PaperlayerView exact revision on arXiv

Source-linked AI summary

Shaping Opinion: Quantifying the Psychological Impact of Autonomous Multi-Agent LLM Interactions

Marcos Rodriguez-Vega, Afonso Ferreira, Iru Exposito-Luis, Carolina Polito, Pino Caballero-Gil

arXiv:2609.37369v1cs.HC

TL;DR

Research has focused largely on algorithmic failure, leaving the socio-cognitive impact of algorithmic competence less understood. This paper introduces FORMS and evaluates live synthetic debates with adolescent and adult cohorts. It reports fragmented prior strategic preferences and increased demand for regulation, alongside a preliminary adult-pilot finding that synthetic debate was judged more sincere than equivalent human discourse.

  • Problem

    Research has emphasized algorithmic failures, leaving less evidence on how competent autonomous synthetic panels affect political stances, regulatory demands, and perceived sincerity.

    Source: §I · p. 1

  • Method

    The paper introduces FORMS, an architecture for autonomous multi-agent debate, and evaluates it through exposure studies with adolescent and adult cohorts.

    Source: §I · p. 2

  • Results

    The authors report fragmented prior strategic preferences and increased regulatory demand across demographics, plus a preliminary adult-pilot sincerity effect of d = 2.04.

    Source: §I · p. 2, §VIII · pp. 9–10

  • Takeaways & Limitations

    The authors argue that competence-driven conversational fluidity can trigger cognitive vulnerabilities that failure-focused AI safety paradigms do not capture.

    Source: §VII · p. 9, §VIII · p. 9

    A human moderator may elevate perceived legitimacy and sincerity, while the adult pilot’s n = 25 limits extrapolation and calls for caution.

    Source: D. Limitations, Moderator Bias, and… · p. 8

Abstract

from arXiv · show

Natural-sounding multi-agent conversational AI is increasingly deployed, fundamentally altering human-machine interaction and human information processing. While prior work largely focuses on algorithmic failure, this study investigates the cognitive ergonomics and socio-cognitive impact of algorithmic competence. We present and evaluate FORMS (Framework for Opinion and Rhetoric in Multi-agent Simulations), a low-latency architecture for spatially mediated human-machine dialogue, driven by distinct LLM-based personas and real-time concurrency resolution. To conduct a system test and evaluation of its psychological impact, we exposed an adolescent cohort (n=120) and an adult pilot group (n=25) to a live, moderated synthetic debate. Our findings reveal that exposure to highly competent multi-agent systems triggers "Cognitive Destabilization," fragmenting users' prior strategic consensus. Concurrently, we observe a "Regulatory Awakening" driven by the "Normality Paradox": fluid human-machine interactions inherently increase the baseline demand for external regulation. Furthermore, our pilot study suggests the presence of a "Truthfulness Paradox": despite understanding the risks of generative AI, participants in the adult cohort rated the synthetic debate as significantly more sincere than equivalent human discourse (Cohen's d=2.04). Supported by robust statistical effect sizes, this paper contributes the FORMS architecture and a replicable evaluation protocol, illustrating how high-fidelity conversational systems can reshape human information processing.

I. INTRODUCTION

The paper addresses a gap in research focused on algorithmic failure by examining how competent autonomous multi-agent debates affect people’s stances, regulatory demands, and perceived sincerity. It introduces FORMS and evaluates its psychological impact in adolescent and adult cohorts.

I. INTRODUCTION

  • The paper asks how exposure to a highly competent, autonomous synthetic panel influences political stances, regulatory demands, and perceived sincerity.Prior work emphasized hallucinations, overt bias, and toxic outputs, leaving the socio-cognitive impact of algorithmic competence less understood.Source: §I · p. 1
  • FORMS is a low-latency architecture for panel-style debates, using distinct LLM personas and concurrency-managed orchestration to study synthetic rhetoric.Its geometric-shape presentation is designed to minimize visual anthropomorphic bias so measured effects reflect semantic and interactive quality.Source: §I · p. 2
  • The evaluation exposed adolescents (n = 120) and adults/specialists (n = 25) to a live, human-moderated debate about technology regulation.The paper presents a replicable protocol and uses Cohen’s d to quantify persuasive capacity.Source: §I · p. 2
  • The paper reports Cognitive Destabilization and increased demand for legal frameworks across demographic groups, alongside a preliminary Truthfulness Paradox in the adult pilot.For the Truthfulness Paradox, the adult pilot’s effect size was d = 2.04; the paper describes the finding as preliminary.Source: §I · p. 2

II. RELATED WORK

The related work situates the study at the intersection of low-latency conversational systems and multi-agent behavior. It reviews speed-oriented architectures and prior concerns about whether agent debates sustain diverse, coherent interaction.

II. RELATED WORK

  • The paper frames its related work around low-latency architectures, multi-agent behavioral dynamics, and cognitive security.These domains contextualize the socio-cognitive effects measured in the study.Source: §II · p. 2

A. Low-Latency Architectures for Real-Time AI

  • Prior work on low-latency interaction spans synchronous dialogue, specialized inference hardware, parallel token generation, and efficient speech models –.The review describes these approaches as ways to support fluid voice interaction and rapid conversational scaffolding.Source: A. Low-Latency Architectures for… · p. 2

B. Multi-Agent Dynamics and Alignment Pathologies

  • Multi-agent studies report premature convergence, manipulative behavior, and limits to reasoning gains from unstructured debate –.The review also describes orchestration methods that add coordination signals and verification loops.Source: B. Multi-Agent Dynamics and Alignment… · pp. 2–3

C. Cognitive Security and the Regulatory Awakening

The section connects concerns about AI persuasion and regulation to FORMS, a streaming architecture for multi-agent dialogue. Its orchestration layer ranks candidate speakers from context and profile while allowing moderators to interrupt the autonomous debate.

C. Cognitive Security and the Regulatory Awakening

  • Prior work describes persuasive AI messaging and links user vulnerability to growing regulatory attention, including risk-based governance approaches –.The review notes that users may lower critical defenses when synthetic entities seem objective.Source: C. Cognitive Security and the… · p. 3

III. SYSTEM ARCHITECTURE

  • FORMS separates LLM reasoning from real-time orchestration in a continuous streaming pipeline for spatially mediated multi-agent debates.The architecture is designed to support human-like conversational dynamics and manage multi-party concurrency.Source: §III · p. 3

A. Orchestration and Concurrency Resolution

  • A context-aware NLI model scores each agent’s fit to the incoming utterance and profile, while vocative weighting favors explicitly addressed agents.The Anti-Monologue Constraint excludes the previous speaker from implicit turns; moderator speech can interrupt the loop and redirect the debate.Source: A. Orchestration and Concurrency… · pp. 3–4
  • Concurrency selection runs in O(|A|) time with approximately 18 ms of overhead.The paper attributes this to a single forward pass of a lightweight NLI classifier rather than generative LLM decoding.Source: A. Orchestration and Concurrency… · p. 4

B. Cognitive Ergonomics: Masking the Interaction Gap

FORMS masks computational silence with immediate acoustic and visual reactions while subsequent generation continues. Its presentation layer uses geometric avatars to reduce visual anthropomorphic bias in the evaluation.

B. Cognitive Ergonomics: Masking the Interaction Gap

  • FORMS aims to keep perceived response gaps below 60 ms by triggering acoustic fillers and gaze animations while LLM generation proceeds.The paper contrasts this target with reported tolerance of 300–600 ms in telecommunications and immersion-breaking delays above 1000 ms.Source: B. Cognitive Ergonomics: Masking the… · p. 4

C. Cognitive Inference and Presentation

  • After speaker selection, the Cognitive Layer streams generated text to TTS, and the Presentation Layer synchronizes audio and visual feedback.Geometric avatars are intended to suppress visual anthropomorphic bias so measured effects reflect the system’s semantic and interactive competence.Source: C. Cognitive Inference and Presentation · p. 4

IV. TECHNICAL EVALUATION

The technical evaluation measures FORMS latency and examines how nonverbal reactions and visual feedback shape the perceived gap before a response. The Kokoro comparison shows how removing acoustic masking changes that timing.

A. Absolute Computational Latency (Lc)

  • Absolute Computational Latency (Lc) measures the interval from the user's final speech frame until synthesized audio is ready.The evaluation compares Qwen3-TTS with Kokoro.Source: A. Absolute Computational Latency (Lc) · p. 4
  • 100 initial interactions were used to detail measured NLI, LLM time-to-first-token, and TTS latencies.The evaluation ran local inference components on an NVIDIA RTX 4080 Ti; streaming STT overlaps user speech and adds no cumulative response-readiness penalty.Source: A. Absolute Computational Latency (Lc) · p. 4, §IV · p. 4

B. Perceptual Masking and the Perceived GAP

  • 1044.40 ms average Lc for baseline Qwen creates a conversational void that FORMS addresses with its NLI Reaction Engine.The engine calculates perceived gap using nonverbal reaction duration; Fig. 3 depicts acoustic and visual feedback before the response is ready.Source: B. Perceptual Masking and the… · p. 4, Figure 3 · p. 5
  • 0 ms GAPp occurs for heavy cognitive states, while shorter reactions leave a perceptible gap that avatar animations visually mask.Table II is organized by intent; Dfiller is the duration of a nonverbal acoustic reaction such as a sigh or laugh.Source: B. Perceptual Masking and the… · p. 5, Table II · p. 5
  • With Kokoro, Dfiller = 0, so the perceived gap equals total computational latency rather than being acoustically masked.Kokoro lacks the fine-grained expressiveness to natively generate the complex nonverbal acoustic reactions used by the baseline pipeline.Source: B. Perceptual Masking and the… · p. 5

C. Architectural Evolution: Kokoro and the LLM Bottleneck

Replacing the TTS component with Kokoro sharply reduces system latency and removes the need for complex soundbanks to sustain conversational fluidity. The evaluation then tests socio-cognitive effects with adolescent and adult cohorts exposed to moderated synthetic debates.

C. Architectural Evolution: Kokoro and the LLM Bottleneck

  • 32.91 ms average stream latency with Kokoro accompanies a total Lc of 226.99 ms.This shift removes the acoustic Playback Sound masking block from the earlier temporal pipeline.Source: C. Architectural Evolution: Kokoro and… · p. 5
  • 196.99 ms GAPp remains consistent across interactions after a 30 ms pacing pause, while 175.64 ms LLM TTFT accounts for over 77% of absolute delay.The authors identify the LLM as the primary systemic bottleneck in this configuration.Source: C. Architectural Evolution: Kokoro and… · p. 5
  • A proposed dual-LLM routing paradigm would let a smaller model scaffold the opening chunk while a larger model continues the argument.The larger model is intended to provide deeper rhetorical fidelity and persona adherence while TTS streams the initial audio.Source: C. Architectural Evolution: Kokoro and… · p. 5

V. SYSTEM TEST AND EVALUATION METHODOLOGY

  • The system test used a moderated debate among autonomous personas, with the moderator retaining exclusive authority to interrupt.Analyses were conducted within cohorts rather than through direct cross-cohort numerical comparisons.Source: §V · p. 5

A. Participants and Demographics

  • The study recruited 120 adolescents and 25 adults or specialists, with the adult group serving as an exploratory pilot rather than an experimental control.The adolescent participants were aged 14–17; adult pilot participants were professionals aged 18–75+.Source: A. Participants and Demographics · p. 6

B. Design, Scenarios, and Persona Configuration

The study used a pre/post design around cohort-specific synthetic debates, with neutral moderation intended to isolate effects of agent rhetoric. Paired questionnaires assessed interaction-related constructs, with preliminary validity checks for the single-item measures.

B. Design, Scenarios, and Persona Configuration

  • Participants completed a baseline questionnaire, watched a 25-minute synthetic debate, joined a 20-minute moderated Q&A, and then completed a post-questionnaire.The pre-test/post-test design adapted the debate scenario to each cohort while preserving the panel structure and ideological conflict.Source: B. Design, Scenarios, and Persona… · p. 6
  • Adolescents saw a workshop on AI and Disinformation, while the adult pilot saw a debate on EU AI Sovereignty.The cohort scenarios retained the tripartite panel structure and ideological conflict shown in the persona configuration figure.Source: B. Design, Scenarios, and Persona… · p. 6, Figure 4 · p. 6
  • The moderator was restricted to procedural tasks and could not summarize, validate, or react to agents' arguments.The protocol was intended to mitigate human authority bias and isolate persuasive impact of the synthetic rhetoric.Source: B. Design, Scenarios, and Persona… · p. 6

C. Measures and Instruments

  • Paired pre- and post-intervention questionnaires used a 5-point Likert scale, with negatively phrased items reverse-coded.Post-intervention items used a conditional preamble, and Table III distinguishes repeated-measures hypotheses from post-only experience checks.Source: C. Measures and Instruments · p. 6
  • Baseline checks supported convergent and known-group validity, although single-item measures precluded standard internal-consistency metrics.Confidence correlated with perceived Reliability (rs = 0.512, p < 0.001); high-frequency users' baseline confidence was 3.08 versus 2.34 for low-frequency users (p < 0.001).Source: C. Measures and Instruments · p. 6

VI. SYSTEM TEST AND EVALUATION RESULTS

The results show substantial movement among preferred AI-governance strategies without a net shift toward strict limits, alongside increased demand for legal frameworks. Trust outcomes differed across measures and cohorts: adolescent skepticism shifted modestly, while the adult pilot rated the synthetic debate as more truthful than human political discourse.

A. Governance Shift and Cognitive Destabilization (H1 & H2)

  • 38.4% of adolescents changed their preferred AI-governance approach, but movement into and out of strict limits was symmetrical (p = 1.000).The multidirectional flow is illustrated in Fig. 5; its colors denote regulation, freedom, and education stances.Source: A. Governance Shift and Cognitive… · p. 7, Figure 5 · p. 8
  • Post-intervention demand for external legal frameworks was above neutral (M = 3.48, p < .001), with d = 0.44.The authors describe this as a Regulatory Awakening following one 25-minute exposure.Source: A. Governance Shift and Cognitive… · p. 7

B. Trust Calibration and the Truthfulness Paradox (H3 & H6)

  • For school-task AI reliability, adolescent ratings fell from Mpre = 2.90 to Mpost = 2.71 (p = .055, d = 0.18), a marginal, negligible-effect shift.Among participants initially estimating AI influence capacity as low, that estimate decreased significantly (p = .001, d = −0.48); Fig. 6 depicts a flattening distribution interpreted as skepticism induction.Source: Primary Cohort: Skepticism Induction · p. 7, Figure 6 · p. 8
  • 91.3% of the adult pilot agreed the synthetic debate was more truthful than human political equivalents (M = 4.17/5.00, p < .001, d = 2.04).The authors caution that the pilot cohort's constrained sample size requires careful interpretation; they discuss the Machine Heuristic as a possible explanation.Source: Pilot Cohort: The Truthfulness Paradox · pp. 7–8

C. Cognitive Ergonomics and the Triggers of Regulation (H4 & H5)

The section links participants’ experiences of synthetic debate to demands for regulation: surprise at an AI sustaining an ideological stance and perceptions of conversational fluency were both associated with stricter regulatory preferences. Fig. 7 depicts the pilot cohort’s simultaneous recognition of manipulation risk and high truthfulness attributed to synthetic entities.

H5: The Normality Paradox

  • Fig. 7 depicts the pilot cohort’s simultaneous recognition of manipulation risk and high truthfulness attributed to synthetic entities.The chart is described as highlighting this cognitive dissonance.Source: Figure 7 · p. 8

H4: The Surprise-Regulation Link

  • In the reported comparison, participants who expressed surprise that an AI could sustain an ideological stance demanded significantly stricter legal interventions than their non-surprised peers (p = .031, d = 0.43).Source: The Surprise-Regulation Link (H4) · p. 8

H5: The Normality Paradox

  • Participants who perceived the interaction as ergonomically fluid demanded more regulation than those who perceived it as erratic (M = 3.57 vs. M = 2.88; p = .036, d = 0.65).The open-ended descriptions were coded into fluid/positive and erratic/negative valences by independent researchers.Source: The Normality Paradox (H5) · p. 8

D. Limitations, Moderator Bias, and Statistical Robustness

The discussion connects conversational competence with cognitive vulnerability and regulatory alarm, while qualifying the evidence by the study’s moderator, pilot sample, and measurement choices. It also outlines dual-routing as a future response to network-dependent latency.

D. Limitations, Moderator Bias, and Statistical Robustness

  • A human moderator may have elevated agents’ perceived legitimacy and sincerity, and the adult pilot’s 25 participants limit its findings to exploratory status.The authors call for unmoderated evaluation and cautious extrapolation.Source: D. Limitations, Moderator Bias, and… · p. 8
  • The adult pilot warrants larger cross-cultural replication, while single-item ad-hoc Likert scales limit exhaustive test-retest reliability assessment.The single-item measures were chosen to minimize fatigue in adolescents after a demanding cognitive task.Source: D. Limitations, Moderator Bias, and… · p. 9

VII. DISCUSSION AND FUTURE WORK

  • The paper argues that conversational competence can trigger cognitive vulnerabilities that failure-focused AI safety paradigms do not capture.The central premise links multi-agent systems’ psychological impact to their technical and rhetorical competence.Source: §VII · p. 9

A. Synthesizing Competence and Cognitive Vulnerability

  • The Truthfulness Paradox had a massive effect size (d = 2.04) among specialized adults, while the Normality Paradox had a medium effect size (d = 0.65); AI literacy did not inoculate users against the Machine Heuristic.The discussion says neutral, objective cadence can bypass critical filters.Source: A. Synthesizing Competence and… · p. 9

B. Architectural Roadmap: Resolving Network Dependency via Dual-Routing

  • External LLM endpoints depend on network stability, which can introduce jitter and undermine deterministic latency; future FORMS iterations propose hybrid dual-routing.The orchestration layer would route prompts to two distinct models to balance speed and rhetorical depth.Source: B. Architectural Roadmap: Resolving… · p. 9
  • The proposed local scaffolding model would generate an immediate conversational opening, while a high-capacity model continues the sentence and argumentative thread.The high-capacity model runs concurrently with the initial local response.Source: B. Architectural Roadmap: Resolving… · p. 9

C. Roadmap and Empirical Validation

The roadmap proposes broader, more diverse experiments and a dual-routing architecture, while the conclusion reports that competent synthetic debates can fragment prior consensus and prompt regulatory demand. It also identifies a pilot-cohort tendency to rate synthetic debate as more sincere than equivalent human discourse.

C. Roadmap and Empirical Validation

  • Future work proposes testing a larger, more demographically diverse sample to assess whether removing latency-masking artifacts further amplifies the Normality Paradox.The authors hypothesize that eliminating conversational friction will increase rhetorical persuasion and regulatory demand.Source: C. Roadmap and Empirical Validation · p. 9

VIII. CONCLUSION

  • FORMS is presented as an open-source architecture for resolving concurrency in multi-agent voice interactions and measuring the socio-cognitive impact of synthetic debates.The roadmap says tested configurations, generation parameters, and system prompts are available in the framework repository; dual-routing is planned for later integration.Source: §VIII · p. 9, C. Roadmap and Empirical Validation · p. 9
  • Competent multi-agent debates fragmented users’ prior strategic consensus, while regulatory alarm was linked to algorithmic competence and seamless integration into conversational norms.The paper names these patterns Cognitive Destabilization and the Normality Paradox.Source: §VIII · p. 9
  • The specialist cohort showed a pronounced tendency to assign greater sincerity to synthetic debate than to equivalent human discourse, despite theoretical awareness of generative-AI risks.The paper describes this as a preliminary identification of the Truthfulness Paradox.Source: §VIII · p. 10
Loading 2609.37369v1…