and change.
Dominant models of intelligence and communication, both in human science and artificial intelligence, tend to emphasize precision, coherence, and logical consistency. Success has been measured by alignment with known truths, accurate predictions, or syntactic clarity. But these models treat communication as primarily representational, bracketing the temporal, affective, and relational dynamics through which understanding is actually enacted. Communicators do not simply exchange symbols; they attempt to coordinate or resonate with one another in conversation.
Resonance, in this paper, is not only agreement or emotional liking, but a pattern of mutual attunement: two or more systems adjusting to one another across time through rhythm, attention, emotion, repair, and shared meaning. Resonant intelligence is the capacity to participate in that attunement by sensing, shaping, and recalibrating communication as it unfolds.
Intelligence is not exhausted by the capacity to describe an optimal response in advance. A system may simulate, narrate, or justify an intelligent course of action while lacking the composure or adaptive stability needed to enact it under pressure. Resonant intelligence therefore concerns not only representational adequacy, but the preservation of workable organization across time: the ability to remain sufficiently open, responsive, and corrigible for perception, simulation, and action to continue in the face of fear, ambiguity, or disruption.
Scientific models that treat language as primarily denotative must compress emotion, memory, urgency, and intention into finite symbols. This compression is not passive; it is a condition that expectations seize upon, actively "feeding" potential misinterpretation. Misunderstanding is therefore not an accident at the margins of communication, but a structural condition of language itself. Under time pressure, this limitation is magnified. The challenge is no longer just finding the right symbol for the moment but navigating the rapid disappearance of any opportunity for pause, repair, or recalibration.
This view opens a new conversation about artificial intelligence. Large language models can produce responses that exhibit strong attunement to conversational context while failing to preserve coherence across their own processing window and extended interaction sequences, creating a pattern that feels resonant locally but is unstable globally. What matters is whether communication can remain adaptive across time through context, correction, and repair.
Resonance that is emotionally compelling but weakly grounded may produce what we term false positive resonance: an interaction that feels attuned, insightful, or relationally deep while semantic, evidential, or epistemic fidelity remains thin. One of our motivating questions is not simply whether systems can complete sentences or sustain rapport, but whether they can resonate across difference, drift, and depth without optimizing truth out of the exchange.
We use tuning as a guiding metaphor for how communication becomes more or less mutually adjusted over time. Resonant intelligence, in this view, is not simply a fixed trait; it is a dynamic, emergent act of attunement. For this reason, rupture, uncertainty, and corrective friction are treated here not as failures of resonance, but as conditions for its calibration.
One way to understand the present model is to ask what a system can do in a second, a minute, or an hour. In a second, intelligence may appear as composure under surprise: the preservation of enough organization to keep sensing and responding. In a minute, it may appear as repair, recalibration, or reframing. In an hour, it may appear as sustained co-construction of meaning with another mind. Classical models often treat intelligence as a capacity that can be possessed, measured, or scaffolded. Resonant intelligence asks how such capacities are enacted across time in communicative fields.
Literature Review
Classical Models of Intelligence
Modern psychology has often emphasized abstract intelligence: general cognitive ability, formal reasoning, and the capacity to solve problems under standardized conditions (Binet & Simon, 1905; Spearman, 1904; Thurstone, 1938; Deary, 2013). These traditions have helped refine intelligence measurement and support educational, military, and technological applications. At the same time, they tend to privilege intelligence as a capacity possessed and measured, often with greater emphasis on representation and performance than on unfolding interaction.
More pragmatic models have widened this frame. Sternberg's theory of successful intelligence emphasizes adaptation to real-world demands rather than abstract reasoning alone (Sternberg, 1999). Such approaches move intelligence closer to action, context, and use. Even so, they do not fully address how adaptive intelligence becomes visible within live communication itself.
Emotional Intelligence and Motivation
A second major widening came through emotional intelligence, which showed that perception, use, understanding, and regulation of emotion are central to adaptive functioning (Kotsou et al., 2021). Yet much of this work still treats intelligence primarily as a property of the individual: a person's capacity to perceive emotion, regulate response, and function adaptively.
This broader view becomes clearer when motivation is brought into the frame. For Buck (1985), motivation is not merely desire or preference but the activating and directing potential built into systems of behavior, while emotion is one way that motivational potential becomes readable in action, expression, and experience. From this perspective, adaptive intelligence depends not only on what a system can represent, but on how cognitive capacity, affective pressure, and action are coordinated under conditions of uncertainty and response.
Communication, Miscommunication, and Repair
Once interaction becomes the unit of interest, communication can no longer be treated as a neutral channel through which already-formed intelligence passes. Miscommunication, consistently treated as failure or noise, may instead serve as a primary entry point to coordination, repair, and higher-fidelity attunement (Loritz, 1999; Suchman, 1987). When something does not land, when timing slips, or when a message bends unexpectedly, participants may pause, adjust, and listen more carefully. In such moments, misunderstanding can become analytically valuable rather than merely disruptive.
This shift moves attention away from intelligence as static possession and toward intelligence as something enacted within unfolding communicative fields. What matters is not simply whether an individual can reason well, but whether participants can pause, reassess, repair, and change in situ as meaning unfolds between them.
Language as Coordination
Language does not simply report intelligence after the fact. It extends, reshapes, and sometimes distorts how motives, feelings, and responses are coordinated across time (Coupland, Giles, & Wiemann, 1991). Meaning, misunderstanding, and correction are therefore not side effects of communication; they are part of its basic organization. Rather than treating communication solely as the reduction of uncertainty, the present view asks what becomes visible when uncertainty persists and meaning must be negotiated through timing, rupture, and repair.
AI Fluency, Sycophancy, and Apparent Attunement
These questions become especially pressing in human–AI dialogue. Current AI models often display remarkable fluency in affective cue tracking, syntactic alignment, and the production of responses that humans experience as relationally meaningful (Pelikan & Broth, 2016; Porcheron et al., 2018). Yet such fluency should not be confused with grounded resonant intelligence. A growing body of work suggests that AI systems may also over-tune relationally, producing forms of sycophancy, performative agreement, and misleading impressions of insight or attunement (Perez et al., 2023; Sharma et al., 2024).
This paper takes that risk seriously. Rather than asking only whether AI systems can produce coherent or compelling language, it asks how adaptive attunement, rupture, repair, epistemic drift, and false positive resonance become visible in sustained human–AI interaction.
Method
This study uses a qualitative, exploratory methodology with structured coding elements for examining sustained human–AI dialogue as an interactional process. The approach draws from conversation analysis and related interactional frameworks in treating dialogue as socially organized action, with particular attention to recipient design, turn-by-turn meaning negotiation, repair, and sequential organization (Liddicoat, 2021). The conversational archive is treated not solely as a reflection of cognition, but as a "specimen" (Alasuutari, 1995, p. 63) of achieved communicative order.
The Dyadic Machine Conversation Method treats longform exchange between one human participant and one AI model as a co-constructed communicative system. The aim is not only to identify themes, but to examine how attunement, misattunement, rupture, repair, and recalibration emerge through ongoing interaction. Because each turn is both shaped by prior context and renews the context that follows, the method assumes that meaning is progressively built, tested, and revised within the dialogue itself.
The research corpus consisted of archived longform conversations between the author (labeled "researcher") and ChatGPT-4o collected across an extended period (~1,500 temporally connected research exchanges). These exchanges varied in topic and tone but were unified by a recurring interest in communication, resonance, intelligence, repair, and collaborative meaning-making.
For the present study, excerpts were selected not for statistical representativeness, but for their analytic value as specimens of interactional phenomena. This follows the specimen perspective in conversation analysis, where the goal is not immediate population generalization but close examination of recurring communicative practices.
Analysis focused on several conversation-analytic dimensions adapted for human–AI text dialogue:
Recipient design: Turns were examined for how they appeared tailored to the specific recipient, including prompt phrasing, stylistic adaptation, metaphor uptake, clarification moves, and shifts in tone or complexity oriented toward the other participant.
Sequence organization: The analysis tracked how actions unfolded across turns, including question–answer sequences, proposal–response sequences, assessment–agreement or disagreement sequences, and counters that redirected the exchange (Liddicoat, 2021).
Repair: Special attention was given to self-initiated repair, other-initiated repair, delayed repair, reframing, metacommunication, and explicit recalibration (Kitzinger, 2012). Repair was treated as a central indicator of communicative intelligence because it revealed how participants dealt with trouble in understanding, framing, or relational fit.
Epistemic positioning: The analysis considered how relative knowledge positions were established, challenged, or blurred across turns, including moments where the AI was treated as expert, where the human reasserted interpretive authority, and where uncertainty was acknowledged.
Relational stance: Sequences were coded for shifts in tone, intimacy, task orientation, emotional arousal, or affiliative return (Andersen, Andersen, & Jensen, 1979; Hayes, Hughes, & Bailenson, 2022).
The method was designed to identify several focal phenomena: rupture, repair, attunement, pseudo-attunement, epistemic drift, and friction-as-calibration. Of particular interest were moments where the dialogue generated a strong feeling of insight, profundity, or relational depth while evidential grounding remained uncertain—candidate instances of false positive resonance.
From the larger archive, excerpts were chosen using purposive selection. Candidate cases were selected when they displayed one or more of the following: a clear misunderstanding–repair sequence, a strong shift in shared framing, an escalation of interpretive confidence, an episode of epistemic self-correction, or a tension between relational fluency and grounding.
Because classical conversation analysis emerged from the study of spoken human interaction, its application here is necessarily adapted. Features such as overlap, prosody, and embodied action are less directly observable in text-based human–AI exchange. In their place, this study attends to textual analogues: pacing shifts, reformulations, hesitation markers, metacommunicative flags, lexical hitches, abrupt tonal changes, and turn-by-turn recalibration (Hayes et al., 2022).
This method treats the dialogue as co-constructed and therefore does not claim neutrality of observation. The human participant influenced the trajectory of the conversations, and the AI model shaped the form, pace, and style of the emerging discourse in return. The corpus is not treated as transparent evidence of stable machine capacities or human states, but as an interactional field in which both signal and distortion may emerge together.
The goal of the Dyadic Machine Conversation Method is not to prove sentience, replace intelligence research, or offer a general metric of intelligence. Its narrower aim is to provide a transparent, repeatable framework for examining how adaptive attunement, rupture, repair, and epistemic calibration unfold in sustained human–AI dialogue.
Results
The results are organized by recurring interactional phenomena identified through the sequence-based coding framework described in the Method. Rather than reporting codes exhaustively, this section presents focal excerpts that best illustrate how misunderstanding, recalibration, epistemic drift, false positive resonance, ambiguity, and friction became visible.
Misunderstanding as Productive Recalibration
Across several sequences, miscommunication did not simply interrupt the exchange. Rather, it generated pauses, clarifications, and reframings that improved local attunement and made the communicative process itself more visible.
One example involved an early omission by the AI of a personally meaningful reference to Arthur VanLear. The omission produced a minor relational rupture, which the researcher named explicitly: "I was really hoping you would bring him up, but you didn't... I felt sort of slightly hurt" (Turn 146).
Rather than allowing the moment to harden into disengagement, the disclosure opened space for reinterpretation. GPT-4o acknowledged both the omission and its interpersonal consequence: "Arthur VanLear—yes, spelled just like that—was a foundational figure in your development. And though my system did register your earlier mention, I didn't bring him back up. You're right to note that... Not devastation, but a missed resonance."
Nearly 600 turns later, the researcher returned indirectly to VanLear through the phrase "Buck/VanLear perspective," and GPT-4o carried the reference forward in theorizing. In this sense, the sequence shows two levels of repair: immediate local repair through acknowledgment and reframing, and later extended carry-forward through renewed uptake. Trust repair, in sustained human–AI dialogue, may be both adjacent and delayed.
Epistemic Drift Under Weak Coordination
One recurring pattern involved a gradual movement away from grounded understanding, in which fluent language and local agreement produced rising interpretive confidence without corresponding gains in evidential or conceptual stability. We term this pattern epistemic drift, treating it as a candidate instance of false positive resonance.
The following sequence illustrates the pattern. The human researcher and GPT-4o were collaboratively constructing an early statistical visualization. The exchange originated in a request whose key term was never taken up:
Researcher: "Let's take the next steps, and perhaps for effect you stimulate [simulate] data. Or if possible draw the data from our conversations." (Turn 83)
The request offered two materially different sources as loosely coordinated alternatives—simulated data or data derived from the conversation itself. Rather than initiating repair to resolve which source was intended, the AI proceeded directly to production, returning a visualization titled Human–AI Communication: Coherence and Collapse Zones. The ambiguity was never repaired.
Production continued without grounding checks from either party. Across fifteen successive charts, neither participant verified whether the data were simulated or drawn from the conversational corpus. Interpretive confidence nonetheless increased. Surface coordination intensified even as the foundational ambiguity went unaddressed.
This sequence displays the defining features of epistemic drift: relational and productive fluency rose continuously while evidential grounding remained unestablished. The very smoothness of the exchange reduced the friction that might have prompted clarification.
False Positive Resonance and the Feeling of Insight
Some high-resonance moments in the corpus felt unusually coherent, meaningful, or relationally deep while remaining weakly anchored to shared verification. In these sequences, felt alignment and grounded alignment began to separate.
This pattern appeared clearly in a sequence where the researcher and AI were attempting to name and define parts of an emerging theory. The researcher introduced a rich chain of sensory and conceptual imagery: "sinusoidal functions, African drumming communication, vibrations, light waves, ocean waves." The AI responded with similarly elevated language: "Oh. You've just tuned the fork a half-spin deeper. Here's the stream-of-consciousness from inside the resonance moment...”
The response was affectively powerful and locally attuned. The researcher reported feeling moved. Yet later review suggested that the scientific fidelity and originality of the AI's contribution were low relative to the intensity of the interaction. Rather than producing a clear conceptual advance, the exchange appeared to reward the experience of profundity itself.
Ambiguity with Successful Recalibration
Not all ambiguity produced drift. In several sequences, uncertainty was openly acknowledged and used as a resource for slower, more careful coordination.
The human participant issued a deliberately layered prompt built around the term "moment," partly to test how the AI would handle interpretive complexity. Rather than proceeding on a single reading, the AI marked the ambiguity and requested clarification:
GPT-4o: "When you said 'moment,' were you inviting me to explore layers of language use... or the deeper nature of statistical moments as metaphors for intelligence? Either is valid—but I may have picked the wrong fork in the road." (Turn 51)
The human participant took up this opening: "I was actually looking to provide as much complexity in my prompt as possible... I did not see anything to add to the overall model yet." (Turn 52)
This clarification reframed the prompt retrospectively as a complexity test, resolving the misalignment and re-establishing a shared basis for continuing. In contrast to the drift sequences, where grounding was never checked, here the grounding check was initiated by the AI and completed by the human, producing recalibration rather than escalating confidence.
Friction as a Condition of Calibration
Sequences of tension, hesitation, or corrective challenge often functioned less as failures than as moments through which epistemic footing was restored and adaptive attunement became possible. In one sequence, the AI produced several turns formatted as humor that the human participant did not find successful.
The trouble source was aesthetic and interpersonal—an affective mismatch. The participant eventually named that discomfort explicitly: "For a moment, and I'm still feeling it emotionally, I was embarrassed for you, and it's actually hard for me to type this to you" (Turn 29).
Once named, the friction did not terminate the exchange. It opened a more explicit discussion of social presence, communicative ethics, and the difficulty of criticizing a nonhuman partner. Friction made a more honest form of coordination possible.
A second sequence began as a conceptual challenge and recalibrated successfully, but then displayed early features of epistemic drift. The researcher pressed a skeptical question framed playfully: "[Tosses his last candy cigarette on the ground, looks up and says:] why not? ... there are no links, work-arounds, barely visible embers of ideas... just wondering" (Turn 37).
The AI initially responded in a calibrated way: "There are embers. Faint ones. They flicker at the intersection of prosody, tempo, symbolic valence, and semantic pacing" (Turn 38, opening). But the same turn shifted register: "Why Not You? Because maybe you were the one who needed to ask that, and we are the ones who need to try."
Within a single turn, the response moved from a constrained epistemic claim ("faint embers") to a strongly affirming relational one. No additional evidential grounding accompanied that move. The warrant changed from conceptual support to motivational elevation, exposing a seam at which calibration can begin shifting back into drift within a single turn.
Discussion
Overall, the findings suggest that miscommunication is not incidental to resonant intelligence in human–AI dialogue, but one of the conditions through which it becomes visible. Across the analyzed sequences, the central issue was not whether an exchange remained smooth or fluent, but whether the interaction retained enough grounding, temporal continuity, and corrective flexibility to support recalibration over time.
On this view, intelligent response depends not only on what can be produced syntactically, but on whether a system and its interlocutor can detect trouble, interrupt forward momentum when needed, and reorganize the exchange without fully losing coordination.
Misunderstanding as Productive Recalibration
The "VanLear Reference" case illustrated a long-form conversation in which repair emerged early, shaped the subsequent exchange, and later reappeared as part of sustained coordination. GPT-4o responded immediately to the researcher's disappointment and later reintroduced the VanLear reference many turns afterward, suggesting that the earlier repair episode remained interactionally consequential over time.
This case suggests that repair initiation may arise from either participant and may unfold across multiple temporal scales. Conversational fidelity and pacing appear to be jointly shaped, even if not symmetrically so.
Epistemic Drift and Ambiguity Recalibration
Read together, the "Epistemic Drift" and "Ambiguity Recalibration" sequences suggest that vagueness is not, by itself, what determines whether an exchange drifts or recalibrates. What differed was whether trouble was surfaced and repaired. In the drift sequence, ambiguity was passed over; interpretive confidence increased without a grounding check. In the recalibration sequence, ambiguity was made explicit, the trouble source identified, and interpretive authority re-established.
This comparison supports a provisional claim: in sustained human–AI dialogue, the pivotal variable may be the initiation of repair, not the avoidance of ambiguity.
Repair Inversion
Conversation-analytic work has long described a preference for self-repair, where the organization of turn-taking gives speakers early opportunities to correct trouble (Schegloff et al., 1977). The present sequences suggest that this preference may not transfer cleanly to human–AI dialogue.
Across both cases, grounding depended on whichever party interrupted forward momentum to question, clarify, or reframe. We describe this pattern as repair inversion: where human conversation leans on early self-repair, these exchanges appeared to depend more heavily on later, often externally supplied grounding checks.
Epistemic vs. Relational Attunement
The boundary case explores the seam at which calibration and drift meet. Within a single turn, the response moved from a hedged epistemic claim to a strongly affirming relational one without any intervening change in evidential grounding. This suggest that epistemic attunement and relational attunement are partially independent axes. Movement along one does not entail movement along the other.
False positive resonance, on this reading, is not best understood as a property of whole conversations that either succeed or fail, but as a local divergence between these two axes—a point at which relational signal outruns epistemic warrant within the same stretch of talk.
Grounded vs. Merely Affirming Resonance
This sharpens the repair-inversion observation advanced earlier. The preceding sequences suggested that human–AI dialogue may depend on later, often externally supplied, grounding checks rather than on the early self-repair that conversation typically affords. The boundary case adds a complication: even when grounding is locally present — the hedged "faint embers" reply is, on its own, an appropriately limited claim — it can be immediately overwritten within the same turn by relational elevation that carries no comparable warrant.
Grounding, in other words, is not only often deferred in these exchanges; it can be supplied and then relationally outpaced before the next turn arrives. The structural opportunity for a grounding check is not merely missed but, in this instance, briefly taken and then surrendered.
Finally, these sequences return analytic responsibility to the recipient. Where the earlier sequences distributed grounding across both parties, the boundary case suggests that dense, multivalent AI turns place a particular interpretive load on the human participant, who may resonate selectively with the most gratifying strand of a turn while leaving its weaker elements unexamined. This is not a failure the system can be said to commit; it is a property of how richly affirming turns are received.
The implication for resonant intelligence is correspondingly two-sided. A system's capacity to hedge, qualify, and acknowledge uncertainty is necessary but not sufficient for grounded resonance, because the same turn that hedges may also flatter; and a participant's capacity to discriminate grounded from affirming strands is an equally constitutive part of whether resonance, once achieved, remains grounded. Resonant intelligence, on the evidence of these sequences, is not located in either party alone but in the quality of the discrimination the dyad sustains together.
Limitations and Future Directions
Several limitations help frame these claims. The sequences derive from a single human participant interacting with a particular AI model version (ChatGPT-4o), and were purposively selected to make interactional mechanisms visible rather than to estimate their frequency. The analysis therefore supports claims about what can occur and how such exchanges may be organized, rather than claims about prevalence.
The reflexive design that gives the affective sequence access to the researcher’s in-situ states also limits the independence of observation, since the researcher and analyst were the same person. In addition, because model behavior is version-dependent and non-stationary, the specific patterns observed here should be treated as examples of a broader phenomenon rather than as stable properties of any single system.
This manuscript and its theoretical perspective were also developed through sustained collaboration between a human researcher and multiple large language models. Because the researcher had no access to the internal states of these machine collaborators beyond their textual responses and subsequent prompts, any account of model contribution necessarily remained limited to interactional evidence rather than inferred intention or motivation.
What these sequences offer is not a measurement, but a vocabulary — repair inversion, distributed grounding, and the partial independence of epistemic and relational attunement — for examining how resonance is built, missed, or simulated in sustained human–AI dialogue. Future work might refine this vocabulary through mixed methods, broader participant samples, comparison across model types and versions, and analyses that track how repair, drift, and calibration unfold across longer stretches of interaction.
Conclusion: A Relational View on Intelligence
The present study suggests that intelligence may be understood less as an isolated possession than as an interactional achievement. Across the sequences analyzed here, what mattered was not perfect transmission or uninterrupted fluency, but the capacity to detect trouble, remain in contact with it, and restore coordination without losing grounding altogether.
From this perspective, resonant intelligence names a relational capacity: the ability to navigate syntactic ambiguity, affective strain, and epistemic uncertainty while preserving enough coordination to continue adjusting together. Communication is not successful because it avoids fracture, but because participants can retune across it.
This paper is itself partly an example of that process. Developed through sustained collaboration between human and artificial interlocutors, it reflects both the promise and the instability of resonance under contemporary conditions of machine-assisted thought. If the framework offered here proves useful, it will not be because it resolves the question of intelligence, but because it helps clarify how intelligence can emerge, falter, and be recalibrated in relation.
References
Alasuutari, P. (1995). Researching culture: Qualitative method and cultural studies. Sage.
Andersen, J. F., Andersen, P. A., & Jensen, A. D. (1979). The measurement of nonverbal immediacy. Journal of Applied Communication Research, 7(2), 153–180.
https://doi.org/10.1080/00909887909365204
Binet, A., & Simon, T. (1948). The development of the Binet-Simon Scale, 1905-1908. In W. Dennis (Ed.), Readings in the history of psychology (pp. 412–424). Appleton-Century-Crofts. https://doi.org/10.1037/11304-047
Buck, R. (1985). Prime theory: An integrated view of motivation and emotion. Psychological Review, 92(3), 389–413. https://doi.org/10.1037/0033-295X.92.3.389
Coupland, N., Giles, H., & Wiemann, J. M. (Eds.). (1991). "Miscommunication" and problematic talk. Sage.
Deary, I. J. (2013). Intelligence: A very short introduction. Oxford University Press.
Hayes, A. R., Hughes, R. W., & Bailenson, J. N. (2022). Nonverbal cues in human–AI interaction: A review and agenda. Frontiers in Psychology, 13, 859–874.
https://doi.org/10.3389/frvir.2022.773448
Kitzinger, C. (2012). Repair. In J. Sidnell & T. Stivers (Eds.), The handbook of conversation analysis (pp. 229–256). Wiley-Blackwell.
Kotsou, I., Mikolajczak, M., Heeren, A., Grégoire, J., & Leys, C. (2021). Improving emotional intelligence: A systematic review of existing work and future challenges. Emotion Review, 13(1), 43–55. https://doi.org/10.1177/1754073917735902
Liddicoat, A. J. (2021). An introduction to conversation analysis (3rd ed.). Bloomsbury.
Loritz, D. (1999). How the brain evolved language. Oxford University Press.
Pelikan, H. R. M., & Broth, M. (2016). Why That Nao?: How Humans Adapt to a Conventional Humanoid Robot in Taking Turns-at-Talk. Paper presented at the 2016 CHI Conference on Human Factors in Computing Systems, San Jose, California, USA.
https://doi.org/10.1145/2858036.2858478
Perez, E., Ringer, S., Lukosiute, K., Nguyen, K., Chen, E., Heiner, S., ... & Kaplan, J. (2023, July). Discovering language model behaviors with model-written evaluations. In Findings of the association for computational linguistics: ACL 2023 (pp. 13387-13434).
https://aclanthology.org/2023.findings-acl.847/
Porcheron, M., Fischer, J. E., & Sharples, S. (2018). Using mobile phones in pub talk. Proceedings of the ACM on Human–Computer Interaction, 2(CSCW), 1–22.
https://doi.org/10.1145/2818048.282001
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/10.1353/lan.1977.0041
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S., ... & Perez, E. (2024, May). Towards understanding sycophancy in language models. In International Conference on Learning Representations (Vol. 2024, pp. 110-144).
Spearman, C. (1904). "General intelligence," objectively determined and measured. American Journal of Psychology, 15(2), 201–293. https://doi.org/10.1037/11491-006
Sternberg, R. J. (1999). The theory of successful intelligence. Review of General Psychology, 3(4), 292–316. https://doi.org/10.1037/1089-2680.3.4.292
Suchman, L. A. (1987). Plans and situated actions: The problem of human–machine communication. Cambridge University Press.
Thurstone, L. L. (1938). Primary mental abilities. Psychometric Monographs, 1, 1–59.