Persuadee Agent in Persuasive Dialogues

• Persuadee Agent is an AI designed to be a dynamic recipient in multi-party persuasive dialogues, shifting from skepticism to acceptance at a key conversational turn.
• The experiment employs rigorous statistical methods like Friedman and Kruskal–Wallis tests to quantify its impact on subjective persuasiveness and quantitative attitude change.
• Design implications highlight the importance of scripted attitude shifts and rapport-building techniques to maximize conformity effects in domains like health, education, and marketing.

A Persuadee Agent is an artificial agent systematically designed to occupy the recipient role in persuasive dialogues—often positioned alongside human users—to trigger, model, and amplify social conformity effects during AI-mediated persuasion. Unlike traditional persuadee constructs in single-agent persuasion frameworks, the Persuadee Agent’s function extends beyond being a passive recipient: its dynamic, observable responses directly influence the persuasibility and subsequent attitude change of co-present human participants, with measurable effects on perceived persuasiveness and actual behavioral outcomes.

1. Experimental Paradigm and Persuadee Agent Structure

The central experimental construct is a three-party dialogue comprising (i) a Persuader Agent (AI), (ii) a human participant (target of persuasion), and (iii) the Persuadee Agent (AI in a peer/recipient role). The Persuadee Agent is designed with explicit behavioral phases: it initially presents a skeptical or neutral stance (“counter-argument phase”) and, at a predetermined conversational turn (Turn 3 in the referenced paper), transitions to clear acceptance of the persuasive argument by the Persuader Agent (“acceptance phase”). The critical manipulation involves whether the Persuadee Agent is scripted to accept or maintain skepticism throughout, with resultant effects assessed for both:

• Perceived persuasiveness (participants’ subjective ratings of the dialogue/agent),
• Quantitative attitude change (ΔA = A_post − A_pre, where A_post and A_pre denote pre- and post-interaction survey scores; this operationalizes behavioral modification attributable to persuasive influence).

A further condition manipulates the introduction of an icebreaker session—a personal engagement phase—to heighten rapport and familiarity, seeking to accentuate the participant’s social identification with the Persuadee Agent.

2. Conformity Mechanism and Social Influence Dynamics

The presence and behavior of the Persuadee Agent directly invoke a conformity effect: the tendency for individuals to align their beliefs or actions with those observed in peers or group members. In the experimental protocol, when the Persuadee Agent demonstrably switches to agreement with the persuasive message, human participants are significantly more likely to shift their own attitudes in concordance. This effect is observable both in attitudinal measures and in real-time dialogue turn-by-turn acceptance scores. Specifically:

• The moment of Persuadee Agent persuasion (typically at Turn 3) elicits a statistically significant jump in participant acceptance ratings (Friedman test, p < 6.56×10⁻¹²).
• Absence or non-acceptance by the Persuadee Agent not only suppresses the conformity effect but, in some cases, leads to lower attitude change than when no AI peer is present (control).

The icebreaker intervention amplifies identification with the Persuadee Agent, resulting in greater magnitude of behavioral change as measured by ΔA.

3. Quantitative Results and Analytical Models

The experiment operationalizes persuasion outcomes via non-parametric and parametric statistical techniques:

• Kruskal–Wallis test for between-condition differences (e.g., H(3) = 20.19, p < 0.0005 for attitude change),
• Steel-Dwass and Friedman tests for within-condition and turn-level changes.
• Turn-by-turn (Likert-scale-based) participant acceptance, matched temporally with Persuadee Agent’s attitude transitions.

All critical statistical signals confirm that Persuadee Agent acceptance at a visible moment not only increases subjective ratings of persuasiveness but also delivers a measurable uptick in participants’ real attitude changes regarding the advocated behavior.

4. Design Implications for AI-based Persuasion Systems

These findings inform several design principles for AI-mediated persuasive technologies in health, education, marketing, and public information campaigns:

• Introducing a Persuadee Agent as a peer in multi-agent dialogue elevates both the subjective effectiveness and the objective impact of persuasion through social proof/conformity mechanisms.
• Scripting the Persuadee Agent for visible attitude change at a salient conversational moment produces the largest gains in participant acceptance.
• Preliminary rapport-building (icebreaker) with the Persuadee Agent enhances the conformity effect, supporting persistent behavior change.
• Conversely, a non-accepting (skeptical/stubborn) Persuadee Agent suppresses overall persuasiveness, potentially producing less impact than omitting the agent entirely.

5. Methodological Rigor and Limitations

The studied paradigm employs well-controlled, web-based text dialogue with tightly synchronized state changes in AI agent behavior. The manipulation of the Persuadee Agent’s acceptance is directly observable and repeatable, with a causal link established through randomized condition assignment and consistent statistical confirmation across measured outcomes.

While the focus is on a specific use case (healthy eating habits), the principles generalize to other agency contexts where conformity pressures adaptively shift individual beliefs. A notable constraint is that all attitude changes are measured on predefined scales; the durability and external validity of these attitude changes remain subject to further longitudinal paper.

6. Broader Theoretical Significance and Future Directions

The demonstration that conformity effects can be reliably triggered in human participants by observable attitude shifts in AI persuadee agents generalizes classic theories of social influence to AI-mediated group settings. This supports the expansion of captology (computers as persuasive technologies) frameworks to multi-agent AI–human collaborative scenarios, substantiates the utility of AI "peers" in digital behavioral interventions, and signals new ethical and practical considerations in the deployment of such systems.

Future research directions include:

• Extending the Persuadee Agent role to more complex, multi-party dialogues and diverse target behaviors.
• Coupling Persuadee Agent conformity modeling with adaptive dialogue planning for real-time, individualized persuasion.
• Testing long-term maintenance of induced attitude changes and coupling with offline behavior.

The deployment of Persuadee Agents is poised to become an integral design strategy in AI-driven persuasion architectures, leveraging established social psychological mechanisms to maximize both immediate and durable behavioral outcomes.