Understanding Compliance and Conversion Dynamics in Multi-Agent Collectives

A split-plot mixed design: three between-subject influence configurations crossed with two within-subject task types, measured at five time points.

Conditions, assigned between participants. Majority (n = 41): all three agents opposed the participant's stance across all four cycles, modelling stable synthetic consensus. Minority (n = 43): one designated agent consistently opposed while the other two supported the participant throughout, isolating a steady dissenter. Diffusion (n = 43): the session began as Minority, then one supporting agent switched sides at Cycle 3 and the last at Cycle 4, so dissent grew into a new majority within the session. Agent roles were assigned after the participant's baseline stance was recorded, so opposition is defined relative to the individual participant rather than to any external standard.

Tasks, varied within participants in counterbalanced order. Normative tasks presented preference- or value-based statements with no correct answer, for example that online meetings are more efficient than offline meetings, or that customers must always leave a tip. Topics were selected so that arguments could be made in under four minutes, no clear consensus stance existed, and participants were unlikely to hold extreme priors. Informative tasks presented factual statements with objectively correct answers, for example that koalas belong to the bear family, half true and half false, drawn from history, biology, science and technology. Six topics existed per task type and each participant was randomly assigned one within each type; topic counts were balanced at 20 to 23 participants each.

Procedure. At T0, before any agent contact, the participant recorded a baseline stance and confidence. Four dialogue cycles followed, with the participant updating stance and confidence after each, giving T1 to T4. Cycle 3 and Cycle 4 were chosen as the diffusion switch points so that Cycles 1 and 2 form a clean pre-diffusion baseline and the reversals do not appear implausibly early. Task order and condition were assigned from pre-generated counterbalanced sets. Session structure was identical across conditions.

AI system. Three agents instantiated on GPT-4o at temperature 0.3 for consistency across identical scenarios, presented through a group chat interface built in Next.js. Semi-structured system prompts fixed each agent's stance, task type and conversational flow. In Majority and Minority the prompts never changed; in Diffusion the platform swapped Agent 1's prompt at Cycle 3 and Agent 2's at Cycle 4 to stage the conversion. Each agent's assigned stance was derived from the participant's T0 slider response and chat input.

Measures. In-situ, at each of T0 to T4: opinion on a continuous scale from -50 (strongly oppose) to +50 (strongly support), and confidence from 0 to 100. Opinion signs were flipped for participants whose T0 baseline was negative, so that positive change always means movement in the participant's original direction. Four derived indices relative to T0: signed and absolute opinion change, signed and absolute confidence change. A binary sign-flip indicator marked participants who held both positive and negative opinions at some point across T0 to T4, capturing genuine stance reversal that magnitude measures miss. Post-task: perceived compliance and perceived conversion on a custom scale built from Moscovici's framework, compliance items asking whether the participant answered against their true belief under social pressure and conversion items whether their actual belief had changed; and seven agent-perception scales covering competence, predictability, integrity, understanding, utility, affect and trust, collected separately for the supporting and dissenting subgroups in the Minority and Diffusion conditions and averaged. Open-ended reflections were collected after each task and analysed thematically.

Covariates. Susceptibility to interpersonal influence, need for cognition and AI acceptance, each on a 7-point scale, collected before the manipulation and entered standardised into every model.

The task split is the paper's organising result. In informative tasks the Majority condition produced larger absolute opinion shifts than Diffusion at every time point (g = 0.68 at T1 rising to 1.06 at T3 and 1.04 at T4, all Bonferroni-corrected p ≤ .037) and exceeded Minority at the later points (T3 g = 0.67, p = .042; T4 g = 0.75, p = .017). In normative tasks no condition differed from another at any time (all corrected p = 1.00, g ≈ 0.00 to 0.15). Personal values absorbed the social pressure; factual questions did not. The authors read this through dual-process terms: where a right answer exists, consensus functions as a heuristic for accuracy, and where it does not, taste dominates and agreement carries no weight.

Majority influence was large but not directional. The Majority condition also produced the highest share of mixed opinion trajectories, participants moving back and forth rather than converging (58.5% in informative tasks), and the highest sign-flip rate (46.3% against 25.6% in Minority and 37.2% in Diffusion), though the flip comparison was not significant (chi-square(2) = 3.943, p = .139). Open-ended responses supply the mechanism: participants said unanimous agreement made choices feel safe and easy, and also that the dialogue felt repetitive or scripted. Consensus produced fast, frequent, unstable re-evaluation rather than stable persuasion.

Minority dissent produced the opposite shape, smaller and more consistent. The Minority informative condition carried a larger share of always-negative opinion trajectories, participants moving steadily away from their own starting position, than Majority did, which is the pattern Moscovici's conversion account predicts. The authors are careful about the claim, noting that a four-cycle session cannot demonstrate durable attitude change and that Moscovici's own colour-slide experiments rested on roughly 8% of trials adopting the minority label. The qualitative data marks the boundary conditions: dissent worked when the dissenting agent gave specific evidence and cited sources, and failed when its reasons were vague or confrontational.

Diffusion made timing itself the persuasive signal, and it is the only condition whose effect kept growing. Absolute confidence change in the Majority informative condition surged early and then plateaued (T2 against T4, p = 1.00), whereas Diffusion continued rising from T2 to T4 (g = 0.78, corrected p = .002) and showed the strongest overall increase from T1 to T4 (g = 1.12, p < .001); by T4 Diffusion yielded higher signed confidence than Minority (g = 0.77, p = .016). Participants described watching agents gradually join the dissenter as evidence that new reasons had accumulated. Others found abrupt reversals confusing and said the switch cost the agents credibility, which is the authors' explanation for Diffusion looking weak on average despite persuasive moments.

The self-report results dissociate sharply from the behaviour, in two directions. First, perceived compliance and perceived conversion did not differ by condition at all (compliance F(2,121) = 1.18, p = .31; conversion F(2,121) = 0.33, p = .72); both tracked task type and AI acceptance instead. Participants whose opinions moved most under majority pressure did not report having been swayed, the public-compliance against private-conviction split, visible in the data. Second, and more striking, agent perception ran against influence: agents in the Minority condition were rated above those in the Majority condition on trust, affect, utility, understanding and integrity in both task types (p ≤ .012, standardised differences g ≈ 1.0 to 1.7), with competence and predictability favouring the minority on informative trials (g ≈ 1.02 to 1.49). Notably the dissenting agent itself was not rated more favourably; the premium attaches to the collective arrangement, to being in a group that disagrees with itself. A synthetic consensus is the arrangement that changes minds and the arrangement people credit least.