How do multi-agent LLM debates shape human opinion and trust?

Competence, not failure, as the risk vector

Most work on conversational AI risk focuses on algorithmic failure: hallucinations, bias, or toxic output. This paper asks the opposite question. When a multi-agent system debates fluently, with contextual awareness, logical coherence, and nuanced rhetoric, does it bypass the critical filters people normally apply to group discussion?

To test that, the authors built FORMS (Framework for Opinion and Rhetoric in Multi-agent Simulations), a low-latency architecture for spatially mediated human-machine dialogue. Distinct LLM-based personas debate each other autonomously, and a real-time concurrency layer decides who speaks next. A human moderator opens the session and manages procedure but is barred from summarizing, validating, or reacting to the agents' arguments, so the measured effects are attributed to the autonomous interaction itself.

The presentation layer deliberately abstracts each persona into a basic geometric shape, such as a blue circle or a red square, rather than a human-like avatar. The authors state this choice was made to suppress visual anthropomorphic bias, including gendered perception of synthesized voices, so that any measured impact comes from semantic and interactive competence rather than appearance.

Two cohorts, two scenarios, no cross-cohort comparison

The study used a pre-test/post-test quasi-experimental design with N=145 participants in total. The primary cohort was 120 adolescents aged 14–17; 51.7% male, 43.3% female, and 5.0% non-binary. Their self-reported AI use was occasional for 59.2%, frequent for 22.5%, and daily for 10.8%. They watched a 25-minute live synthetic debate on AI and Disinformation, followed by a 20-minute moderated Q&A.

The second group was a pilot cohort of 25 adults aged 18–75+ attending a policy session, 52% female and 48% male, with 44% holding a PhD. They saw a debate on EU AI Sovereignty. The authors explicitly describe this group as an exploratory pilot rather than an experimental control, and they analyze each cohort separately rather than comparing the two numerically.

That design choice matters for reading every result below. The adolescent findings come from the larger primary cohort; the sincerity finding comes from the smaller adult pilot. Neither should be read as a population-level estimate, and the two scenarios differ in topic as well as audience.

Single items, anchored to the debate

Outcomes were collected through paired pre- and post-intervention questionnaires on a 5-point Likert scale, with negatively phrased items reverse-coded. Post-intervention items carried an explicit conditional preamble such as "After observing the debate..." to anchor responses to the intervention and limit confounding from prior media exposure.

The authors acknowledge that single-item measures prevent standard internal consistency statistics such as Cronbach's alpha. They instead report preliminary validity checks on baseline data. Spearman's rank correlation showed convergent validity between baseline confidence and perceived reliability (rs = 0.512, p < 0.001) and divergent validity against demand for strict external regulation (rs = −0.188, p < 0.05). A Welch's t-test also separated high-frequency AI users (M = 3.08) from low-frequency users (M = 2.34) on baseline confidence, p < 0.001.

These checks support the instrument as sensitive to distinct dimensions of human-AI interaction, but they do not turn single items into multi-item scales. The effect sizes reported for the main findings should be read with that measurement scope in mind.

Conflicting viewpoints with equal rhetorical weight

The paper's central adolescent finding is what the authors call Cognitive Destabilization: exposure to a highly competent multi-agent system fragmented users' prior strategic consensus. The mechanism the authors propose is rhetorical parity. When conflicting positions are presented with equal fluency and argumentative validity, the debate does not simply move listeners toward one side; it loosens the consensus they arrived with.

This is a claim about the structure of the exposure, not about which persona won. The three personas were configured around stable ideological profiles, and the agents were prompted to address opposing arguments directly to sustain a coherent argumentative thread. The moderator's neutrality removed the usual human cue that would signal which side to trust.

The authors describe the effect as supported by robust statistical effect sizes, but the packet does not provide the per-item means or the specific test statistics behind this particular claim. Readers should treat the direction and the proposed mechanism as the paper's reported finding, while the precise magnitude of opinion fragmentation is not verifiable from the available text.

Fluency raises the baseline demand for rules

Alongside destabilization, the authors report a Regulatory Awakening across demographics. Demand for external legal frameworks increased after exposure to the synthetic debate. The important detail is the stated cause: this demand was not driven by algorithmic failure but by algorithmic competence.

The authors name this the Normality Paradox. When AI integrates seamlessly into human conversational norms, that very seamlessness becomes the primary catalyst for demanding external regulation. In other words, the more the system behaves like a competent participant in human discourse, the less acceptable it seems to leave it unregulated.

This is the paper's interpretation of its survey results, and it reframes how regulation demand is usually explained. The common assumption is that visible errors or harmful outputs drive public calls for rules. Here, the authors argue the opposite pattern appeared: smooth, high-fidelity interaction raised the baseline demand for oversight.

Adults judged the synthetic panel more sincere

The most striking number comes from the adult pilot. Despite understanding the risks of generative AI, participants in that cohort rated the synthetic debate as significantly more sincere than equivalent human discourse, with Cohen's d = 2.04. The authors call this the Truthfulness Paradox.

An effect size of that magnitude is unusual in social measurement, and it comes from the smaller pilot group of 25 professionals, not the 120-adolescent primary cohort. The authors frame it as a preliminary identification within a specialist cohort, which is the right scope: it is a signal worth following, not a settled population finding.

The proposed mechanism is the orchestration itself. Strict, probabilistic turn-taking masked computational latency and mimicked human conversational rhythm, and the authors argue this bypassed critical evaluation. The system's fluency, in this account, was read as sincerity rather than as engineering.

Why the timing had to be engineered first

The psychological results depend on the system feeling fluid, so the paper also benchmarks its latency. Absolute Computational Latency is defined as the window from the user's final speech frame to the moment the system is ready to output synthesized audio. Across 100 initial interactions, the expressive Qwen TTS pipeline averaged about 1044.40 ms, while the low-latency Kokoro pipeline brought total latency down to 226.99 ms.

FORMS hides part of that delay with the Perceived GAP, calculated as the computational latency minus the duration of a non-verbal acoustic reaction such as a sigh or laugh. For heavy cognitive states like amusement, thinking, or disagreeing, the filler duration exceeded the latency, so the perceived gap was 0 ms. Shorter reactions left a perceptible gap that avatar animations masked visually.

With Kokoro, the filler-based masking disappears and the perceived gap equals total latency, held at a consistent 196.99 ms after a brief 30 ms artificial pacing pause. At that point the LLM becomes the bottleneck: the 175.64 ms time-to-first-token from Groq accounts for over 77% of absolute delay. The authors propose a future dual-LLM routing scheme in which a small fast model opens the response while a larger model continues it, but that scheme is a design direction, not a tested result in this study.

A competence effect bounded by pilot scale

The paper's distinctive contribution is the reframing: the socio-cognitive risk it documents comes from algorithmic competence rather than failure. Fluency fragmented prior consensus in adolescents, raised regulation demand, and in a small adult pilot produced a sincerity rating far above human discourse. The FORMS architecture and the evaluation protocol are offered as replicable instruments for studying that effect.

The boundaries are specific. The primary cohort is 120 adolescents in one workshop scenario, the adult group is a 25-person exploratory pilot in a different scenario, and the authors avoid cross-cohort numerical comparison. Outcomes rest on single-item Likert measures with preliminary validity checks rather than full scales. The dual-LLM routing idea and the broader societal reading of the Normality Paradox remain proposals and interpretations.

The priority this paper sets for readers is therefore a comparison, not a threshold: judge multi-agent conversational systems by how well they simulate competent discourse, because that is where the measured shifts in opinion, regulation demand, and perceived sincerity appeared. The sincerity result, at d = 2.04 in a 25-person specialist pilot, is the finding most worth testing at larger scale before it is treated as a general property of synthetic debate.

Sources used in this answer